How China Won (2025 State of AI report) - CourtneyR.dev's 2nd Brain

How China Won (2025 State of AI report) 🔗 **Source:** [How China Won (2025 State of AI report)](https://www.youtube.com/watch?v=3_71Nog8JZw) > [!note]+ Document Notes > tags: [ai-trends, open-source-ai, deep-seek] > author: Theo - t3dotgg > source: YouTube > reference: https://www.youtube.com/watch?v=3_71Nog8JZw > date created: 2025-01-24 10:30 > > # 🏷️ Title: How China Won (2025 State of AI report) >  ## 🧰 Action Inbox - [H1] Review OpenRouter analytics for T3 Chat usage patterns — data-driven optimization #analytics - [H2] Evaluate Chinese open-weight models for production use — cost vs performance analysis #evaluation - [H3] Implement WorkOS for enterprise authentication — business deals dependency #integration ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [OpenRouter 2025 State of AI Report](https://openrouter.ai/reports/2025-state-of-ai) - [T3 Chat](https://t3.gg/chat) ## 📝 Summary - OpenRouter analyzed 100+ trillion tokens of AI inference data throughout 2025 to reveal major shifts in AI usage patterns - Chinese open-weight models went from 1.2% to 30% market share, led by DeepSeek, Quen, and Moonshot's Kimi K2 - Role-playing dominates open-weight model usage at 52%, followed by programming at second place - Over 50% of tokens now route through reasoning-optimized models, indicating the rise of agentic inference - Early adopters show much higher retention rates than later users (glass slipper effect) ## 📖 Main Content The 2025 State of AI report from OpenRouter and A16Z reveals dramatic shifts in the AI landscape, particularly the rapid rise of Chinese open-weight models and unexpected usage patterns. The analysis covers over 100 trillion tokens of real-world inference data, providing unique insights into how AI models are actually being used in practice. The most striking finding is the meteoric rise of Chinese open-weight models, which grew from just 1.2% of usage at the end of 2024 to nearly 30% by late 2025. This growth was driven by releases from DeepSeek, Quen, Moonshot (Kimi K2), and other Chinese AI labs that delivered competitive performance at attractive pricing. Contrary to assumptions about AI being primarily used for productivity, role-playing scenarios dominate open-weight model usage at 52%, with programming as the second largest category. This suggests that creative, interactive use cases drive significant adoption, particularly among cost-sensitive users who gravitate toward free or low-cost models. The report also documents the shift toward agentic inference, with over 50% of tokens now routing through reasoning-optimized models. Average prompt length has increased 4x throughout 2025, primarily driven by programming use cases that require longer context and more complex interactions. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/3_71Nog8JZw?t=0) — [H1] 2025 AI Landscape Overview - Key idea: DeepSeek R1 and reasoning models fundamentally shifted the industry - Evidence: "OpenAI introduced 01 which brought reasoning models to the entire world" #### ⏱️ [02:45](https://youtu.be/3_71Nog8JZw?t=165) — [H2] OpenRouter's Unique Data Position - Key idea: OpenRouter hosts all models giving unprecedented usage insights - Evidence: "They do some rough top level analytics on all of the things that are running through open router" #### ⏱️ [06:30](https://youtu.be/3_71Nog8JZw?t=390) — [H1] Chinese Models Market Share Explosion - Key idea: Chinese open-weight models went from 1.2% to 30% usage in one year - Evidence: "At the end of 2024, the weekly share of Chinese openweight models was only 1.2%. Now it's nearly 30%" #### ⏱️ [09:15](https://youtu.be/3_71Nog8JZw?t=555) — [H2] Open vs Closed Weight Distribution - Key idea: Despite OpenRouter's bias toward open models, 70%+ usage remains closed-weight - Evidence: "Even a place like open router is still seeing them as the minority of the usage for AI" #### ⏱️ [12:20](https://youtu.be/3_71Nog8JZw?t=740) — [H1] DeepSeek V3 as Inflection Point - Key idea: DeepSeek V3 was the "holy shit" moment before R1's reasoning breakthrough - Evidence: "V3 was the holy [ __ ] there's something special happening here moment" #### ⏱️ [15:45](https://youtu.be/3_71Nog8JZw?t=945) — [H2] GPT-OSS Underrated Performance - Key idea: GPT-OSS excels in specific domains despite being overshadowed by GPT-5 - Evidence: "It's number one in technology. It's number one in finance. It's number one in science" #### ⏱️ [20:30](https://youtu.be/3_71Nog8JZw?t=1230) — [H1] Role-playing Dominates Open-Weight Usage - Key idea: 52% of open-weight model usage is role-playing, not productivity tasks - Evidence: "Roleplaying is around 52% of the use of openweight models on open router" #### ⏱️ [25:10](https://youtu.be/3_71Nog8JZw?t=1510) — [H2] Price Sensitivity in Role-playing Community - Key idea: Role-playing users gravitate toward free models like DeepSeek - Evidence: "The weebs cannot be stopped" and free model usage patterns #### ⏱️ [30:45](https://youtu.be/3_71Nog8JZw?t=1845) — [H1] Programming Usage Growth - Key idea: Programming tasks have grown dramatically, especially for open-weight models - Evidence: "Programming is winning in the space now" with 39% of Chinese model usage #### ⏱️ [35:20](https://youtu.be/3_71Nog8JZw?t=2120) — [H1] Agentic Inference Becomes Default - Key idea: Over 50% of tokens now route through reasoning-optimized models - Evidence: "We recently crossed the point where over half of the tokens that Open Router is watching get generated are reasoning tokens" #### ⏱️ [40:15](https://youtu.be/3_71Nog8JZw?t=2415) — [H2] Tool Calling Adoption - Key idea: Only 15% of traffic involves tool calling despite growth in agentic use cases - Evidence: "Only 15% or so of the traffic going through open router has a tool getting invoked" #### ⏱️ [45:30](https://youtu.be/3_71Nog8JZw?t=2730) — [H1] Prompt Length Explosion - Key idea: Average prompt tokens increased 4x throughout 2025, driven by programming - Evidence: "The average number of tokens in a request is four times higher than it was at the beginning of the year" #### ⏱️ [50:10](https://youtu.be/3_71Nog8JZw?t=3010) — [H2] Glass Slipper Effect - Key idea: Early adopters of models show much higher retention than later users - Evidence: "People who tried Claude 4 in May, right when it dropped, much more likely to still be using it" #### ⏱️ [55:45](https://youtu.be/3_71Nog8JZw?t=3345) — [H2] Geographic and Cost Patterns - Key idea: North America still dominates usage, cost is not the primary driver of adoption - Evidence: "Quality and capabilities often trump cost" ## 📚 References - [OpenRouter Platform](https://openrouter.ai/) - [A16Z AI Research](https://a16z.com/) - [T3 Chat](https://t3.gg/chat) - [WorkOS Authentication](https://workos.com/) ## 🗣️ Paraphrase - Chinese AI labs achieved remarkable market penetration in 2025, growing from negligible usage to commanding nearly one-third of open-weight model traffic through competitive releases like DeepSeek V3, Quen 3, and Kimi K2 - Creative use cases, particularly role-playing scenarios, drive the majority of open-weight model adoption rather than the productivity tasks many assume dominate AI usage, suggesting different user segments have distinct model preferences - The AI industry has shifted toward agentic workflows where models perform multi-step reasoning and tool interactions, with average request complexity increasing dramatically throughout 2025 ## 📰 Newsletter **The Great AI Shift of 2025: What Open-Source Usage Data Reveals** This week's deep dive into OpenRouter's 2025 State of AI report reveals some surprising trends that WordPress developers and AI practitioners should understand. **Key Takeaway**: Chinese open-weight models went from 1.2% to 30% market share in just one year, while role-playing (not productivity) dominates usage at 52%. This suggests the AI landscape is more diverse and creative than many assume. **For WordPress Developers**: The 4x growth in prompt complexity, driven primarily by programming tasks, indicates AI coding assistants are becoming more sophisticated. Consider how longer context windows and reasoning capabilities might enhance your development workflow. **Practical Applications**: - Open-weight models like Kimi K2 now offer compelling alternatives to closed models - Tool calling and agentic workflows are becoming standard - Early adoption of new models shows significantly higher retention rates Resources: [OpenRouter](https://openrouter.ai/), [AI Model Comparisons](https://artificialanalysis.ai/), [Coding AI Tools](https://cursor.sh/) ## ✅ Action Items #### DevRel - Developer advocacy: Create content around open-weight vs closed-weight model selection criteria, benchmarking demos, cost-performance analysis - Education: Workshop on agentic inference patterns, tool calling best practices, reasoning model integration - Developer experience: Improve documentation for model selection, add usage analytics dashboards - Developer marketing: Case studies highlighting unexpected AI use cases beyond productivity - Community management: Facilitate discussions around model bias and selection criteria #### OSPO - Policy & Governance: Develop guidelines for open-weight model evaluation and deployment - License Compliance: Review Chinese open-weight model licensing terms and compliance requirements - Security Management: Assess