OpenAI Fights Back - CourtneyR.dev's 2nd Brain

OpenAI Fights Back ![rw-book-cover](https://i.ytimg.com/vi/vu8X3YroB-w/sddefault.jpg) 🔗 **Source:** [OpenAI Fights Back](https://www.youtube.com/watch?v=vu8X3YroB-w) > [!note]+ Document Notes > tags: [gpt-6-1-sol, ai-model-pricing, coding-agents, model-benchmarks, prompt-caching] > author: Theo - t3․gg > source: Theo - t3․gg (YouTube) > reference: https://www.youtube.com/watch?v=vu8X3YroB-w > date created: 2026-09-29 16:31 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: OpenAI fights back >  ## 🧭 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) - [🔄 Refactor appendix](#-refactor-appendix--openai-fights-back) ## 📝 Summary - **What happened.** Theo (channel "Theo - t3․gg") published a 30:58 commentary on 2026-09-29 covering OpenAI's GPT 6.1 model. The video description spells the model "Sol 6.1"; the auto-captions render it "Soul" throughout, and every quote below preserves the caption spelling. The description frames it as OpenAI's answer to Anthropic's Opus 5.5, "like Astra, but less dumb, and way way cheaper." - **The claim.** The host says the model lists at $2 per million input tokens and $10 per million output tokens, cuts cache-read pricing to $0.10 per million (a 95% discount, up from 90%), scores at or near the top of Terminal Bench and Deep SWE, and is far less erratic than GPT6 Astra. The host still keeps Opus 5.5 as the daily driver. - **The verdict.** Strong at code review, audits, computer use and scoped work. Weak at long unattended rewrites, judgment when managing a fleet of machines, and front-end design ("no taste"). The host plans to have Opus call 6.1 Soul for investigation and review, and says the model will effectively replace Sonnet 5.5 for them. - **Why it matters.** Cache-read price dominates agent workloads, so a halved cache price cuts real spend more than the headline token prices suggest. The host also reads the pricing and a subscription-limit change as a sign the "subsidization era" is ending. - **Key stakeholders.** OpenAI (vendor), Anthropic (Opus 5.5, Sonnet 5.5, named comparison), Depot (sponsor), Tibo (OpenAI-side poster of the subscription tweet), testers Julius and Ben, developers paying for agent inference. - **Risks and tradeoffs.** The host has early access and says OpenAI didn't pay for the video but did set the publish timing. Benchmark numbers come from the host's own reruns on "random VMs" with three tasks dropped and one run incomplete. Most figures are spoken or shown on screen only, so they're unverified here. Several quoted numbers are internally inconsistent (see Logical validity). - **Bottom line.** A credible hands-on review with real caveats and no independent benchmark behind it. The durable idea is architectural: price per completed task depends on cache reads and token efficiency, not list price. ## 📖 Main content **The setup.** Anthropic released Opus 5.5 the week before. OpenAI answered with GPT6 Soul and GPT6 Luna, which the host found unremarkable and skipped. GPT 6.1 Soul arrives about a week after 6.0 Soul, and the host argues the timing means it wasn't planned as a point release. Evidence offered: the model is "significantly smarter" than GPT6, its tokens per second are slower (which the host reads as a bigger model, with a stated "hard to know for sure"), and a tweet from Tibo restores a $200 Pro subscription with usage recalculated to net out at about half the API-dollar value of the old plan. **The price argument.** List price is $2 in, $10 out, described as one fifth of Astra and the same as GPT6 Soul. The cache-read cut to $0.10 per million matters more. The host quotes Opus 5.5 (used as a reviewer) saying agent workloads are 96% cache reads. Rerunning the host's own session history at the new prices, Opus estimates $1,550 instead of a $5,700 guess at Opus-like pricing. The host also cites rumored margins as high as 95% on inference and says the subscription change follows from thinner margins. **The behavior argument.** The host's main complaint about Astra was "spikiness": strong peaks with random collapses into very poor output. 6.1 Soul is described as "mostly, not entirely" fixed. Peaks are lower than Astra's, computer use is a little weaker but good enough, and two other testers (Julius, Ben) reportedly agree. The host trusted it to prepare wire transfers from email invoices without an error. **The benchmark argument.** On Terminal Bench and Deep SWE the model's low and high effort settings score near Astra's high setting at far lower cost. The host compares it with Jev Router (a cost-routing product on Open Router, named in the captions "Jev router"), which sent about 60% of requests to Deepseek 4.1 Flash and cost more than Astra on low while taking 20 minutes against 4.6. The host's explanation: a router that can't reason can't judge how hard a task is from a bare prompt. **Where it loses.** A TypeScript-to-Rust compiler port made no progress under 6.1 Soul over about four days. Opus restarted it and finished in roughly a night. The generated Fish Slop game looks better than earlier versions but plays worse, runs at a lower frame rate, and surrounds the game with 20-plus lines of filler UI text. The host calls the model "unacceptably garbage at UI." **Where it wins.** Code review and audits. In one PR review it found regressions Opus and Fable missed. In a later audit it took the Rust port from 83.7% to 100% and found that 1.3 million of 1.8 million lines were unused, because Opus had rewritten the code in a new crate and left the old code behind. **Contested and unverifiable in this source.** Every price, score and cost figure is spoken or read from an on-screen chart, and the host says the benchmark numbers on screen were still changing. Two claims conflict with each other: "73x cheaper" (00:04:22) against "70th the price" (00:20:58), and a $1.38 versus $512 cost gap described as "four to 5x" (00:09:25). The subscription usage figures ($8 to $9,000 on Claude, over $12,000 on Codex) are the host's estimates and the host says "hard to know for sure." The claim that OpenAI may keep less margin than before rests on the host's inference from timing and a tweet, and the host says they're "walking a thin line." Treat all of it as claims. **Sponsor segment.