Anthropic Releases Opus 5.5, OpenAI Unveils New Models  🔗 **Source:** [Anthropic Releases Opus 5.5, OpenAI Unveils New Models](https://share.snipd.com/episode/f63d7e19-200e-42b0-91e6-9ae2c143d20d) ## Highlights - **Anthropic Cuts Opus Pricing While Accelerating Releases** - Anthropic priced Opus 5.5 at $20 per million output tokens, down from $25, while claiming stronger coding and knowledge performance. - The rapid release, only two months after Opus 5, shows Anthropic is pursuing frontier improvements at an unusually fast pace. Transcript: Jaeden Schafer Okay, let's talk about what's going on with Anthropic. What's interesting on this new model that they've put out is that the output tokens are going to drop from $20 per million, or they're going to drop to $20 per million from 25. So that's a 20% price improvement. This new model is also going to be better than the larger Fable model on a bunch of different coding and knowledge tasks. A lot of times when they're announcing these things, it's kind of their internal benchmarks. So I would just say, you know, a word of a word of caution there. Not necessarily something that's going to be, you know, universally agreed upon. A lot of times when they drop the model, they just like to tell everyone how good they internally have found their model. I'm not, you know, I'm not saying they're doing that, but this is kind of, kind of common. One thing that is interesting though, Opus 5.5 is coming out just two months after Opus 5, which was in July 24th that came out. So, I mean, they're making pretty, pretty solid and steady improvements on the models. Two months later, they have, you know, after five, they have 5.5. So they are moving. They're, they're really pushing as far as a company that's trying to quote unquote pace the frontier it feels like they are moving at a really fast speed sonnet 5.5 and 5.5 haiku are also Going to come in the coming weeks ([Time 0:01:44](https://share.snipd.com/snip/20c57c3d-b05b-4ab7-bb99-c722b5865ded)) - **Anthropic’s Expanding Model Lineup Creates Confusion** - Anthropic’s growing lineup now spans Haiku, Sonnet, Opus, Fable, and Mythos, making model selection increasingly complicated. - Jaeden Schafer suspects Fable and Mythos may share a base model but differ through guardrails or access limitations. Transcript: Jaeden Schafer And they both will apparently have very similar improvements what's interesting to me and also i don't know i actually don't love this is the fact that and maybe it just goes to show like How drastically different the training process is for the models or maybe the data sets that they're using but the thing i don't love is that we went from anthropic having like three Models three different tiers you know low medium high and inside of each tier you can kind of like crank up the intelligence level and then now we have added the whole new level which was We went from just having that you know haiku sonnet opus then we added fable and on top of that we have mythos which is like the super hacking coding one that you know no one can get access To anyway so really anthropic has like five models out there although i suspect that mythos and fable are like some variation of the same model just with different guardrails or limitations But they give them different names so it's like no don't worry we only give the dangerous one to people that are approved but in any case I mean they basically have five models and it gets Really annoying and complicated when you're trying to remember the names. OpenAI was accused of this for a very long time because we had you know GPT-3 and GPT-3 and then we had you know all sorts of random names in there, GPT-4. Like, I mean, it just, it got a little ridiculous with the naming conventions. And so it seems like OpenAI is copying Anthropic in a sense that they have kind of the small, medium, and large model sizes. It feels like Gemini did the same thing. So we have kind of the same, I guess we have the same idea. ([Time 0:02:54](https://share.snipd.com/snip/18211f9d-e026-44ab-8ce3-4a329ecd8b3d)) - **Opus 5.5 Brings Frontier Skills Downmarket** - Anthropic says Opus 5.5 matches Fable 5.1 on most tasks while costing 40% less and leading in agentic coding, computer use, and knowledge work. - External evaluators, including Meta and Frontier Design, tested the model before release. Transcript: Jaeden Schafer Cyber security capabilities are matching mythos and that is triggering a bunch of the same safeguards for anthropic that they're applying towards fable models so apparently this Model is pretty good as you'd suspect these things keep getting better and better they had a tweet that they put out today where they said