security implications of open-weight model adoption - Software Inventory: Track open-weight vs closed-weight model usage across organization - Developer Support: Provide training on model selection criteria and bias considerations - Community Engagement: Engage with open-weight model communities and maintainers #### Community - Maintainers: Support development of model evaluation frameworks and benchmarking tools - Contributors: Contribute to open-weight model ecosystem through testing and feedback - Sponsors: Consider funding open-weight model development and infrastructure - Ecosystem: Coordinate standards for model evaluation and selection across projects #### WordPress - Core AI: Evaluate open-weight models for WordPress AI features, establish model selection criteria - Core Performance: Analyze inference costs and performance implications of different model types - Community: Educate WordPress community about AI model options and trade-offs - Documentation: Create guides for AI model integration and selection - Hosting: Provide guidance on AI model hosting and inference optimization ## 🧪 Hypotheses - Open-weight models will continue gaining market share as quality improves and costs decrease - Role-playing and creative use cases will drive continued demand for less-constrained AI models - Average prompt complexity will continue growing as agentic workflows become standard - Early adopter retention patterns indicate timing of model adoption significantly impacts long-term usage ### 📊 Methodology - Longitudinal analysis of model adoption patterns across different user cohorts - Usage pattern analysis across geographic regions and use case categories - Cost-benefit analysis comparing open-weight vs closed-weight model adoption - Retention analysis measuring user stickiness across model release cycles ### ❓ Questions - How will the competitive dynamics between Chinese and Western AI labs evolve in 2026? - What factors drive the surprising dominance of role-playing use cases in open-weight models? - How will reasoning model capabilities impact traditional software development workflows? ### 🔮 Future Research - Investigate correlation between model licensing terms and adoption patterns - Analyze impact of geopolitical factors on AI model selection and usage - Study evolution of agentic workflows and their infrastructure requirements ### 💡 Implications - Practice: Developers should evaluate open-weight alternatives for cost-sensitive applications - Policy: Organizations need frameworks for evaluating models from different geographic origins - Security: Open-weight model adoption requires different security considerations than closed models - Education: AI literacy should include understanding of model selection criteria beyond pure performance - Funding: Investment patterns may shift toward infrastructure supporting diverse model ecosystems ### 🧠 Learning Styles Assessment Reading level estimate: Grade 12-14 (technical content with domain-specific terminology) #### 👀 Visual learners - Present: Multiple charts showing usage trends, market share evolution, geographic distribution - Improve: Interactive visualizations, model comparison matrices, workflow diagrams #### 🔊 Aural learners - Present: Video format with detailed narration and explanation - Improve: Podcast version, audio summaries of key findings #### 🗣️ Verbal learners - Present: Clear explanations of technical concepts, analogies like "glass slipper effect" - Improve: More real-world examples, case studies, metaphors for complex patterns #### 🧩 Kinesthetic learners - Present: References to hands-on model testing and evaluation - Improve: Step-by-step model evaluation guides, interactive demos #### 🧠 Logical learners - Present: Data-driven analysis, clear causal relationships, statistical trends - Improve: Decision trees for model selection, quantitative frameworks #### 🧑‍🤝‍🧑 Social learners - Present: Community discussions, shared experiences with different models - Improve: Collaborative model evaluation projects, community benchmarks #### 🧘‍♀️ Solitary learners - Present: Detailed technical analysis suitable for independent study - Improve: Self-assessment tools, personal model selection worksheets #### 🌿 Naturalistic learners - Present: Discussion of ecosystem dynamics and competitive evolution - Improve: Analogies to natural selection and ecosystem adaptation ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Well-structured headings, clear chart descriptions - Improve: More detailed alt text for charts, data tables as alternatives to visuals #### Keyboard navigation - Present: Standard video player controls - Improve: Timestamped navigation, keyboard shortcuts for chart sections #### Contrast and readability - Present: High contrast video interface - Improve: Transcript formatting with better visual hierarchy #### Text resizing and reflow - Present: Responsive video player - Improve: Scalable chart graphics, reflowable transcript text #### Media and captions - Present: Video with audio narration - Improve: Closed captions, downloadable transcript, chapter markers #### Forms and interactions - Present: Basic video controls - Improve: Interactive chart exploration, accessible data filtering ### 🌍 DEIB Assessment #### Cultural breadth - Present: Global perspective on Chinese vs Western AI development - Improve: More diverse geographic representation beyond US/China binary #### Inclusive language - Present: Technical language appropriate for developer audience - Improve: Define more technical terms, avoid cultural assumptions #### Neurodiversity support - Present: Clear structure with timestamped sections - Improve: Visual summaries, bullet-point key takeaways #### Representation - Present: Focus on technical capabilities rather than personal characteristics - Improve: Acknowledge diverse global AI research contributions #### --- ## Full Transcript [00:00] 2025 has been a crazy year for AI. At [00:02] the end of 2024, OpenAI introduced 01 [00:05] which brought reasoning models to the [00:06] entire world and right when 2025 [00:09] started, Deepseek rose with Deepseek R1, [00:11] an openw weight model that allowed for [00:13] reasoning to be done for everyone. These [00:16] changes shifted the whole industry and [00:18] we've seen crazy things happening since. [00:21] So crazy that our friends over at Open [00:23] Router took the time to break it all [00:25] down. They have unique insights on [00:27] everything going on because they host [00:29] and let you use every model. They do [00:32] some rough top level analytics on all of [00:34] the things that are running through open [00:36] router which gives them a shitload of [00:38] interesting data to use to help break [00:40] down trends. From the weird use cases [00:42] for LMS and spoiler programming is [00:44] actually number two. You'll be very [00:46] surprised what number one is to the [00:48] split across open and closed weight [00:49] models which has been really cool to [00:51] see. It's over 30% open weight models [00:54] overall now, which is nuts. And the [00:56] different releases that got us here are [00:58] really cool. The absurd growth in token [01:01] usage overall. It's been crazy. [01:03] Seriously, there's so much cool [ __ ] in [01:04] this paper. I can't wait to break it all [01:06] down for you. But as a person who spent [01:08] a lot of money on tokens in 2025, we do [01:10] have bills to pay. So, we're going to do [01:11] a quick break for today's sponsor. Stop [01:13] me if you've heard this one before. I [01:15] rolled my own off and I regret it. For [01:16] T3 Chat, I built my own offplatform [01:18] using open source packages, and it [01:20] worked a bit for the most part. But as [01:22] soon as we wanted to do anything even [01:24] vaguely complex, it fell apart. We lost [01:26] multiple potential deals with businesses [01:29] that wanted to adopt T3 chat because we [01:31] had no way for them to authenticate [01:32] their business users. That has all been [01:34] fixed because we made the move to work [01:36] OS. We are very thankful we did. It has [01:38] allowed for integrations that were [01:39] impossible for us to deal with before. [01:41] We're actually really liking a lot of [01:42] the other subproducts, too. Things like [01:44] Vault are super cool for encrypting user [01:46] data. Things like API keys if users want [01:48] to bring their own a API key to your [01:50] service. Vault makes it trivial to do [01:52] that in a way that is safe and not [01:53] exposing the data. The admin portal is [01:55] the killer feature though. This makes it [01:56] so easy for anybody at any company to [01:59] set up your platform. If the IT team at [02:01] Microsoft has to onboard for your [02:03] product, good luck. Have fun with all [02:05] those calls if you're not using work OS. [02:07] And if you are, you send them a link, [02:08] they click on whatever off method they [02:10] want to use, and they're good to go. If [02:12] you've never had to deal with ADP or [02:14] SAML or OCTA before, I envy you. And now [02:16] you won't have to because you know about [02:18] work OS. Speaking of miserable O [02:20] stories, they're also working really [02:21] hard in figuring out how to off your MCP [02:24] servers, which is a a non-trivial task [02:26] if you've watched my videos on the [02:27] topic. I'm so thankful that works