** Depot sponsors 00:01:29 to about 00:03:00. The transcript gives the link as "swive.link/devo"; the description gives https://soydev.link/depot. The Depot claims (CI up to 10 times faster, Docker builds up to 40 times faster) are spoken sponsor copy with no source. ## 🗒️ Notes #### ⏱️ [00:00:04](https://www.youtube.com/watch?v=vu8X3YroB-w&t=4) Recap, early-access disclosure - Opus 5.5 landed the prior week; OpenAI answered with GPT6 Soul and GPT6 Luna, which the host found unimpressive. - GPT 6.1 Soul is the third OpenAI model in about two weeks. The host filmed at 2 a.m. and says the take is unfiltered. - Disclosure: early access, no payment, OpenAI controls "when" but not "what and how." #### ⏱️ [00:01:29](https://www.youtube.com/watch?v=vu8X3YroB-w&t=89) Sponsor: Depot - Sponsor segment. Pitch: faster CI (up to 10x), faster Docker builds (up to 40x), bare-metal AMD Epyc hardware, failure analysis agents can read. These are sponsor claims, unverified. - Host anecdote: switching a project to Bun caused CI failures, and Depot made the failures easier to read. - Caption garbles: "depot metal," "swive.link/devo." Description link is https://soydev.link/depot. #### ⏱️ [00:02:57](https://www.youtube.com/watch?v=vu8X3YroB-w&t=177) Benchmarks and their caveats - Host says the numbers on screen are "unlikely to be exactly accurate" because the host is rerunning Terminal Bench personally: two failed runs, a third in progress, no medium run initially, an incomplete max run backfilled from X high. - Claims Terminal Bench is state of the art for the model, and Deep SWE too, with an odd shape: scores fall at X high and max and stay flat at low and high. - Cost comparison on a log scale: 6.1 Soul on low costs $0.21 (caption: "21"), Astra on low $1.46, Opus 5.5 on max $14.65 for the same score. Host says "73x cheaper," which conflicts with the later "70th the price." #### ⏱️ [00:05:05](https://www.youtube.com/watch?v=vu8X3YroB-w&t=305) Why this isn't a point release - Timing inference: GPT6 Soul shipped a Tuesday, 6.1 Soul ships the next Tuesday. - Two behavioral tells: "significantly smarter" than GPT6, and meaningfully slower tokens per second (read as a bigger model, with uncertainty stated). - Tibo's tweet: the $200 Pro plan returns with usage recalculated to net out at half the API-dollar spend of the old plan. The host connects this to an alleged internal "code red" over Opus 5.5. #### ⏱️ [00:06:31](https://www.youtube.com/watch?v=vu8X3YroB-w&t=391) Pricing and the cache-read cut - Host estimates monthly usage on $200 plans: Claude Code $8 to $9,000, Codex over $12,000, Astra closer to $8,000. All estimates, described as hard to log. - Pricing: $2 per million in, $10 per million out; half the price of 5.6 Soul after discounts (as stated), same as GPT6 Soul, one fifth of Astra. - Cache read: $0.10 per million, a change from OpenAI's usual 10% of the input price. The host says it's the first such cut they know of. > [!quote] [00:07:14] > That makes it a 90% discount and now it's a 95% discount. #### ⏱️ [00:07:58](https://www.youtube.com/watch?v=vu8X3YroB-w&t=478) Margins and subscription math - Rumored inference margins as high as 95%, per the host, "for every $10 you spend, they only have to spend 50" (cents implied). - Host's reading: the cache cut eats margin, so subscription subsidies had to shrink. Details were "left out there for us to reverse engineer." #### ⏱️ [00:08:42](https://www.youtube.com/watch?v=vu8X3YroB-w&t=522) Benchmark setup and the cost gap - The host ran Terminal Bench on random VMs on their own network and dropped three tasks that need an H100. - Most expensive 6.1 Soul run: about $1.38 per task. Cheapest Opus 5.5 run: "$512." The host calls this a four to 5x gap, which the two figures don't support, so the units are unclear. - Host says none of this is real code work and sets up the behavior discussion. #### ⏱️ [00:10:09](https://www.youtube.com/watch?v=vu8X3YroB-w&t=609) Spikiness, Astra vs 6.1 - Astra was smarter than Anthropic's best in the host's view, but spiked into the worst output of the year. The host churned to Fable 5.1 and then Opus 5.5 for daily work. > [!quote] [00:10:09] > It is also dumb. It is smart and dumb at the same time. - 6.1 Soul: "mostly, not entirely" fixed. Lower peaks, slightly below Opus 5.5 on the 3D-capability work, slightly weaker than Astra on computer use. - Anecdote (00:11:33): the model prepared wire transfers for overdue invoices from email, opened a Chrome tab per payee, and got nothing wrong. The host left the send click to themselves. - Testers: Julius called it incredible, Ben called it "incredibly boring." The host counts "boring" as the right place to be. #### ⏱️ [00:12:16](https://www.youtube.com/watch?v=vu8X3YroB-w&t=736) Where it's the default - Weirder behavior at the edge of its capability, so the host still reaches for Opus "a decent bit." - Computer use: good enough that the host wouldn't think twice at the price gap. The host expects a faster "Ultra" mode (caption "Ultraast") to make it six times faster. - Deep code review: the "Rottweiler nature" of OpenAI models, digging until every flaw surfaces. #### ⏱️ [00:13:40](https://www.youtube.com/watch?v=vu8X3YroB-w&t=820) Audit and improvement tasks - Orchestrator v2 audit: near-identical to Astra, slightly higher in one analysis, at about half the price (297 vs 584, units not stated). Opus and Sonnet cost less but were less thorough. - T3 Code improvement task: 6.1 scored 87.4 against Astra's 83.8 and Grok 4.7's 80.7. Costs: Opus 5.5 $5, Sonnet 5.5 almost $9, 6.1 $2.15. - Host's summary: token price under Sonnet's, token use still efficient. #### ⏱️ [00:14:26](https://www.youtube.com/watch?v=vu8X3YroB-w&t=866) Opus as blind reviewer - The host had Opus 5.5 review the model's work under a different name, then edited the history to reveal it. Opus called it frontier tier in a blind 10-model bench (rank and score garbled in captions as "A7.4"). - Cost of that bench: about $7 for 6.1, $15 for Sonnet 5.5, $50 for Opus. (Caption reads "$7 to run versus 15 for Sonic 55 and 50 for Opus.") - Opus's call: incredible for scoped work, would choose it over Astra, about level with Opus for long unattended building, but it follows process rules even when they stall progress and doesn't ask for help. #### ⏱️ [00:15:53](https://www.youtube.com/watch?v=vu8X3YroB-w&t=953) The Rust port and fleet management - A TypeScript-to-Rust compiler port with this model made no progress over about four days and burned many tokens. Opus restarted from scratch and had it working in a day, after months and "hundreds of thousands of dollars" in earlier attempts (host's figure). - Fleet management (the host's machines running agents): a couple of poor judgment calls. The host says Opus also errs there and that Astra made the fewest such mistakes, and adds that they're "done using Astra" after this model. - Verdict for unattended heavy rewrites: Anthropic is "comically far ahead." #### ⏱️ [00:17:19](https://www.youtube.com/watch?v=vu8X3YroB-w&t=1039) Opus guesses the price - Opus guessed $5 in, $0.50 cache, $30 out, then $3, $0.30, $20 as an adoption price. Told the real price, it called it "very aggressive." > [!quote] [00:17:19] > Agent workloads are 96% cash reads. - The claim comes from Opus, not a measurement in the video. #### ⏱️ [00:18:02](https://www.youtube.com/watch?v=vu8X3YroB-w&t=1082) What the cache price changes - Rerun on the host's sessions: $5,700 guessed cost becomes $1,550. Heavy PR reviews: expected up to $10, actual up to $3. - Opus's takeaways: default