introducing claude opus 5.5 the first model In our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 for most tasks and costs 40% less to run than Opus 5. And that's basically the headline here is, I mean, Fable 5.1 just came out and they're like, look, it's just as good as Fable 5.1, which to be honest, I'm using Fable 5. I'm not even using Fable 5.1. I'm using Fable 5 for most everything on Anthropic because it just burns through too many tokens to use the best models. So yeah, I think that just going with this is going to probably give you some good boosts. They said, Opus 5.5 is our first model since we called for pacing the frontier. As with previous models, it was tested by external evaluators before release, including Meta and Frontier Design. Our most comprehensive alignment tested achieves the strongest score to date. Opus 5.5 is a major step up from Opus 5, leading on agentic coding, ([Time 0:04:43](https://share.snipd.com/snip/568aecac-14d9-454f-86bb-847664d2875c)) - **Faster Responses Matter More Than Lower Model Costs** - Model providers often improve older systems’ efficiency while newer frontier models remain expensive and compute-heavy. - Jaeden Schafer considers Opus 5.5’s 30% faster output more practically valuable than its headline cost reduction. Transcript: Jaeden Schafer Computer use, and knowledge work. Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that. Kind of interesting though because i mean i feel like we keep just like getting you know the colossal like a super inefficient model that's like super smart and super good and that's Great whatever but then they're always like look you know we we just made it so 5.5 is like more efficient than five and it's like like whatever i mean it's just a variation of the older, Like you're just, you're giving us more efficient, but like older models, but the newer models, it's not like when, you know, opening eyes Astra or like Mythos comes out. They're like, here's Mythos, the smartest model. And it's like way more efficient. Like no, the newest models are always just as inefficient. And then as they optimize the older models, they get more efficient, but they're kind of always in the same place. And we're slowly, I guess, just getting more intelligent. I don't know if there's much to complain about. I mean, it's great. This is what progress looks like. But I just don't feel like it's this big headline like it costs 40% less. It's like fantastic until two months later when the next one comes out. And then I'm not going to be using this 40% less one because it's going to seem dumb compared to the new one. So whatever. I digress here. They said, at its default effort settings, Opus 5.5 delivers frontier results for a fraction of the cost per task, often beating other models running at their highest settings. It also generates outputs more than 30% faster than Opus 5. And for that, I will actually say any day of the week, you can tell me how much more efficient or inefficient the model is. But if it is 30% faster, I will say congratulations. I'm actually really excited about that. ([Time 0:05:56](https://share.snipd.com/snip/272c16d6-ebad-4d64-8f28-bdcbc6ffc10a)) - **OpenAI Answers Anthropic With Cheaper Specialized Models** - OpenAI released GPT-6 Sol and Luna shortly after Anthropic’s announcement, cutting API prices by 50% versus GPT-5.6 promotional pricing. - Sol targets complex coding, while Luna handles high-volume summarization and extraction and is available to free users. Transcript: Jaeden Schafer Okay, we also have big news though out of OpenAI, who not to be outdone, released their own improvements on their tier two i guess you could call them models and get this this all came out An hour and a half or 90 minutes after anthropic put theirs out opening i came out and put their their uh new update out as well so they shipped gpt6 soul and luna it's half the api price so I think opening or anthropic went down 40%. OpeningEye is like, look, we can beat you. We'll do 50. We'll do half the price, 50%. It is the quote unquote, smaller sibling of the GPT-6 Astra. And it is going to be a lot smarter, have the price cut. And of course, yeah, this is coming just a couple hours after Anthropoc. So there is definitely a little bit of friendly competition going on right here. Solz is right now targeting complex coding work. Luna is aimed at high volume clerical tasks. So it's doing things like summarization and extraction. Both of the models are live in chat GPT work codecs and the chat GPT API, which hallelujah, I really do appreciate when we release the model and get it in the API right away. ([Time 0:08:37](https://share.snipd.com/snip/d4d894f1-0419-4ed1-a9ce-128a1b689d56)) - **Running