is [02:29] figuring all of this out because I was [02:31] so tired of doing it myself. I highly [02:33] recommend checking them out if you [02:34] haven't at soy. Oo. Shout out to Open [02:37] Router A6 for putting the time into [02:38] doing this type of research. They're one [02:40] of the few companies that's well [02:41] positioned to get this type of data [02:43] since Open Router lets you route to all [02:46] the different models. They get insights [02:47] for all the different use cases and [02:49] they've been collecting some very very [02:51] interesting anonymized data throughout. [02:53] With the release of the first widely [02:54] adopted reasoning model 01 on December [02:56] 5th of 2024, the field shifted from [02:58] single pass pattern generation to [02:59] multi-step deliberation inference, [03:01] accelerating deployment experimentation [03:03] and new classes of apps. As the shift [03:05] unfolded at a rapid pace, our empirical [03:07] understanding of how these models have [03:08] actually been used in practice has [03:10] lagged behind. In this work, we [03:12] leveraged the open router platform, [03:13] which is an AI inference. You know what [03:14] open router is if you're here. They've [03:16] done over 100 trillion tokens of real [03:18] world LLM inference across tasks, [03:19] geographies, and time. In our empirical [03:21] study, we observe substantial adoption [03:23] of openweight models, the outsized [03:25] popularity of creative roleplay beyond [03:27] just productivity tasks that many assume [03:29] dominate LLMs, as well as coding [03:30] assistance categories, plus the rise of [03:32] agentic inference. They also have [03:34] identified different cohorts and groups [03:36] of users that stick around more. The [03:38] glass slipper effect of people who [03:41] showed up early sticking around much [03:43] longer than people who are showing up [03:44] later. So those early people to LMS [03:47] stick around. Findings underscore the [03:49] way developers and end users engage with [03:51] LM in the wild is complex and [03:52] multifaceted. We discuss implications [03:54] for model builders, AI devs, [03:56] infrastructure providers, and outline [03:57] how a datadriven understanding of usage [03:59] can inform better design and deployment [04:00] of LM systems. Should be fun. As always, [04:03] the whole thing is linked in the [04:04] description if you want to read it. We [04:05] will be skimming and finding useful [04:07] parts throughout. Great research [04:08] regardless. You know, they're going back [04:10] in time when they're talking about [04:12] Sonnet 2.1 and anthropics improvements [04:15] on RA. Good old RAG, good old days. So, [04:18] the different sections are open versus [04:19] closed source models, agentic inference, [04:21] category taxonomy, geography, effective [04:24] cost versus usage dynamics, and [04:26] retention patterns. Should be fun. All [04:28] of this is metadata, so there's no user [04:30] content data exposed at all. They used [04:32] the Google tag classifier for content [04:34] categorization. No direct access to user [04:36] prompts or model outputs was available [04:37] for the study. Instead, open router [04:39] performs internal categorization on a [04:41] random sample comprising of [04:42] approximately 0.25% of all prompts and [04:45] responses through a non-proprietary [04:47] module, the Google tag classifier. While [04:49] it represents only a fraction of total [04:50] activity, the underlying data set [04:52] remains substantial given the overall [04:54] query volume presented by open router. [04:56] Google tag classifier interfaces with [04:57] Google cloud natural languages classify [04:59] text content classification APIs. The [05:02] API applies a hierarchal language [05:03] agnostic tonomy to textual input [05:05] returning one or more category paths [05:07] like computers, electronics, [05:08] programming, arts, entertainment, role [05:10] playing, games, etc. with corresponding [05:12] confidence scores in the range of 0 to [05:14] one. The classifier operates directly on [05:16] prompt data up to the first 1,000 [05:17] characters. Classifier is deployed [05:19] within open routers infra ensuring that [05:21] classifications remain anonymous and are [05:22] not linked to individual customers. Huge [05:24] categories of confidence scores below [05:26] the default threshold of 0.5 are [05:27] excluded from further analysis. So if it [05:29] doesn't get scored highly enough in a [05:31] thing, it just ignores it. Cool. So they [05:33] have made these buckets for the [05:35] categories and show which tags link to [05:37] each. I don't know if this was public [05:38] before, but this is really cool cuz I [05:39] look at that data a lot. Programming, [05:41] roleplay, translation, general Q&A, [05:43] knowledge, productivity, writing, [05:45] education, literature, creative writing, [05:47] adult, and others. They also break out [05:50] between open source and proprietary [05:51] variants, origin uh variants, Chinese [05:54] versus everywhere else because China's [05:56] killing it that hard, especially in the [05:58] openweight world. Prompt versus [05:59] completion tokens, which is interesting. [06:01] Prompt tokens being the ones you send, [06:02] completion tokens being the ones you [06:04] receive. And we have geographical [06:06] segmentation. Should be fun. They use [06:08] billing location. Let's take a look at [06:10] the data. Here is one of the most fun [06:13] charts. This is token share by source [06:16] type and origin. uh light blue is [06:18] Chinese openweight models, middle blue [06:21] is anywhere else rest of world [06:23] openweight models and then dark blue is [06:25] closed weight models. So even though [06:27] open router is like main value prop is [06:29] that it's easier to access every model [06:32] with that you would imagine that the [06:34] main use case is open weight models. I [06:35] know it is for us the majority of our [06:37] use case for open router int3 chat is [06:39] models that are hosted by Chinese labs [06:41] where we don't want to use the Chinese [06:42] infra. want to use American providers [06:44] and other providers that have better [06:46] data retention policies. Open router [06:48] almost feels essential if you want to [06:49] use models like Kimmy K2 for example. So [06:52] seeing them still having vast majorities [06:55] of their traffic being closed weight is [06:57] nuts. There was a brief moment around [06:59] when Quen 3 and GBT OSS dropped where [07:04] over 30% of the inference they were [07:06] seeing was open weight but now it's back [07:09] in the 20enter range or so. Very [07:12] interesting. Again, their data is going [07:14] to be biased towards the open weight [07:16] options because they are an easy way to [07:18] use them, test them, and integrate them. [07:20] But even then, they're seeing 70% plus [07:24] using closedweight models. So, as cool [07:26] as openweight models are, even a place [07:28] like open router is still seeing them as [07:30] the minority of the usage for AI. Very [07:33] interesting. I would have expected [07:34] bigger for them in particular. It's also [07:36] crazy to see that at the start of 2025 [07:39] there was very very little usage of [07:42] Chinese openweight models and that very [07:44] quickly changed and by the end of the [07:45] year the majority of openweight models [07:47] are Chinese at least open weight usage. [07:49] While proprietary models especially [07:51] those from major North American [07:52] providers still serve the majority of [07:53] tokens openweight models have grown [07:55] steadily reaching approximately 1/3 of [07:57] usage by late 2025. This expansion is [08:00] not incidental. Usage spikes align with [08:02] major open model releases like Deepseek [08:04] V3 and Kimmy K2. It's hard to understate [08:06] how big of a deal V3 was. I know [08:08] everybody was excited about Deep Seek [08:09] R1, but V3 was the holy [ __ ] there's [08:12] something special happening here moment. [08:14] And Kimmy K2 is honestly similar in this [08:16] regard. I like that they called that out [08:18] here. There's a reason it's now the [08:19] default on T3 chat. If you go to an [08:22] anonymous tab on T3 Chat, Kimmy K2 is [08:25] the default now because it is the most [08:26] pleasant model to talk to that I've [08:28] personally ever used. I genuinely really [08:30] really like Kimmy K2 and I highly [08:32] recommend it if you're looking for a [08:33] good model to talk to. That's why we [08:35] made it our default. They're very very [08:38] competitive. There's also other [08:39] competitive launches like DCV3 further [08:42] editions as well as the GPT OSS models [08:44] and these have been adopted rapidly and [08:46] sustained their gains. GBT OSS is also [08:49] criminally underrated for various [08:51] things. I use it a lot more than y'all [08:52] would probably guess. Since the GPT OSS [08:55] model came out around the same time as [08:58] 04 did, if I recall, it was between 04 [09:00] and GPT5. Because of that, it kind of [09:03] got missed, I feel like, and no one [09:05] really embraced the things it was great [09:07] at. Like we see here, it's the number [09:09] one model for legal questions in the [09:11] legal category on Open Router. It's [09:13] number one in