for scoped work; bloated system prompts barely matter under cache pricing; long loops still a bad idea, but for wasted time, not money; expects the price to rise. - Three rounds of work for about half of one Sonnet round, because average tokens read per request were about 110,000 against 360,000 for Sonnet 5. #### ⏱️ [00:19:29](https://www.youtube.com/watch?v=vu8X3YroB-w&t=1169) Jev Router comparison - The host forked Deep SWE to test a new Open Router product, Jev Router, and calls the idea "incredibly stupid." > [!quote] [00:19:29] > It turns out a model that cannot reason, that is given a prompt and no context, cannot make a good decision around how hard the problem is. - Jev Router sent about 60% of requests to Deepseek 4.1 Flash, cost more than Astra on low, took 20 minutes vs 4.6, and averaged 104 steps vs 19. - 6.1 Soul on low: 4.8 minutes, $0.21, second-highest score the host has seen. Opus 5.5 on max: 50 minutes per task, tying that score. #### ⏱️ [00:21:41](https://www.youtube.com/watch?v=vu8X3YroB-w&t=1301) Recommended setup, Fish Slop begins - If your work fits what 6.1 does well, use it for everything; otherwise keep Opus and let Opus call 6.1 for root-causing, codebase analysis, triage and reviewing Opus's own work. - Fish Slop game demo: high graphic fidelity (fish, submarine, propellers, plants), but the game is wrapped in filler UI and the sound is blaring. #### ⏱️ [00:23:09](https://www.youtube.com/watch?v=vu8X3YroB-w&t=1389) Fish Slop gameplay and UI text - Movement is worse than the Opus and Sonnet versions, frame rate is lower, and the host lists filler strings such as "Take a breather paused" and "A little underwater chaos" (about 20 in all). - The host sent a screenshot to OpenAI and "crashed out in the Slack." #### ⏱️ [00:24:38](https://www.youtube.com/watch?v=vu8X3YroB-w&t=1478) Front-end and 3D verdict - Front-end and design verdict: the model has no taste and regressed again; bring your own. - Strength: Blender via CLI from a screenshot. Weakness: game mechanics feel poor. First test run hit a game-over in 30 seconds with no reason given. - The game cost about $5 to generate; the host expects fast generation to change game development. #### ⏱️ [00:26:03](https://www.youtube.com/watch?v=vu8X3YroB-w&t=1563) Real PR review - The host couldn't use the model on open-source repos during the testing window, only internal projects (Lakebed) and audits. - On a T3 Code PR (a worktree setup window that vanished when leaving a thread), 6.1 found that follow-up messages stayed blocked after the agent started and that recovered setup progress vanished too early. Fable and Opus had missed both. It used code reading plus computer use. - The host says they don't trust the model to write the code they land but do trust it to review. #### ⏱️ [00:27:28](https://www.youtube.com/watch?v=vu8X3YroB-w&t=1648) Opus and Soul in tandem - The host's pattern: Opus implements, Soul reviews and digs into details. Opus is the more pleasant collaborator and doesn't get blocked as often. > [!quote] [00:27:28] > I like using them in tandem because I find Soul to be way better at reviewing and digging into the details - Rust port audit: Soul took it from 83.7% to 100% in under a day after months of stalled work by other models. Critique found 1.3 million of 1.8 million lines unused, because Opus rewrote the code from scratch in another crate and left the old code behind. #### ⏱️ [00:29:35](https://www.youtube.com/watch?v=vu8X3YroB-w&t=1775) Closing verdict - Good enough that the host has almost no reason to use Sonnet now, and plans to configure Claude Code so Opus calls Soul for review. - Not the Opus-dethroning release the host wanted. No plans to cancel Claude subscriptions, more use of Codex subscriptions planned. - Closing prediction: this "potentially mark[s] the start of the end of the subsidization era." Video ends about 00:30:58 with a subscribe request and no promo or link. ## 📚 References The transcript itself contains no links; the following come from the video description or are primary sources for named entities. - [OpenAI fights back — Theo - t3․gg](https://www.youtube.com/watch?v=vu8X3YroB-w) — the source video, 30:58, uploaded 2026-09-29. - [Depot](https://soydev.link/depot) — sponsor link from the description. The transcript renders it "swive.link/devo." - [Sponsor the channel](https://soydev.link/sponsor-me) — description link, not discussed in the transcript. - [t3.gg](https://t3.gg) — channel home page, from the description. - OpenAI GPT 6.1 "Sol" — the description spells it "Sol 6.1"; captions say "Soul." Find the official model page or launch post from OpenAI before citing prices. - Anthropic Opus 5.5 and Sonnet 5.5 — named comparison models. No URLs in the source. - Terminal Bench and Deep SWE — the benchmarks named at 00:03:39 and 00:04:22. The host runs personal reruns, so the on-screen numbers aren't the benchmarks' published leaderboards. - Open Router "Jev Router" — the routing product tested at 00:19:29. The name matches the "Jev" model discussed in other channels, so confirm whether Open Router's router is built on that model before reusing the claim. - Caption garbles worth correcting before reuse: "Codec" is Codex, "cloud code" is Claude Code, "Quad" is Claude, "Grock 4.7" is likely Grok 4.7, "Deep Su" is Deep SWE, "Sonic 55" is Sonnet 5.5, "cash reads" is cache reads, "Ultraast" is unclear, "Fish slop" is the host's game demo name. ## 🗣️ Paraphrase Theo argues that OpenAI's newest model is the first in the GPT6 line that the host is comfortable using every day. It keeps Astra's strength at code review and finding bugs, drops most of Astra's random collapses, and costs a fraction as much. The pricing is the surprise. Input and output prices are one fifth of Astra's, and the cache-read price fell from 10% of input to 5%, which matters most because agent workloads mostly re-read cached context. The host hedges throughout. Benchmark numbers are self-run and still changing. Some are inconsistent. The model loses to Opus on long unattended rewrites, judgment when managing several machines, and any front-end work, where it wraps a good-looking game in filler text and plays badly. The host's practical answer isn't to switch. It's to keep Opus as the main collaborator and have Opus hand investigation and review to the OpenAI model, which found regressions and dead code Opus had missed. The host closes with a warning that the cheap-subscription era may be ending. ## 📰 Newsletter **A cheaper reviewer, not a new boss.