Out Of Credits Changes Jaeden’s Model Workflow** - Jaeden Schafer regularly exhausts his ChatGPT work-model credits within two days while using computer-use tools. - When limits arrive, he switches to Claude or Meta’s free Muse, which he says resembles Claude Cowork and ChatGPT Work. Transcript: Jaeden Schafer I've been playing with a handful of them. These ones just came out, though, about three hours ago. So they're pretty new. People are sharing some of the examples of things that they're doing with them. But at the end of the day, I think these examples, like anytime we have these kind of like, I guess you call them tier two models or tier two and tier three models coming out. It's a little bit less exciting, in my opinion, because the like when they're like, Oh, look, we have the brand new Astra model out. The thing is like, you know, and then everyone's trying their demos are like, Look, I got to one shot this crazy game. And I rebuilt GT, you know, GTA six with it, and blah, blah, blah. And you know, people are just doing their wild demos, because it's the best models, they're pushing it to limits when these new ones come out, no one's like, guys check out the way that I like saved 40 on like the task that i was already doing that's kind of like boring and mundane like i think that those are real improvements in the industry that people are getting a lot Of value out of them but you just see less exciting less exciting stuff and at the end of the day when you're low on tokens which happens to me constantly like i'm using chatPT work to get A lot of my work done. It's basically, I think my like prime, you know, computer use model right now, but I'll burn through all of my tokens in two days. And then I have to wait for reset, which I just inevitably switch over to Claude or now Muse. ([Time 0:11:25](https://share.snipd.com/snip/55a4e3e7-94f6-46c1-b999-892c0e52187b)) - **Boring AI Improvements May Deliver The Most Value** - Tier-two models receive less attention because they enable mundane savings rather than spectacular demos, despite offering substantial practical value. - Their strongest use case is affordable, large-scale work that does not require the best reasoning model. Transcript: Jaeden Schafer But at the end of the day, I think these examples, like anytime we have these kind of like, I guess you call them tier two models or tier two and tier three models coming out. It's a little bit less exciting, in my opinion, because the like when they're like, Oh, look, we have the brand new Astra model out. The thing is like, you know, and then everyone's trying their demos are like, Look, I got to one shot this crazy game. And I rebuilt GT, you know, GTA six with it, and blah, blah, blah. And you know, people are just doing their wild demos, because it's the best models, they're pushing it to limits when these new ones come out, no one's like, guys check out the way that I like saved 40 on like the task that i was already doing that's kind of like boring and mundane like i think that those are real improvements in the industry that people are getting a lot Of value out of them but you just see less exciting less exciting stuff and at the end of the day when you're low on tokens which happens to me constantly like i'm using chatPT work to get A lot of my work done. It's basically, I think my like prime, you know, computer use model right now, but I'll burn through all of my tokens in two days. And then I have to wait for reset, which I just inevitably switch over to Claude or now Muse. Muse is actually a big shout out to Muse. They're doing amazing. Never thought I'd say that about a meta product. Go try them out. If you haven't already, it's free and it's like Claude co-work and chachip t work but in any case yeah the the better models they have the splashier demos but i think once for me like when I run out of all the credits on the best models like the astras i'm gonna go down to these new models that they've launched i'm gonna be using gpt6 soul and gpt6 luna i, depending on how many Credits, if I'm really, you know, if I'm really hurting bad with 5% credits left, I'll probably just, you know, switch everything off to Luna and only try to ask a few questions here and There. I think probably a lot of people, if they're using it with these kind of subsidized subscription plans, they'll be doing that. And it's kind of the best use case for a lot of these. But for big enterprises, when you have massive scale projects and you don't need the best reasoning model, I think that these models will do great for a lot of different tasks. ([Time 0:11:33](https://share.snipd.com/snip/7a8acca4-5d4e-4111-a209-2b12b333cd67))