technology. It's number [09:15] one in finance. It's number one in [09:16] science. Hell, it's number two in [09:17] programming. Everything else is quite a [09:19] bit further behind on. I use it for data [09:21] stuff all of the time because on the [09:24] fast providers, if we scroll a little [09:26] bit, you'll see some of them like base [09:29] 10 or where is my boys at Grock or [09:32] Grock? Grock's pulling 511 tokens per [09:35] second. That's 10 times what you get for [09:38] something like GPT5. Insane. There are [09:41] some places that can pull it over a [09:42] thousand. Cerebras is pulling 2,000 [09:44] tokens per second with it. That is [09:45] unbelievably fast. And the value of that [09:48] speed isn't just like, oh, it's really [09:50] fast. It's the speed plus the price plus [09:53] its ability to do things like follow the [09:55] shape of an object well make it great [09:57] for parsing absurd amounts of data. We [09:59] use this for doing sentiment analysis on [10:01] the comments on my videos. We use this [10:03] for analyzing the results from [10:04] SnitchBench. We use this for a ton of [10:06] things. It's a great model. It also can [10:09] run on my computer at reasonable speeds. [10:11] If I open up LM Studio, since my MacBook [10:14] has a ton of memory on it, it can run [10:18] these big models pretty fast. It's able [10:20] to run the GBTOSS 20 bill version at [10:23] like 90 or so TPS, but the 120 bill [10:26] version, I assumed it wouldn't be able [10:27] to do [ __ ] with cuz that's a huge model. [10:30] That is a 60 gig model. I couldn't run [10:32] that faster than like 18 TPS on my [10:35] RTX5090, [10:36] but on here on my M4 Max MacBook, it [10:40] runs quite a bit faster. You have to [10:42] wait for it to load into memory, but [10:44] that's the key. Since it's 60 gigs and [10:46] my laptop has 128 gigs, it can load the [10:49] whole thing into memory. No, this is not [10:52] an M5. The M5 does not have a model with [10:55] this much RAM. I'm on an M4 that I just [10:57] bought because we don't have an M5 Pro [10:59] yet. And this is why you cannot get this [11:01] speed on an M5 right now as write an [11:04] essay about why Rust is better than C++. [11:07] And that is faster than reading speed. [11:10] That's flying. [11:12] This is going this fast not just because [11:14] my processor is fast. It's mostly [11:15] because my RAM is insane and I have [11:18] memory that works with my GPU. Doesn't [11:20] matter how much RAM you put on your [11:22] desktop if you're running a 5090 because [11:24] 5090 has a certain amount of VRAM that [11:26] you can't really expand. So, if you [11:28] can't fit the model in your VRAM, it's [11:30] going to run like [ __ ] I can fit it in [11:31] my VRAM because I'm on a MacBook, which [11:33] means I have unified memory. So, the RAM [11:35] is the same as the VRAM. So, I'm able to [11:37] pull a faster speed here than I can on [11:39] my giant beefy gaming desktop with a [11:41] 5090 on it. Yeah, that was 71 tokens per [11:43] second. That's faster than GPT5 over [11:46] API. It's awesome. That's so cool. I had [11:49] to use half my RAM for it, but now that [11:50] I've ejected this, it should be totally [11:52] fine after it's done beachballing. So, [11:54] it used all my memory. [11:56] And that's all local. That's entirely [11:58] offline. So yeah, GBDOSS criminally [12:01] underrated, but if you know, you know. [12:03] And it definitely caused a spike in [12:05] usage of openweight models. As we see [12:07] here, Kimmy K2, Quen 3, and GPTOSS all [12:10] caused the spike that led to openweight [12:13] models being over 30% of the usage. And [12:15] if we look at the overall token usage [12:18] instead of just as a percent, you'll see [12:20] Open Router had a very good year. Good [12:22] for those guys. At the end of 2024, the [12:25] weekly share of Chinese openweight [12:26] models was only 1.2%. [12:29] Now it's nearly 30% of total usage among [12:31] all models in some weeks. Over the [12:33] 1-year window, they average [12:34] approximately 13% of weekly token [12:36] volume. That being Chinese models are at [12:38] 13%. With strong growth concentrated the [12:40] second half of 2025. This is very [12:42] interesting. It turns out that Deep Seek [12:44] wasn't the point where everyone started [12:46] using these Chinese models. It was [12:48] around halfway through the year when we [12:49] got Kimmy and Mini Max and Quen got good [12:52] and all these other drops happened for [12:55] non-Chinese. Open source models averaged [12:57] 13.7% [12:59] and proprietary models maintained their [13:01] 70% share roughly. This pattern has [13:04] materially reshaped the open- source [13:05] segment and progressed global [13:07] competition across the LM landscape. [13:08] Absolutely. It's been a crazy year for [13:10] openweight models. The fact that the [13:12] default model on T3 chat is an [13:13] openweight model and that like half of [13:15] my personal inference is going through [13:16] openweight models is just kind of crazy. [13:18] Key open- source players. This is by [13:21] number of tokens. Even though I just [13:23] said Deepseek wasn't the main player [13:25] that got this bump, they still are the [13:27] single biggest one pulling 14.37 [13:30] trillion tokens over open router from [13:33] November 2024 to November 2025. That is [13:36] crazy. Quen's at 5.6, six, which is also [13:40] nuts, especially as I personally don't [13:42] find those models particularly great. [13:44] Meta and Mistl did okay here as well. [13:46] Open AAI, especially when you say the [13:48] fact that they put out their first open [13:49] weight model halfway through the year. [13:51] Them being this high is impressive. [13:53] Miniax again coming halfway through the [13:55] year with their first relevant model [13:57] being that high is crazy. Zai doing very [14:00] well as well. They're the GLM models. [14:02] And then Moonshot, my underrated goats [14:04] at 0.92. I'm sad they didn't hit the one [14:06] trill. We're working our way towards [14:08] that with Kimmy as the default now. Hope [14:10] you guys can climb this real fast. Also, [14:13] Miniax was free for a bit too. That's [14:15] another thing to account for with Open [14:16] Router. Models that are provided for [14:18] free tend to get boosted really, really [14:21] hard because a lot of open routers like [14:23] heavy users are cost-sensitive. That's [14:25] also a big part of why the role playing [14:27] community seems to really like DeepSeek. [14:29] It's not so much that they like the [14:30] model. It's mostly that they provide [14:32] them for free and that target and that [14:34] demographic is very cost-sensitive. that [14:36] they love using free models. This is [14:39] also why XAI always brags that they're [14:41] so popular on Open Router. It's cuz they [14:43] provided Grock for free for so long when [14:45] they put out the fast version. [14:47] Obviously, when you're using something [14:48] for free, there is some value exchange [14:50] there. They're almost certainly tracking [14:52] all of the things you send and using [14:53] that for post training. So, know that [14:55] when you use the free models, there's no [14:57] way they're just giving that out for [14:58] free out of the kindness of their heart. [15:00] And now we have the decline of Deep [15:01] Seek's dominance. Oh boy, this will be a [15:04] fun chart. Yeah, as we see here, Deep [15:06] Seek was the king of these openweight [15:09] models for a while, quickly eating into [15:11] Llama's previous like strongholds, but [15:15] then as the summer hit and we got more [15:17] of these Chinese openweight models like [15:19] Moonshot with K2 or Mini Max or ZI, [15:24] you'll see their share just getting [15:26] eaten up fast. But then they dropped [15:29] V3.1 of Deepseek and won a huge chunk of [15:32] it back and they got to this point here [15:34] where it's actually is like the peak of [15:37] Yeah, no, it's not that the peak of Deep [15:39] Seek's ownership was definitely here. [15:41] They lost a whole bunch. They dropped [15:43] V3.1 in September, climbed back for a [15:46] bit, and now it's dwindling. Right now [15:48] it's probably up again cuz 3.2 is nuts, [15:50] but they also sucked at hosting that for [15:51] a while, so who knows? Their near [15:53] monopoly on open weights was shattered [15:56] by the summer inflection. The market has [15:58] since become both broader and deeper [15:59] with usage diversifying significantly. [16:01] New entrance like Quen's models, [16:03] Miniax's M2, Moonshot's Kimmy K2, and [16:05] OpenAI's GPTOSS. I don't know why they [16:08] are just skipping Zai and GLM here [16:10] because those models did really well, [16:12] too. They actually performed higher than [16:14] Moonshot there. Wonder why they skipped [16:16] that. Interesting. Regardless, they all [16:19] did very, very well. By late 25, the [16:21] competitive balance had shifted from [16:23] near monopoly to a pluralistic mix. No [16:25] single model exceeds 25% of open source [16:28] tokens and the token share is now [16:30] distributed more evenly across five to [16:32] seven models. The practical implication [16:34] is that users are finding value in a [16:35] wider array of options rather than [16:37] defaulting to one best choice. Yep. And [16:40] this