** Theo's 31-minute hands-on with OpenAI's GPT 6.1 (spelled "Soul" in the auto-captions, "Sol" in the description) has one takeaway that applies well beyond the model: for agents, cache-read price matters more than the headline token price. The host says OpenAI halved it to $0.10 per million and cites Opus 5.5 estimating that agent workloads are 96% cache reads. The host's own session history dropped from $5,700 to $1,550 under the new prices in that estimate. Where it lands for WordPress work: use it as the second pair of eyes. The host had it review a real pull request and it caught two regressions that two other frontier models missed. Their verdict is "Opus builds, Soul reviews." That maps cleanly onto plugin-review queues, Core patch triage and docs audits, where a thorough, cheap reviewer is worth more than a fast author. Grain of salt: the benchmark scores come from the host's own reruns on mixed hardware, and one figure ("73x cheaper") disagrees with a later one ("70th the price"). Read it as a strong field report, not a leaderboard. Links: [the video](https://www.youtube.com/watch?v=vu8X3YroB-w). ## ✅ Action items #### DevRel - Developer advocacy: write a "builder plus reviewer" demo that pairs two models on one PR, and publish the token and cache accounting per run. - Education: teach that cost per completed task depends on cache reads and tokens per request, using the 110,000 versus 360,000 example, labeled as the host's estimate. - Developer experience: add a "which model reviews, which model writes" section to agent quickstarts, with a config example. - Developer marketing: don't repeat 70x or 73x cheaper in any Pantheon or community material without a benchmark citation. - Community management: when this video circulates, add a note that scores are self-run and prices are the host's, pending OpenAI's published numbers. - Internal/external/cross-project: confirm current list prices and cache-read rates from the vendors' own pages before any internal cost model uses them. #### OSPO - Policy and governance: decide whether AI review output may block a merge, given the host's report that a review caught regressions others missed. - License compliance: check the terms of any new reviewer model for training on submitted code before pointing it at private repos. - Security management: add the model to any code-review threat model. Reviewing untrusted PRs feeds attacker-controlled text into an agent with tool and computer access. - Software inventory: record the reviewer model and version as a dependency, since behavior changed between 6.0 and 6.1 in about a week. - Developer support: publish guidance on when a cheap reviewer is enough and when a human must review anyway. - Community engagement: no direct action from this source; nearest relevance: vendor-subsidized subscription plans that open-source maintainers depend on may shrink. - Strategic alignment: set a KPI for review cost per merged PR so a model swap is measurable. #### Community - Maintainers: pilot an AI reviewer on one repo for 30 days and log real bugs found versus false alarms. - Contributors: document that AI review comments are advisory, and how to contest one. - Sponsors: if a vendor sponsors AI tooling, ask for the cost and quality numbers before and after. - Ecosystem: no direct action from this source; nearest relevance: coordinated disclosure when an AI reviewer finds a vulnerability in a shared dependency. #### WordPress - Core: no direct action; nearest relevance: patch review throughput. - Core Performance: no direct action from this source. - Core AI: add "reviewer versus author" as a separate evaluation category in responsible-AI guardrails, since the host reports opposite strengths. - **Core Program Team:** flag that model behavior and pricing changed within a week, so any AI-assisted workflow needs version pinning and a review date. - Mobile: no direct action from this source. - Playground: build a browser demo that runs two review passes on the same diff and shows the differences. - Community: no direct action from this source. - Training: draft a lesson on reading AI vendor claims, using the 70x versus 73x mismatch as the example. - Marketing: no direct action from this source. - Support: trial an AI reviewer on drafted support replies, measured against human edits. - Design: note the host's finding that the model has "no taste," so design review still needs a human. - Documentation: test an AI audit pass on a docs section and count real errors found. - Photos: no direct action from this source. - TV: no direct action from this source. - Polyglots: no direct action from this source. - Accessibility: run AI review alongside axe on one plugin and record what each catches. The video gives no accessibility data. - Meta: no direct action from this source; nearest relevance: reviewer tooling cost for .org queues. - CLI: no direct action from this source. - Hosting: no direct action from this source; nearest relevance: the sponsor's CI-speed claims are unverified. - Tide: test an AI reviewer against known-labeled findings before trusting its severity calls. - Openverse: no direct action from this source. - Themes: trial AI review on theme submissions with a human gate. - Plugins: trial AI review on plugin submissions with a human gate on anything security-related. #### The WPOCC - Programs: no direct action from this source; nearest relevance: subscription-funded tooling costs may rise. - Projects: no direct action from this source. - Funding experiments: if a microgrant funds AI-assisted review, require published cost per reviewed PR and defect-catch rate. ### FAIR - Technical independence: a hosted, proprietary reviewer is a single-vendor dependency. Document it before adopting. - Policies and covenants: require AI review findings to be logged and auditable, not just the pass or fail outcome. - Productized governance artifacts: add a "model version and pricing date" row to operator scorecards. ## 🧪 Hypotheses 1. On a held-out set of 100 real PRs with known defects, GPT 6.1 finds at least 20% more true defects than Opus 5.5 while costing at least 3x less per review. 2. Cache-read price explains more than 70% of the total cost gap between 6.1 and Astra on agent workloads, and list token price explains less than 30%. 3. The rate of catastrophic single-run failures (task score below half the median) is lower for 6.1 than for Astra across 50 repeated runs of the same tasks. 4. A router that picks models from a bare prompt with no context (the Jev Router case) costs more per solved task than a fixed mid-tier model on Deep SWE. 5. On long unattended rewrites, Opus 5.5 completes more tasks per dollar than 6.1 once retry and restart costs are counted. 6. A two-model pipeline (Opus writes, 6.1 reviews) reduces regressions per merged PR by at least 25% against Opus alone. ## 📊 Methodology - **H1 and H6.** Build a labeled set of 100 merged PRs with later-confirmed regressions. Run each reviewer on the pre-merge diff with identical prompts and record true defects found, false alarms and cost from invoices. Compare the pipeline arm against Opus alone. - **H2.** Log tokens by category (uncached input, cached input, output) for the same 30 agent sessions on both models. Compute cost under each price sheet and attribute the difference. - **H3.