isn't just like, oh, I prefer this [16:42] model or I prefer this model. This is [16:44] also this one is cheaper and can handle [16:47] more tokens. This one can do reasoning [16:48] with images. This one is really good at [16:51] tool calling, but this one isn't. This [16:53] one handles reasoning well and this one [16:54] doesn't. This one writes well. There's a [16:56] lot of reasons to prefer different [16:57] models. Like I obviously like talking to [16:59] Kimmy K2, but I like doing data analysis [17:01] with GPOSS. There's good use cases and [17:04] bad use cases for all of these. You just [17:06] got to play with them to know. Overall, [17:09] the open source model ecosystem is now [17:10] highly dynamic. Some key insights [17:12] include the top tier diversity where now [17:14] there's tons of different families of [17:16] openweight models. The rapid scaling of [17:18] new entrance, people show up and [17:20] suddenly blow up. And the iterative [17:22] advancement, the longevity of DC's [17:24] presence at the top underscores the [17:26] continuous improvement is critical. [17:28] Their successive releases like chat v3, [17:29] R1, and now v3.1 and 3.2 are keeping it [17:33] competitive even as challenggers emerge. [17:35] OSS models that stagnate in development [17:37] tend to lose shares to those with [17:38] frequent updates at the frontier or [17:39] domain specific finetunes. The open [17:41] source arena resembles a competitive [17:43] ecosystem where innovation cycles are [17:45] rapid and leadership is not guaranteed. [17:47] It feels like the early days of [17:48] closedweight models in the openweight [17:50] world and it's awesome to see. The model [17:52] size versus market fit medium is the new [17:54] small. Very interesting. I definitely [17:57] felt this myself. For example, with [17:59] GBToss that is a medium model because I [18:02] can run it on my laptop. Large [18:04] openweight models are things that you [18:06] cannot run on consumer hardware that [18:07] require specialized hardware. And small [18:09] models are things that you can run on [18:11] cheaper hardware like my 5090 instead of [18:13] my MacBook. I know that sounds silly, [18:15] but the amount of RAM used is an [18:17] important key to the differences here. [18:19] And it is cool to see how many small [18:21] openweight models got useful. It's still [18:23] crazy that there are so many large [18:24] openweight models that are good too [18:27] because there's like the only people who [18:29] benefit from a large openweight model [18:31] are the hosts that can use it because [18:33] they have enough H100s around. So seeing [18:36] large openweight models do much better [18:38] at the start of the year makes sense. [18:39] Something like Deepseek R1 is a very [18:42] large model. You're not running that [18:43] without a distillation of it locally at [18:45] all. But now we're getting more stuff [18:47] like 120 bill models that are reasonable [18:50] to run, but also seeing some growth in [18:52] the small models too. That said, we're [18:55] getting some crazy big open weight [18:57] models. I know Kimmy, for example, is [18:58] like one trillion parameters. Good luck [19:00] running that locally. Their definitions, [19:02] oh, interesting. By their definition, [19:04] GPT OSS is large at 120 bill. I guess [19:07] that's fair, but just interesting to [19:09] see. Small is under 15 bill per, medium [19:11] is 15 to 70, and large is 70 plus. I [19:15] would have cut this differently. I would [19:16] have done small as anything under 30 and [19:18] large is anything o over 150. But I get [19:21] it. Especially now with RAM prices, it [19:23] makes sense to knock these down. The [19:25] data on developer and user behavior [19:27] tells us a nuance story though. The [19:29] figure shows that while the number of [19:31] models across all categories has grown, [19:32] the usage has shifted notably. Small [19:34] models are losing favor while medium and [19:36] large ones are starting to capture more [19:38] value. This is very interesting. The [19:40] size of the model helps determine a lot [19:42] of different things. Obviously, the [19:43] amount of knowledge baked into it, the [19:45] size of it affects that a lot, but it [19:47] also affects the price a ton, too. These [19:50] large models are much more expensive [19:51] than the small ones because you can run [19:52] them quicker and on cheaper hardware. [19:55] So, the interest in small openweight [19:57] models is maintaining and medium is [20:00] growing meaningfully because the medium [20:01] ones are really good price to [20:03] performance value. The small market is [20:05] declining overall. Things like Gemma 312 [20:08] bill, which was released in August, saw [20:10] rapid adoption, but now it's competing [20:12] in a crowded field where users are [20:14] continually seeking the best [20:15] alternative. We're still waiting on [20:16] Gemma 4. I think it'll happen at some [20:18] point this week. Google's been teasing [20:19] it on Twitter all week, but we'll see. [20:22] The medium market, finding the model [20:23] market fit. The medium model category [20:26] tells a clear story of market creation. [20:27] The segment itself was negligible until [20:30] Quen 2.5 coder 32 bill dropped in [20:32] November of last year, which effectively [20:34] established the whole category. The [20:36] segment then matured into a competitive [20:37] ecosystem with the arrival of other [20:39] contenders like Mestral Small in January [20:41] and GPT OSS20 bill in August. Again, I [20:44] would put OSS20 bill as a small model, [20:47] but I understand why they're making this [20:48] distinction. The large model segment has [20:50] been very interesting with Quen 3 235 [20:52] bill or ZI's models as well as the GP [20:56] OSS120 bill all capturing meaningful and [20:59] sustained usage. The plism suggests that [21:01] users are actively benchmarking across [21:03] multiple open large models rather than [21:04] converging on a single standard. Yep, [21:07] absolutely. I play with all of them and [21:09] they all have good use cases. So, I just [21:11] cropped this because I want you guys to [21:13] look and guess without seeing what use [21:17] case do you think is in the yellow. This [21:20] use case is between 40 and 80% of the [21:22] usage of openweight models on open [21:25] router. You've got to be thinking stuff [21:27] like code or maybe chat or knowledge, [21:30] but your guess would be code, right? [21:33] That's what mine would have been. [21:35] Reality is always more interesting cuz [21:37] that is the roleplaying category. Yeah, [21:40] the weebs cannot be stopped. Roleplaying [21:43] is around 52% of the use of openweight [21:47] models on open router. There have been a [21:48] lot of theories as to why this is. The [21:50] one I've grown to believe the most is [21:52] that this category is price sensitive. [21:55] And the free models like Deepseek V3.2 [21:58] free are very well regarded in those [22:00] communities and they are willing to let [22:03] them have all of their data so they can [22:04] get good enough role playinging. [22:06] Interesting. The figure above highlights [22:08] that more than half of all openweight [22:10] models have their usage falling under [22:12] roleplay while programming is the second [22:14] largest category. This indicates that [22:16] users turn to open models primarily for [22:18] creative interactive dialogues like [22:19] storytelling, character roleplay, and [22:21] gaming scenarios and for coding related [22:23] tasks as well. The dominance of [22:25] role-playing hovering at more than 50% [22:27] of all openweight tokens underscores a [22:29] use case where open models have an edge. [22:31] They can be utilized for creativity and [22:33] are often less constrained by content [22:35] filters which make them attractive for [22:37] fantasy or entertainment applications. I [22:39] wish they would show what percentage of [22:41] those are using the free versions of the [22:43] models. That's still my hypothesis, but [22:45] the fact that they are less likely to be [22:47] constrained by the hosts makes a lot of [22:50] sense, too. You can only bake so much [22:51] security into the model itself. Most [22:54] labs put a layer in front of the model [22:56] that will detect if the query is bad and [22:58] block it rather than just baking the [23:00] safety into the model itself because you [23:01] can't adjust that over time. But that [23:03] also means the openweight models if they [23:05] have weird behaviors or they don't tune [23:07] out properly, that's there forever. The [23:09] weights are public. Role-play tasks [23:11] require flexible responses, context [23:13] retention, and emotional nuance, [23:15] attributes that open models can deliver [23:17] effectively without being heavily [23:19] restricted by commercial safety or [23:21] moderation layers. This makes them [23:22] particularly appealing for communities [23:24] experimenting with character-driven [23:26] experiences, fanfictions, interactive [23:28] games, and simulation environments. [23:29] Makes sense. We look at the breakdown [23:32] just within Chinese openweight models. [23:34] It's shifted quite a bit. There was a [23:36] point where programming had overtaken [23:38] roleplay, but now it's a little bit [23:42] closer. Roleplay has knocked its way [23:44] down to around 33% of Chinese tokens, [23:47] which is interesting. And if I recall, a [23:48] significant portion of that is on the [23:50] DeepS free tier. The shift suggests that [23:53] models like Quen and Deepseek are [23:54] increasingly used for codegen and [23:55] infrelated workloads. While highvol [23:57] enterprise users may influence specific [23:59] segments, the overall trend points to [24:01] Chinese open source models competing [24:03] directly in technical and productivity [24:04] domains. And overall over the year, [24:08] technology and programming were 39% and [24:10] role-play was around 33. So looks like [24:13] programming is winning in the space now. [24:15] But what if we break down our usage of [24:17] models in programming by open- source [24:20] versus closed versus not Chinese? Gets [24:23] very interesting again with that surge [24:25] in the summer with Miniax M2, GLM 4.5, [24:28] Kimmy and all of that. We saw a huge [24:31] bump in the percentage of tokens going [24:33] through open router being used for code [24:35] tasks. Massive bump. It flattened a [24:38] little bit. Bumped again in October. I [24:41] curious what openweight model we was put [24:44] out from somewhere else. That bumped [24:46] there might have been a mystral thing. [24:47] We'll see in a sec. And now the [24:49] non-Chinese openweight models are barely [24:51] being used and closed models are still [24:54] winning by far. Let's see how they break [24:56] this down. Chinese open source models in [24:58] blue delivered the majority of open [24:59] source coding help driven by early [25:01] successes like Quen 3 coder. By quarter [25:03] 4, Western OSS models such as Llama 2 [25:06] code and GBT OSS had surged but [25:08] decreased in overall share in recent [25:10] weeks. The oscillation suggests a very [25:12] competitive environment. Yep. Developers [25:15] are open to whatever open source model [25:16] currently provides the best coding [25:18] support. As a limitation, the figure [25:19] doesn't show absolute volumes. Open [25:20] source coding usage grew overall. Also, [25:22] a shrinking blue band doesn't mean [25:24] Chinese open source lost users, only [25:26] relative share. Good to know. Role- [25:28] playing breakdown for open-source [25:30] models. Interesting. [25:33] Non-Chinese openweight models are 43% of [25:37] the role playinging use case. Very [25:39] interesting. The most recent numbers, [25:41] Deepseek V3 is still the most popular. [25:44] R1T2 and V3.2 are close though. Then [25:48] it's Flash. Then it's Deepseek 3.1. [25:52] Then it's Gemini. Then it's Mistral [25:54] Nemo. What? [25:59] What the [ __ ] That does not line up at [26:01] all with what I'm reading here. This [26:04] looked like in the roleplay category, [26:07] non-Chinese open source was doing much [26:09] better, but that does not line up with [26:11] this at all. [26:13] Like others is 42.6%. [26:17] But it's not. [26:19] Yeah. Interesting. Very interesting. And [26:22] you can see here when Grock 4.1 Fast [26:24] first came out and was free, it just [26:26] slaughtered. And as soon as it stopped [26:27] being free, it vanished. [26:30] Again, proving my theory that this [26:32] community wants the cheap [ __ ] There [26:35] you go. 4.1 Fast was number one when it [26:37] was free by a lot. And as soon as it [26:40] stopped being free at mid December, it [26:42] immediately vanished almost entirely [26:44] from the chart. Yeah, I'm surprised they [26:47] didn't break this down a bit more. [26:50] I want more info on the openweight [26:53] non-Chinese models that are being used [26:54] here because they're just not sharing [26:56] enough about it. Very interesting. I [26:59] wish we had more info. Enough about open [27:01] weight in China. Let's talk about the [27:03] rise of a gentic inference. This is the [27:06] amount of tokens that are reasoning [27:08] tokens. And we recently crossed the [27:11] point where over half of the tokens that [27:13] Open Router is watching get generated [27:15] are reasoning tokens. I was saying for a [27:17] bit that the average token isn't read by [27:19] a human. This really emphasizes it. [27:21] Unless you're reading every point of [27:23] reasoning in every trace you do, which [27:25] you can't with a lot of these models, [27:26] they don't even share the reasoning [27:27] data. You cannot possibly be reading all [27:30] the tokens being generated. Those tokens [27:32] for reasoning are making the actual [27:34] answer more likely to be better. And the [27:35] tokens that are being generated are [27:37] often not even being shown to a human in [27:39] the first place. They're being used to [27:40] command tools to go do things. So yeah, [27:43] that's one of the most interesting parts [27:44] of this chart is it's showing one of the [27:46] many angles at which text generation [27:49] models and LLMs are being used to [27:51] generate things that aren't for a human [27:53] to read. As shown in the figure above, [27:55] the share of total tokens routed through [27:56] reasoning optimized models climbed [27:58] sharply in 2025. What was effectively a [28:00] negligible slice of usage in the early [28:02] Q1 days is now over 50%. The shift [28:05] reflects both sides of the market. Oh, [28:07] this isn't what I thought it was. This [28:08] is share of all tokens routed through [28:10] reasoning models versus routed through [28:12] non-reasoning models. So this isn't what [28:14] percentage of output tokens were or [28:15] weren't reasoning. That would probably [28:17] be much more brutal. This is what [28:19] percentage of requests, not even [28:21] requests, of the tokens sent to open [28:23] router as input. What percent of them [28:26] were sent to a reasoning model versus [28:27] sent to a non-reasoning model? That [28:30] makes more sense. Top used models with [28:32] reasoning. Grot code fast one is number [28:34] one because again they offered it for [28:36] free. It says XAI's APIs suck to use and [28:39] sign up for. Most people opted to use it [28:41] through open router. That makes sense [28:43] that it's so high up, but again, it's [28:44] there because it's free. XA's Graco Fest [28:47] one now drives the largest share of [28:48] reasoning traffic excluding free launch [28:51] access. Interesting. Apparently, even if [28:53] they exclude the free launch access, [28:55] it's still number one. That feels kind [28:56] of [ __ ] This is a notable change [28:58] from only a few weeks ago where 2.5 Pro [29:00] led the category and R1 and Quen 3 were [29:02] also in the top tier. This is very [29:04] interesting that Gemini 2.5 Pro is so [29:07] far ahead in Open Router. Again, my [29:10] conspiracy theory there is that using [29:12] Google's models through Google's APIs is [29:15] the stupidest, most miserable [ __ ] [29:16] thing you can do. I say as somebody who [29:19] does it a lot, it is hell. It is awful. [29:22] I do not recommend anybody deal with [29:24] Google's SDKs or APIs directly when you [29:26] can use something like Open Router. I [29:27] would gladly pay them the [ __ ] 3 to [29:29] 5% fee, whatever it is, to just route my [29:32] traffic through Open Router and not deal [29:34] with Google's [ __ ] Both OpenAI and [29:36] Anthropic have competent APIs and [29:38] competent CLI to use the models through [29:40] if you're doing code and competent [29:42] integrations and editors like cursor, [29:44] windsurf, etc. So the need to use [29:46] something like open router to hit them [29:48] is way way lower. With Google models, [29:50] there is basically no competent way to [29:52] hit them other than open router. So [29:53] we're kind of stuck here. But these are [29:55] the biases we have to account for in [29:56] this report because this report is [29:58] generated by a source that inherently [30:00] sees different patterns of traffic. As [30:02] Clay said in chat, open router just [30:04] works as opposed to Google Cloud's [30:05] [ __ ] Totally agree. Uh somebody [30:08] said to check this graph out. O yeah to [30:12] go back to like Grock getting traffic [30:14] cuz it's free. Here you go. See all that [30:16] purple traffic that just vanished all of [30:18] a sudden? This is when they made that a [30:20] free model. And as soon as it stopped [30:22] being free, it disappeared from the [30:24] chart. You get the idea. A lot of this [30:26] data is inherently biased by the nature [30:28] of the users and the nature of the [30:30] reasons they're using models through [30:31] open router. It's still really good [30:33] info, but it's biased. It's also again [30:36] to really just highlight how biased this [30:38] is. There is no world in which anybody [30:40] actually thinks that GPT5 is less [30:43] popular than GPTOSS. [30:46] Obviously, like duh. Why the hell is GBT [30:50] OSS120B [30:51] higher up on here than GPT5 is? It's [30:55] obvious if you've been listening. It's [30:56] because GBT5 is able to just be hit [30:58] through OpenAI's API totally fine. But [31:00] GBToss isn't even hosted by OpenAI. You [31:05] cannot use the GBOSs models by hitting [31:07] the OpenAI infra. You have to use them [31:10] elsewhere, which means you have to deal [31:12] with other providers or use Open Router [31:14] and it will route you to the best [31:15] provider, which