** Run 50 repeats of 20 Terminal Bench tasks per model on identical hardware. Define "collapse" before running, and report the distribution, not a mean. - **H4.** Rerun Deep SWE with the router against three fixed models. Report cost, wall time and steps per task, and publish the task list, including any tasks dropped. - **H5.** Give both models the same 10 rewrite tasks with no human help for 48 hours. Count completed tasks and total spend, including restarts. - **Controls.** Fix prompts, effort settings and tools. Record model names and versions with dates, because the host reports a new snapshot within a week. ## ❓ Questions - What are OpenAI's official prices for 6.1, and does the $0.10 cache-read rate apply to all plans? - What benchmark harness and hardware produced the on-screen Terminal Bench and Deep SWE numbers, and which three tasks were dropped? - Which figure is right, "73x cheaper" or "70th the price," and what does "$1.38 versus $512" measure? - How was the "96% cache reads" figure derived, and is it a property of agent workloads generally or of the host's usage? - Does the Pro subscription's halved API-dollar value change effective cost for developers who use agents heavily? - Why did the model stall on the Rust port while Opus finished it? Is it process-rule rigidity, context handling, or task fit? - Is "boring" (Ben's word) a stable property, or did the 6.0-to-6.1 jump come from a one-off post-training run? - What happens to the "cheaper reviewer" pattern if the price rises, as the host says Opus predicts? ## 🔮 Future research - **Effective cost per completed task.** No standard metric folds list price, cache rate, token efficiency and retry cost into one number. A shared harness would make claims like "70th the price" comparable. - **Reviewer versus author specialization.** The host's tandem pattern (one model writes, another reviews) is anecdotal. A controlled study on real repos would test whether the pairing beats either model alone. - **Router failure modes.** The Jev Router result suggests routing without context fails. A test with context-aware routing would separate the idea from that implementation. - **Subsidy dynamics.** How subscription usage caps change after a cheaper model ships is observable from published plan terms and could be tracked over time. - **Agent-run wire transfers.** The host had the model prepare payments. Failure rates and safe-handoff patterns for financial agent tasks are unstudied in this source. ## 💡 Implications - **Practice.** Cost planning for agents should start from cache-read price and tokens per request, not list price. A pipeline that separates writing from reviewing can use a cheaper reviewer. - **Policy.** Teams that let AI review block merges need a written rule for disputing a finding and for logging the model version. - **Security.** A reviewer with tool and computer access reading untrusted PRs is an injection surface. Sandbox it and limit credentials. - **Education.** The video is a compact case study in reading vendor and reviewer claims: self-run benchmarks, a model grading its own competitor, and inconsistent multipliers. - **Funding.** The host predicts the end of heavy subscription subsidies. Open-source projects that rely on subsidized plans should budget for API-rate costs. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid Grade Level on the transcript is about grade 8 to 9. Flesch Reading Ease is near 70 for the spoken prose, but dense jargon (cache reads, Terminal Bench, RLHF-adjacent model names) raises effective comprehension to a working-developer level. #### 👀 Visual learners - Present: on-screen benchmark charts on a log scale, cost tables, and the Fish Slop game demo. - Improve: the charts carry the core numbers but the narration reads only some. Add a written cost table and label the log axis aloud. #### 🔊 Aural learners - Present: continuous narration with spoken numbers and clear contrasts between models. - Improve: the host reads several figures too fast ("297 versus 584" with no units). Repeat units and add a spoken recap of the five key prices. #### 🗣️ Verbal learners - Present: analogies ("Rottweiler nature") and a running story about the Rust port. - Improve: define cache reads, tokens per second and spikiness once, early. #### 🧩 Kinesthetic learners - Present: real tasks (PR review, audits, wire-transfer prep) shown as narrated cases. - Improve: no repository, config snippet or step list for "Opus calls Soul in Claude Code." A copy-paste config would let viewers try it. #### 🧠 Logical learners - Present: cost comparisons, score comparisons and a stated rationale for the recommended split. - Improve: some numbers conflict (73x versus 70th). Show the arithmetic ($14.65 divided by $0.21) so the reader can check. #### 🧑‍🤝‍🧑 Social learners - Present: tester quotes from Julius and Ben, and a request for comments. - Improve: no discussion prompt or link to a community thread comparing results. #### 🧘‍♀️ Solitary learners - Present: self-contained and rewatchable, with a clear verdict at the end. - Improve: add a self-check such as "what share of your agent spend is cache reads?" #### 🌿 Naturalistic learners - Present: no explicit nature ties. - Improve: the host mentions electricity cost for subscriptions at 00:07:58. An energy-per-task comparison would connect efficiency to environmental cost. ## 📊 Accessibility assessment #### Screen reader semantics - Present: YouTube supplies landmarks and player controls, and the description holds the sponsor and channel links as text. - Improve: the benchmark charts and Fish Slop screen are visual-only. WCAG 2.2 AA 1.1.1 (Non-text content) and 1.2.5 (Audio description, prerecorded) are unmet for the charts. #### Keyboard navigation - Present: the YouTube player is keyboard operable with visible focus. - Improve: nothing in the video adds barriers. 2.1.1 and 2.4.7 pass through the platform, not the content. #### Contrast and readability - Present: unverifiable from a transcript. - Improve: chart labels and small terminal text often fail 1.4.3 at default resolution. Zoom into the numbers being discussed. #### Text resizing and reflow - Present: the player scales and captions resize independently. - Improve: burned-in chart and UI text can't resize. 