is a great great option. [31:18] So GBToss is best used on Open Router [31:21] the same way Gemini is. So yeah, all [31:24] five of these models suck to use outside [31:26] of Open Router and are great to use with [31:28] Open Router. So it makes a lot of sense [31:30] that they are the top five. That does [31:32] not mean any of these models are doing [31:33] better than whatever the Hellanthropic [31:35] is doing or what OpenAI is doing with [31:37] their closedweight models. It just means [31:39] that on this chart with Open Router, [31:42] Open Router gets a lot more traffic for [31:44] models that suck to use outside of Open [31:46] Router, which makes sense. It's kind of [31:48] funny to say, but Open Router's edge is [31:50] that XAI and Google really suck at [31:53] making good developer experiences, and [31:55] they can provide a better one by routing [31:57] the traffic through a layer that doesn't [31:59] suck. Turns out being a rapper is a good [32:01] business decision. And now we have the [32:04] rising adoption of tool calling. This is [32:06] interesting. I would have expected [32:08] higher but again the use cases are [32:10] interesting [32:12] roleplaying for example you're not going [32:14] to be doing a whole lot of tool calling [32:16] in with programming you're probably [32:17] doing a lot more but only 15% or so of [32:20] the traffic going through open router [32:22] has a tool getting invoked very [32:24] interesting the noticeable spike in May [32:26] in the figure above was largely [32:28] attributable to one sizable account [32:30] whose activity briefly lifted overall [32:32] volumes interesting I I'm trying to [32:34] remember who it was somebody who's using [32:35] open router as their default provider [32:37] and then moved off it and it caused a [32:39] spike. It was one of the open- source [32:41] code tools. Can't remember which it was. [32:43] Regardless, you get the idea. Top 10 [32:45] most used models with tool call finished [32:47] reasoning. So, this is models where they [32:50] are generating responses and ending [32:51] because there's a tool call they're [32:52] waiting for, which is how tool calls [32:54] work. So, this is highest percentage [32:56] tool calls. Cloud force on it is still [32:58] the king of tool calls as we could have [33:00] expected. 3.5 and 3.7 were also great at [33:03] it. You can see how quickly when a new [33:05] model drops from Enthropic, their [33:07] traffic for the old one collapses. Like [33:10] this pink here, this was 3.7. The gray [33:13] is 40 mini. So I guess somehow in May, [33:15] 40 mini traffic grew a ton. We just [33:19] barely got sonnet 4 if I recall. Yeah. [33:22] So sonnet 4 dropped near the end of the [33:24] month and most of the Sonnet 37 traffic [33:27] had already started dying. In June, it [33:29] collapsed even more. And by July, Sonnet [33:31] 37's barely even being touched. And [33:34] Claude 4 has crushed. And then slowly as [33:38] new models like 2.5 Flash dropped and [33:40] got good enough to actually be able to [33:42] do tool calls, it started eating into [33:44] the margins of other things. And then 45 [33:46] dropped and crushed everything. That all [33:48] lines up as expected. But also notice [33:51] how little 2.5 Pro got in terms of the [33:55] ownership of this chart for tool [33:56] calling. Part of this is because Gemini [33:59] 2.5 Pro is really expensive. Part of [34:00] it's because it's not the best at tool [34:02] calls. And part of it's because if [34:03] you're willing to pay those prices [34:04] anyways, you might as well just use [34:05] Claude. I I've never been a big 2.5 Pro [34:07] fan. All it has done is cost us a ton of [34:09] money and a lot of support tickets. I'm [34:11] thankful to watch it slowly die. Also [34:13] notice that Gemini 3 Pro is not coming [34:16] up here at all. It is late edition, so I [34:20] wouldn't expect it to have snuck in [34:21] particularly deeply, but I don't think [34:23] it ever will because I think it got [34:25] trumped really fast. Also, GLM is [34:27] killing it in this, too. The GLM models [34:28] are really good at tool calling, and [34:30] it's the only open weight model really [34:32] holding some ground in here. That's cool [34:34] to see. I somehow missed that initially. [34:35] Good for ZI. And now we have the anatomy [34:38] of the prompts. The prompt tokens grew a [34:41] ton. So, when you send a request, the [34:43] average number of tokens in a request is [34:45] four times higher than it was at the [34:47] beginning of the year. That's nuts. I [34:48] did not expect it to grow that much. [34:50] completioner tokens going up makes more [34:52] sense though because again reasoning [34:53] models generate way more tokens way more [34:57] tokens. So this being much higher makes [35:00] sense but only 3x when we got a 4x for [35:02] the input is interesting and programming [35:04] is the main driver behind prompt token [35:07] growth. So the number of tokens being [35:09] submitted for code has gone up a ton [35:11] over time which is why the number of [35:13] tokens per request has gone up so much. [35:15] That makes a ton of sense. Longer [35:17] sequences and more complex interactions. [35:18] The average sequence length has grown [35:20] over time. So this is number of tokens [35:23] per generation. So prompt and completion [35:25] together and programming is going up way [35:28] faster overall. It's meaningfully bigger [35:30] than everything else. So for everybody [35:32] asking why we don't do coding stuff in [35:34] T3 chat, why we're not like letting you [35:36] link your codebase and whatever, it's [35:38] cuz we would be increasing our cost by 4 [35:40] to 6x and we charge $8 a month. Shut the [35:42] [ __ ] up. Unless you are going to build [35:44] something or pay us way more. It just [35:46] doesn't make sense for what we're trying [35:47] to do, which is a really good chat [35:49] experience for a reasonable price. If we [35:51] let you bring your code base, we're just [35:52] going to increase our cost by four to [35:53] 6x. Go use cursor. It's a good tool. [35:56] Agentic inference is the new default. [35:58] Together, these trends, rising [36:00] reasoning, shares, expanded tool use, [36:01] longer sequences, and programmers outsiz [36:03] complexity suggests that the center of [36:05] gravity LM usage has shifted. The median [36:08] LM request is no longer a simple [36:09] question or an isolated instruction. [36:11] Instead, it's part of a structured [36:12] agent-like loop, invoking external [36:14] tools, reasoning over state, and [36:16] persisting across longer contexts. For [36:18] model providers, this raises the bar for [36:20] default capabilities. Latency, tool [36:22] handling, conduct support, and [36:23] robustness to malformed or adversarial [36:25] tool chains is increasingly critical for [36:28] info operators. Inference platforms must [36:30] now manage not just stateless requests, [36:32] but longunning conversations, execution [36:34] traces, and permission sensitive tool [36:36] integrations. Soon enough, if not [36:38] already, Agentic Inference will be [36:40] taking over the majority of the [36:42] inference. Yes. And it's crazy how many [36:44] of these like GPU rental services don't [36:46] understand this stuff at all and are [36:48] forcing companies like Moonshot to go [36:51] put out benchmarks showing which hosts [36:54] do and don't follow the tool call [36:56] protocols properly. How are people using [36:59] LMS? This will be fun. Again, we saw [37:02] role-playing was a big chunk, but we've [37:04] seen programming in particular, [37:05] programming through open router growing [37:08] massively. [37:10] This isn't just LLM programming usage [37:12] growing, although that is a huge [37:13] portion. I think a significant chunk of [37:14] this is that the openweight models that [37:16] are best used through open router have [37:19] become much much better, especially [37:22] around halfway through the year, which [37:23] is why this giant spike started around [37:26] halfway through the year. I'm curious if [37:28] they agree with my analysis here. [37:29] Openweight models got good at code and [37:31] open router is the best way to use [37:32] openweight models. So they saw unique [37:34] growth here but they might try to [37:36] generalize this across the industry [37:37] which I probably won't agree with. Now [37:39] of course LLMs become embedded in [37:41] developer workflows their role as [37:43] programming tools is being normalized. [37:46] The evolution is implications for model [37:48] development including increased emphasis [37:49] on codecentric training data improved [37:51] reasoning depth for multi-step [37:52] programming tasks and tighter feedback [37:54] loops between models and integrated [37:55] development environments. They are not [37:57] acknowledging the fact that this growth [37:59] is almost certainly unique to them. [38:01] Other places in the industry do not see [38:03] code LM usage go from 11% to 50%. They [38:06] did because the models that they are [38:07] best for got way better at code this [38:10] year. Very specific to them. That said, [38:13] anthropic models are at 60% of the share [38:15] for programming based spend on open [38:17] router. So maybe I can't make this [38:21] statement so definitively, but at the [38:23] same time, there's a lot of providers in [38:24] here that are best used through them. [38:27] And I expect Google to have that big of [38:28] a stronghold there, too. It was very [38:30] interesting to see how quickly these [38:32] other options have grown into gotten [38:35] actual footing in the space, but it [38:37] looks like it's a small enough [38:38] percentage that I should shut my mouth [38:39] on the open weight thing. I might be [38:40] wrong on that. Very interesting. And [38:43] Miniax is taking a good slice, too. very [38:45] very interesting to see if we look [38:48] within categories role play only 15% of [38:51] the roleplay is adult according to them [38:53] but over half is games I know stuff like [38:56] Dungeons and Dragons love having AI [38:59] generated role-playing text so that [39:01] makes sense programming is interesting [39:05] how that's split scripting languages [39:07] versus development tools versus other I [39:10] don't think these splits are useful so I [39:12] am going to skip most of this section [39:14] author Level insights by category. [39:16] Different model authors are utilized in [39:18] different usage patterns. Figure below [39:20] shows distribution of content categories [39:22] for the major model families. So [39:24] anthropic obviously mostly programming. [39:27] A little bit of tech, not a whole lot of [39:29] roleplay, lots of code though. Google's [39:33] most popular categories a little more [39:34] varied. A lot more roleplaying, a lot [39:36] less code, decent bit of tech, good bit [39:39] of science, not much analysis. I would [39:41] have thought that people would use them [39:42] more for analyzing documents and things [39:44] and that would be clearer here by [39:45] showing like legal or finance or [39:47] something. XAI was code until they made [39:50] it free and then it became roleplay in [39:52] technology I guess and now it's not free [39:54] anymore. It's going to go back to just [39:55] code. I know how this works. OpenAI has [39:58] had a very weird distribution where it [40:00] was mostly science until June when they [40:02] got better at coding and then it became [40:04] mostly code with technology still being [40:07] high. I'm curious what technology even [40:09] means in these cases. Let's go back up [40:11] up up and see. They don't even have [40:12] technology broken out here. Oh, they do. [40:15] Personal assistance, business and [40:17] productivity software, web design and [40:19] development. That's code [ __ ] So, I [40:22] guess some code is put in there, too. [40:24] Weird category splits. And then Deepseek [40:27] is the roleplay machine. They should [40:30] rename it to Deepseek and Dragons at [40:31] this point. God damn. And then Quen is [40:34] mostly code with a weird variety of [40:36] other random [ __ ] Fascinating. And then [40:39] we have usage across regions. North [40:41] America versus Asia versus Europe versus [40:43] other mostly North America still, but [40:46] Asia's had a bigger and bigger chunk at [40:48] least on open router. Language [40:50] distribution over 80% is still English. [40:52] Checks out. And then the glass slipper [40:54] phenomena. This is what they're talking [40:55] about before where when somebody tries a [40:56] new model really early, they are much [40:59] more likely to stick with it than if [41:01] they come to it later. So, people who [41:03] tried Claude 4 in May, right when it [41:06] dropped, much more likely to still be [41:08] using it than people who tried it the [41:09] month after. People who had tried Gemini [41:11] 25 Pro right when it dropped in June, [41:13] are comically more likely to still be [41:15] using it than people who tried it later. [41:16] Very interesting. The collection of [41:18] retention charts captures the dynamics [41:19] of the LLM user market across leading [41:21] models. At first glance, the data is [41:22] dominated by high churn in rapid cohort [41:24] decay. Yet, beneath the volatility lies [41:27] a subtler and more consequential signal. [41:29] A small set of early user cohorts [41:30] exhibit durable retention over time. We [41:33] term these foundational cohorts. They're [41:35] not merely early adopters. Oh god, this [41:37] is so LLM written. That's so that [41:40] pattern just screams LLM. These cohorts [41:44] are not merely early adopters. They [41:46] represent users who workloads have [41:48] achieved a deep and persistent workload [41:50] model fit. We can't just product market [41:52] fit everything, guys. Once established, [41:55] this fit creates both economic and [41:57] cognitive inertia that resists [41:58] substitution even as newer models [42:00] emerge. We introduced the Cinderella [42:02] glass slipper effect as a framework to [42:04] describe the phenomena. The hypothesis [42:06] posits that in rapidly evolving AI [42:08] ecosystems, there exists a latent [42:10] distribution of high-v value workloads [42:11] that remain unsolved across successive [42:13] model generations. Each new frontier [42:15] model is effectively tried on against [42:17] these open problems. And when a new [42:19] release model happens to match a [42:20] previously unmet technical and economic [42:22] constraint, it achieves the precise fit, [42:24] the metaphorical glass slipper. Okay, so [42:26] if you have a problem that none of the [42:28] models are solving and you try a new [42:29] model right when it drops and it solves [42:31] it, you just stick with it. That's the [42:33] hypothesis. [42:35] Interesting. Can kind of see it. Hard to [42:37] know for sure. It's also small enough [42:39] numbers it's hard to care. Like llama [42:41] did not have any real retention at all [42:44] in comparison to quad 4 sonnet. But even [42:46] then, like a lot of this could be [42:48] measured by how long did it take for [42:50] them to ship something new. Like Cloud [42:52] Force Sonnet, they shipped 4.1 then 4.5 [42:56] pretty quickly after. Gemini 25 Pro, [42:59] they did not ship anything until over 6 [43:01] months later. So obviously that's going [43:03] to stay flatter. I don't buy this part. [43:05] I think this is a bit of a reach. [43:07] Apparently they're using 40 Mini as the [43:10] golden example here. God, I hate that 16 [43:13] months later people are still using 40 [43:15] Mini. That was the model we shipped T3 [43:16] chat with in January. Also, this is [43:18] [ __ ] 20 Flash never had any fit. 20 [43:22] Flash is one of my favorite models. I [43:23] use it for a ton of [ __ ] I really just [43:26] don't buy this part. And then the [43:29] boomerang effect. The amount of reaches [43:30] they're doing to explain this. No, I [43:31] don't believe this at all. Let's look at [43:33] cost versus usage. This will be much [43:34] more interesting. Cost is X-axis. Total [43:37] tokens is Y-axis. Interesting to see [43:39] translation focus on cheaper. Legal also [43:42] focusing on cheaper overall. Tech and [43:44] marketing leaning more expensive and [43:46] programming leaning right in the middle. [43:48] Very interesting. The scatter plot above [43:51] reveals a distinct segmentation of AI [43:53] use cases, mapping them based on their [43:55] aggregated usage volume against their [43:57] unit cost. Both are logarithmic axes. [44:00] Really interesting to see roleplay [44:02] leaning towards more expensive models [44:04] than legal does. Your lawyer's cheaper [44:06] than your Dungeons and Dragons DM guys. [44:08] You heard it here first. An effective [44:10] cost versus usage of models. This is [44:12] fun, too. You can see some models that [44:15] are expensive [44:16] getting much less usage and some that [44:18] are cheap getting much more. But [44:19] overall, it's pretty evened out. There's [44:21] a very slight trend downwards where [44:23] usage goes down as cost goes up. It's a [44:26] weak overall correlation. Interesting. [44:28] Cool to know that the cost is not a big [44:30] driver of usage overall. Quality and [44:32] capabilities often trump cost. That [44:35] checks out. Being cheap isn't enough. A [44:37] model must also be differentiable and [44:39] sufficiently capable. Yep. Cool. What a [44:44] paper. Here is the end of discussion. [44:46] The ecosystem is now multimodel. The [44:49] usage diversity beyond productivity. [44:50] We're seeing lots of other use cases. [44:52] Agentic use cases are going up a ton. [44:54] The geographical split is widening, but [44:57] still mostly North America based. Cost [44:59] versus usage. Cost does not really [45:02] significantly drive usage in ways that [45:05] you would expect. And the glass slipper [45:07] effect, which I don't agree with. [45:11] This was really interesting research. I [45:14] am pumped that they published this and [45:15] shared all of this data. [45:18] This was fun. Thank you to the Open [45:20] Router team as well as A16 for helping [45:22] out with this. I feel like I learned a [45:24] lot and have a much better idea of the [45:26] whole ecosystem and the greater picture [45:27] of what's going on. Curious how y'all [45:29] feel though. Are readthroughs like this [45:31] useful? Are you interested in the [45:32] greater market stuff or was this just a [45:34] really long boring video? Maybe you [45:36] didn't even get to this part. Let me [45:37] know what y'all think and until next [45:38] time. Peace.