1.4.4 and 1.4.10 are unmet for text inside the frame. #### Media and captions - Present: auto-captions exist, which is how this analysis was produced. - Improve: no chapters, and captions garble product names ("Soul" for Sol, "Codec" for Codex, "cash reads" for cache reads, "Quad" for Claude). That fails 1.2.2 (Captions, prerecorded) in substance. The blaring Fish Slop audio at 00:22:25 is announced but loud, which raises the audio-control concern under 1.4.2. #### Forms and interactions - Present: none in the video beyond a sponsor link and a subscribe request. - Improve: give the sponsor URL as text in the description (it is) and speak it correctly. The captioned "swive.link/devo" doesn't match the description's link. #### Standards reference - Primary gaps against WCAG 2.2 AA: **1.2.2** (inaccurate captions), **1.2.5** (no audio description for charts), **1.1.1** (chart data only visual), **1.4.2** (loud audio without warning), **1.4.4 / 1.4.10** (burned-in text). Corrected captions and chapters are the cheapest fixes. ## 🌍 DEIB assessment #### Cultural breadth - Present: examples come from software teams and an English-language, US-centric developer market. - Improve: add one non-English or non-US use case. The model's language and localization ability goes untested. #### Inclusive language - Present: no exclusionary terms found. The host uses profanity in several places (00:15:53, 00:24:38) and casual hyperbole. - Improve: replace ableist framings such as "dumb" as a model descriptor with "erratic" or "inconsistent," which also carries more information. #### Neurodiversity support - Present: a consistent structure of claim, evidence, caveat, and a clear spoiled verdict at 00:21:41. - Improve: no chapters, and the host says the video is unscripted at 2 a.m. with stumbles. Chapter markers and an on-screen price table would reduce load. #### Representation - Present: one narrator, plus Julius, Ben and Tibo mentioned as testers or posters. - Improve: no third-party voice challenges the host. An independent reviewer would improve balance and credibility. #### Psychological safety - Present: the host discloses uncertainty ("I legitimately don't know if people are going to love it or hate it") and doesn't attack viewers who disagree. - Improve: the fear of being "cancelled" at 00:30:19 frames disagreement as a threat. State the scope of the claims and invite specific counter-tests instead. ## ✅ Logical validity #### Strong reasoning - **Deductive, 00:07:14.** If cache reads are the bulk of agent tokens and the cache-read price halves, then agent cost drops by a large share. The conclusion follows once the 96% figure holds. To clarify: source the 96% from logs, not from Opus. - **Causal, 00:18:02.** Lower average tokens per request (about 110,000 versus 360,000) is offered as the reason 6.1 used under half the input tokens of Sonnet. The mechanism is stated (fewer calls, less output growing the context) and is plausible. - **Comparative, 00:19:29.** The Jev Router test compares cost, time and steps against a fixed model on the same benchmark, three measurable dimensions. The stated reason (no context, no reasoning) is a mechanism, not just an outcome. - **Self-limiting disclosure, 00:03:39 and 00:14:26.** The host admits the benchmark runs were flawed, three tasks were dropped and the max run was incomplete, and that the model's first-draft code had real bugs. These concessions raise credibility. - **Case evidence, 00:26:03 and 00:28:11.** Two real cases (a PR regression, the Rust port audit) show the reviewer finding issues others missed. They're specific and checkable, though single cases. #### Fallacies - **Internal inconsistency, 00:04:22 and 00:20:58.** "73x cheaper" versus "a 70th the price" for the same comparison, and a $1.38-versus-$512 gap called "four to 5x." *Fix:* show the arithmetic and units for each figure. - **Hasty generalization, 00:04:22.** Deep SWE is called "far from a perfect measure" and still anchors the cost claim. One benchmark, self-run, drives the headline. *Fix:* report two independent benchmarks with published harnesses. - **Circular or authority-by-model, 00:14:26 and 00:18:02.** Opus 5.5 grades the competing model and supplies the 96% cache-read figure, then its conclusions are treated as evidence. The host does hedge with "blame Opus." *Fix:* verify the figure from provider invoices and use a neutral judge. - **Speculation from timing, 00:05:05.** "I think it's reasonable to infer that 6.1 Soul was not meant to be 6.1 Soul" rests on release dates, tweet timing and slower token speed. The host flags it as inference. *Fix:* label it as a hypothesis and look for a vendor statement. - **Anecdotal evidence, 00:11:33.** One successful wire-transfer session is offered as a trust signal. *Fix:* report the count of attempts and errors before calling a model trustworthy for payments. - **Undisclosed-interest risk, 00:00:04 and 00:01:29.** The host discloses early access and no payment, and Depot's sponsorship is announced before its segment. OpenAI's control over timing is disclosed. No fallacy, but the two commercial ties (sponsor, early access) shape the framing. *Fix:* keep the same disclosures in the description. - **Ambiguity from captions, throughout.** "Soul" versus "Sol," "cash" versus "cache" and "Jev Router" affect how claims read. *Fix:* corrected captions. #### Tone of voice - Casual, fast and opinionated, with profanity and hyperbole ("comically far ahead," "unacceptably garbage," "absurdly cheap"). The host also flags uncertainty often ("hard to know for sure," "take this as you will"). Superlatives cluster around pricing and the front-end critique, while the review-and-audit section is more specific. It closes with fatigue and worry about reception at 00:30:19. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a response to accusation. This is a product review with no accuser or accusation. For completeness across the three prongs: **Deny** — nothing is denied; the host volunteers flawed runs and front-end failures. **Attack** — the host criticizes products (Jev Router, the model's UI output, Astra) and not people. The host's remark that Opus's price prediction is "hilarious" is mild ridicule of a model, not a person. **Reverse victim/offender** — the closing worry about being "cancelled online" (00:30:19) is a pre-emptive fear, not a claim of victimhood, and no one has accused the host of anything in the source. ## 🎯 Conclusion The model in the video is presented as a cheaper, steadier reviewer and audit tool, not as a replacement for Opus 5.5. Its case rests on two things that hold up as mechanisms: a lower cache-read price applied to workloads that are mostly cache reads, and fewer tokens per request. The host's own experiments (a PR review that caught two regressions, a Rust port audit that found 1.3 million unused lines) support the review claim, and the host's failures (a stalled port, poor UI) mark the limits. The numbers don't survive scrutiny in this source. Benchmarks are self-run with dropped tasks, the "70th" and "73x" figures disagree, key figures came from a model acting as reviewer, and everything is spoken or on-screen only. Depot sponsors part of the video, and the host has early access with OpenAI setting the publish time. **Next steps for readers:** pull vendor prices from official pages, then measure your own split of uncached input, cached input and output tokens before believing any cost multiplier. If you use a writer-and-reviewer pattern, test it on 20 of your own merged PRs first. Confirm the model name spelling ("Sol" per the description) before citing the video. ## 🔄 Refactor Appendix — OpenAI fights back > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 2 Permanent. Resonance filter applied — the fish-slop UI complaints and the subscription-margin speculation were dropped as low-evidence or low-reuse. > > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] Cache-read discount drives agent cost > [!quote] [00:07:14] > That makes it a 90% discount and now it's a 95% discount. **Context & use**: - 00:07:14, right after the host says OpenAI never changed cache-read pricing before. - The host later says agent workloads are mostly cache reads, so this line, not the $2 and $10 headline, drives real spend. The share figure itself is unverified. **Tags**: #xref #articleseed  #### [H2] Agent workloads are mostly cache reads > [!quote] [00:17:19] > Agent workloads are 96% cash reads. **Context & use**: - 00:17:19, quoted from Opus 5.5 while the host asked it to price the new model. - Source is a model's estimate, not a log. Treat as a hypothesis to check against provider invoices. **Tags**: #xref  #### [H3] Erratic peaks made a smarter model unusable > [!quote] [00:10:09] > It is also dumb. It is smart and dumb at the same time. **Context & use**: - 00:10:09, describing Astra and why the host stopped using it daily. - Spikiness, meaning variance in output quality, is the host's stated reason for churn, more than average capability. **Tags**: #xref #aphorisms  #### [H4] A cheap reviewer pairs with a stronger builder > [!quote] [00:27:28] > I like using them in tandem because I find Soul to be way better at reviewing and digging into the details **Context & use**: - 00:27:28, the host's summary of using Opus to implement and 6.1 Soul to review. - Backed by two cases (00:26:03 PR review, 00:28:11 Rust audit), not by a controlled test. **Tags**: #xref #share  #### [H5] Context-free routers can't judge task difficulty > [!quote] [00:19:29] > It turns out a model that cannot reason, that is given a prompt and no context, cannot make a good decision around how hard the problem is. **Context & use**: - 00:19:29, explaining why the Jev Router sent about 60% of requests to Deepseek 4.1 Flash and cost more than Astra on low. - Links to the prior Firecrawl video on Jev as a classifier: routing needs context, and a bare-prompt classifier may fail here. **Tags**: #xref #articleseed  ### 🧠 Atomic Split Pad #### [P1] Cost per task follows cache reads, not list price **Claim** (one sentence, Courtney's voice): When an agent re-reads the same context on each call, the cache-read rate and tokens per request set the bill more than the headline input and output prices do. **Evidence pointer**: [[#H1]] · [[#H2]] **Scope & conditions**: Holds for long-running agent sessions with large reused context. Fails for one-shot prompts, for workloads with little reuse, and for vendors that don't cache. The 96% figure is an unverified model estimate, so measure your own token mix first. **Links**: - [[AI cost architecture]] — because extends - [[Reading vendor benchmarks]] — because contrasts on where the price claim is measured — headline price versus real session spend - [[Agent workflow cost review]] — because applies-in reviewing my own tooling bills  #### [P2] Use one model to build and another to review **Claim** (one sentence, Courtney's voice): A strong implementer and a thorough reviewer are different jobs, so pair a model good at unblocked building with a cheaper model good at digging into details. **Evidence pointer**: [[#H4]] **Scope & conditions**: Supported by two anecdotes from one developer. It might not hold when the reviewer can't read the full context or when both models share the same blind spots. Test on your own merged PRs before trusting it. **Links**: - [[Code review practice]] — because extends - [[Single-model agent loops]] — because contrasts on separation of writing and checking - [[Plugin review queue]] — because applies-in triage of submissions with a human gate  ### ✏️ Fleeting Queue - [F1] Confirm the model's real spelling ("Sol" in the description, "Soul" in captions) and OpenAI's official prices before citing this video — https://www.youtube.com/watch?v=vu8X3YroB-w #fleeting #inbox - [F2] Reconcile "73x cheaper" (00:04:22) with "a 70th the price" (00:20:58): $14.65 divided by $0.21 is about 70 #fleeting #inbox - [F3] Open question: is "Jev Router" on Open Router the same model as the "Jev" from the earlier Firecrawl analysis? Check before linking the two notes #fleeting #inbox - [F4] Measure my own token mix (uncached, cached, output) on one agent workflow to test the 96% cache-read claim #fleeting #inbox - [F5] Article seed: "Cost per completed task, not list price" as a standard for comparing model prices #fleeting #inbox #articleseed ### → Things 3 (max 2 actions) - [ ] Pull my own agent token logs and compute the cache-read share of spend — AI cost review #things - [ ] Verify GPT 6.1 "Sol" prices and cache-read rate on OpenAI's official page before citing the video — research, 15 min #things  📄 **Full text:** [[Full Document Contents/Articles/OpenAI Fights Back.md|OpenAI Fights Back by Theo - t3․gg]] ## Highlights - [00:07:14] cheaper than 5.6 Soul was at launch. Half the price of 5.6 Soul after discounts and the same price as GPT6 Soul. 1/5 the price of Astra. However, this is not the whole story cuz cash reads matter. And the cash read price for this model is going to be 10 cents per mill in. That's a big deal. OpenAI has not changed cash read price as far as I know ever before. It's always been exactly 10% of the normal read price. That makes it a 90% discount and now it's a 95% discount. That means they cut the cache read cost in half, massively reducing the cost for real world agentic use, which to be clear is what we're using these for most of the time. So this makes the model absurdly cheap for [00:07:58] doing realworld code work. That also means that they are almost certainly cutting into their margins. Historically, these margins are rumored to be as high as 95%. Like for every $10 you spend, they only have to spend 50. And as crazy as that sounds, it makes a lot of sense. especially when you consider how expensive it is to make and train these models. But that also gives them wiggle room to change things around a bit, which appears to be what's happening here. That also means that if they were to keep subsidizing the same level that they were on the subscriptions that your electricity cost for your sub would be more than you're paying. So, I get why they have to change this. They've kind of just left the details out there for us to uh reverse engineer. So, uh take this as you will. 6.1 coming so fast seems to ([View Highlight](https://read.readwise.io/read/01m3qebr1eqbsqejf7e5p22dgd)) - Note: [H1] Cache-read discount drives agent cost - Tags: atom/literature/h1 - [00:10:09] I know I cite this diagram a lot, but there's a reason for it. The thing that made me so frustrated with GBD6 Astra wasn't that it was less intelligent than the best models from Anthropic, because it was more intelligent than the best models from Anthropic, and I would argue in many ways still is. But there is a problem. It is also dumb. It is smart and dumb at the same time. GB6 Astra would just randomly spike into the dumbest I've seen a model do this year. Even worse than like some of the small openweight models I play with. It's still so deeply frustrating that Astra does this because on the other end when it does well, it's unbelievable. But these spikes got to the point where I effectively churned. I was only using my codec subs for computer use and I ended up just leaning on to Fable 5.1 [00:10:51] and obviously now Opus 5.5 for almost all of my day-to-day work. So, have they addressed the spikiness? Has GBD 6.1 Soul fixed the problems that I was so frustrated about with Astra? I would say mostly, not entirely, but for the most part, yeah, this is a much better model. Its peaks are not as high. This is not the incredible revolutionary 3D capabilities that we saw with Astra. In fact, I would put it slightly below 5.5 opus in most of those types of things. It is not as good at computer use as Astra, although it is close enough to the point where I have been happy using it for all of my day-to-day work. I actually had 6.1 Soul go through all of my emails and find invoices that I had ([View Highlight](https://read.readwise.io/read/01m3qebwpz5nq6ekwjk1yk9dtk)) - Note: [H3] Erratic peaks made a smarter model unusable - Tags: atom/literature/h3 - [00:17:19] because it is quite fast. And it's not a pro model, which is where it expects those higher like $100 out tiered pricing things to come. Pro models aren't really a thing anymore. We just use Fable and Aster, but you get the idea. This is the funniest part of the whole thread, though. If OpenAI wants people to adopt it, I would expect $3 per mill in, 30 cents for cash, and $20 instead. That would still be a fair price for what it does. To which I responded, if I told you it was $2 in, $10 out, and 10 cents per mill cash read, what would you think? I'd call that very aggressive pricing. For how you use it, it costs about a quarter of what I guessed. The cash price does most of the work. Agent workloads are 96% cash reads. So 10 cents for cash reads matters more than the $2 and $10 [00:18:02] headline prices. When it looked at all of my sessions, its price guess would have been $5,700. But after looking at these new prices, it redid the math and it would have been $1,550. That is a massive decrease. And for all my PR review type tasks, it was expecting those to be up to $10. And it's actually only up to $3. And that's for like heavy PRs with tens of thousands of lines of code. According to Opus, so don't blame me, blame Opus for saying this. First off, Opus says it becomes the default model for scoped work. Second off, it says that bloated system prompts barely matter anymore because of the cash pricing. It's just noise. Third, it says long loops are still a bad idea, but not because of money. It's cuz according to it, the TS ROSport wasted 4 days and made no ([View Highlight](https://read.readwise.io/read/01m3qebtd3whn0swtkjc7tnbnc)) - Note: [H2] Agent workloads are mostly cache reads - Tags: atom/literature/h2 - [00:18:46] progress at all. And Opus even said they'd be suspicious of it lasting. They expect this price to go up in the future. I cannot fathom OpenAI ever increasing the price for a model, but Opus thinking they will is hilarious and shows just how good a value the model is. I love this call out here. GBD 6.1 Soul did three rounds of work for about half the cost of Sonnet's single round. A lot of this comes down to how context was managed, both because 6.1 soul is much more efficient, so it's not doing as many calls that bloat the context. It's not outputting as many tokens that are like building up over time. So the average number of tokens being read per request was only around 110,000 tokens versus 360,000 for Sonnet 5. The result is that Solot used under half as many [00:19:29] input tokens as Sonnet, making this model significantly more efficient. Speaking of efficiency, I want to talk about these deep SWE scores a tiny bit more because this is a weird bench for me to have forked and include in these things. I actually did it for a different reason, not to compare against 6.1 soul, but to compare against a new release from Open Router, Jev Router. Open Router added Jev router to try and optimize costs with your requests. And I thought it was an incredibly stupid idea. Once I started running it against benchmarks, I confirmed it's an incredibly stupid idea. It turns out a model that cannot reason, that is given a prompt and no context, cannot make a good decision around how hard the problem is. and Jev router ended up being Deepseek v4.1 flash router for the ([View Highlight](https://read.readwise.io/read/01m3qec1jdhthwzhjfxk65ynep)) - Note: [H5] Context-free routers can't judge task difficulty - Tags: atom/literature/h3 - [00:27:28] that again over to Claude and it eventually got it good enough and then I finally merged. But that is what I like this model for, and I cannot wait to push its limits for actually coding. Although, I will say from the code that I did have the misfortune of reading, it is harder to justify merging this code than it is for code from Opus. Normally, I would make you guys wait for the Opus versus Soul video or the Sonnet versus Soul video, but I'll just spoil the details now. I like using them in tandem because I find Soul to be way better at reviewing and digging into the details, but I find Opus a more pleasant collaborator and significantly better at actually implementing code without getting blocked constantly throughout its work. And even now with the Rust rewrite of TypeScript, I find myself in a similar pattern where I have Opus 5.5 [00:28:11] just going and going and going, making the codebase work and work well. And then I had Soul come in and do an audit. And this is the funniest part. Remember before I said that I had Soul and Astra working on that TS Rust port for effectively months. I told Soul to come in and it got it from 83.7% to 100% in under a day. I was blown away by that that it had somehow unblocked the work that Astra and Soul were doing as well as 6.1 Soul. And I was absolutely blown away by that, that it had taken the work that 56 soul, 61 soul, and Astra had done over months and got it unblocked where it had been stuck for weeks and finished it. I was much more blown away when I had 61 soul take a look at that work and critique it. And what it ([View Highlight](https://read.readwise.io/read/01m3qebz0jr3pxthhr52w16h5h)) - Note: [H4] A cheap reviewer pairs with a stronger builder - Tags: atom/literature/h3