Anthropic Actually Fixed Opus  🔗 **Source:** [Anthropic Actually Fixed Opus](https://www.youtube.com/watch?v=jgGyX7MPPVg) > [!note]+ Document Notes > tags: [claude-opus, coding-agents, reasoning-effort, inference-cost, model-benchmarks, zero-data-retention] > author: Theo Browne > source: Theo - t3.gg (YouTube) > reference: https://www.youtube.com/watch?v=jgGyX7MPPVg > date created: 2026-09-23 19:03 > > > [!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 Do not quote or attribute any other line to the named author.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: Anthropic Actually Fixed Opus > <!-- deep-read:sections start --> ## 🧭 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--anthropic-actually-fixed-opus) ## 📝 Summary - **What happened.** Theo Browne (Theo - t3.gg) published a 40:14 first-day review of Anthropic's Opus 5.5 on 2026-09-23, the same two-day window in which he says Grok 4.7 and OpenAI's GPT-6 Soul and Luna also shipped (00:02:55). The video has no chapters. Two sponsor reads (WorkOS at 00:01:50-00:03:20, Parallel at 00:23:10-00:24:22) sit inside the runtime and are excluded from the atoms below. - **The claim.** Opus 5.5 repairs the damage Opus 5 did to Theo's trust in the Opus line. He reports a full day of real work, including a merged PR, on roughly 20% of one weekly Claude limit (00:19:23), code he is "not scared to look at, much less merge" (00:39:38), and prose that let him delete his anti-slop instructions from CLAUDE.md (00:10:02). - **The measured part.** Anthropic's launch post, as Theo reads it: Fable 5.1-level performance on most work at 40% lower cost than Opus 5, $4 per million input and $20 per million output tokens (20% cut), cache reads at $0.20 per million (60% cut), about 30% faster, and a tester's 680,000-line migration finished in under a day (00:03:37-00:05:05). Artificial Analysis data, as Theo reads it: a 5-point Intelligence Index lead over Fable 5.1, but nearly 120,000 output tokens per task on max effort against 73,000 for Opus 5 and 27,000 for GPT-6 Astra (00:12:54-00:14:22). - **The anecdotal part.** The 20% weekly-usage figure, a 1%-of-weekly PR workflow (00:20:04), a 6-hour-30-minute max-effort loop that wrote half a markdown plan (00:17:15), roughly five PRs' worth of code-quality impressions (00:36:47), and a 120 FPS Three.js fish game (00:30:16-00:33:13). Theo labels these as day-one observations himself. - **Why it matters.** For anyone running Claude-based coding agents on a subscription, the operational findings are the payload: max effort costs more than 10x the tokens of xhigh for a 1% score gain on Theo's SkateBench (00:18:40); effort level no longer lives in the system prompt, so changing it mid-session keeps the cache (00:12:54); and the model shows context-window paranoia in long runs that no benchmark reaches (00:33:13-00:35:21). - **Key stakeholders.** Anthropic; OpenAI (Codex subscription, GPT-6 Astra, Soul, Luna); xAI (Grok 4.7); Xiaomi (MiMo V2.6); Artificial Analysis as the benchmark aggregator; enterprise buyers gated by zero-data-retention policy; subscription users choosing between a $200 Claude plan and Codex plans. - **Risks and tradeoffs.** One day of testing. Theo says so at 00:36:47 and again at 00:37:30: the model is smaller and will fail more often than Fable on long jobs. The auto-captions garble nearly every product name (Grok as "Gro," Codex as "Codeex" and "codec," Claude Code as "Cloud Code," cache as "cash," Three.js as "3JS"). Several numbers he reads off charts arrive without units in the caption ("$763," "$134," "326"). - **Bottom line.** The sourced numbers support a narrower claim than the title: Opus 5.5 on medium or high effort is the cheapest model per unit of Artificial Analysis score above 50 (00:15:05), and its ZDR eligibility makes it the default enterprise coding model regardless of benchmarks (00:10:02). Whether it holds up past day one on long agentic runs is the open question Theo leaves open. ## 📖 Main content **The arc.** Theo frames the release against a ten-month history. Opus 4.5 arrived in November of the previous year and changed how he wrote code with AI (00:00:00). Opus 5 and Sonnet 5 were, in his words, "iffy at best." He admits he "fell for Opus 5" before its prose slop and dropped requirements made it "more like a liability" (00:00:43). Meanwhile Anthropic's Fable line became his daily driver; he holds five Claude subscriptions because of Fable 5.1 (00:18:40). The question the video asks is whether Opus 5.5 is good enough to spend the Opus half of those subscriptions on. **What Anthropic claims, per Theo's reading of the launch post (00:03:37-00:05:47).** Fable 5.1-level results on most work. 40% cheaper than Opus 5 at default settings on typical workloads. $4 per million input tokens and $20 per million output, a 20% cut. Cache reads at $0.20 per million, a 60% cut. About 30% faster. A tester migrated 680,000 lines of code in under a day. Writing is "clearer and easier to follow"; one tester's quote, "It writes the way I do." Sonnet 5.5 and Haiku 5.5 follow "in the coming weeks," which Theo notes would be the first Haiku release in a year. These are vendor numbers. Theo says he trusts Anthropic's safety claims by default (00:04:20) and does not independently verify the 40% figure, though his own usage observations point the same direction. **What third parties measured, per Theo's reading of the charts (00:05:47-00:08:37, 00:12:54-00:17:15).** On Artificial Analysis's Intelligence Index, Opus 5.5 leads Fable 5.1 by 5 points; Theo notes that gap is about the size of the gap between GPT-6 Astra and a model the captions call "Muse Spark," and reads that as evidence the index compresses real differences (00:13:38). On Terminal Bench 4, Opus 5.5 at medium effort scores above Fable 5.1 at max, at $2.94 per run against $20 (00:06:29). On Cursor Bench, which no longer includes OpenAI models because OpenAI "banned Cursor from using their models," medium again beats Fable 5.1 max (00:07:12). Opus 5.5 at high effort scores 53 on Artificial Analysis against 50.9 for GPT-6 Astra at high, at a higher price (00:07:56). Output tokens per task on max effort: Opus 5 73,000, Fable 78,000, Opus 5.5 almost 120,000, GPT-6 Astra 27,000 (00:13:38). Cost per Artificial Analysis run on max: Opus 5 $5.86, Opus 5.5 almost $6, Fable $7.63 (the caption reads "$763"). Artificial Analysis's own note, as Theo relays it: the model would have been 80% more expensive per run without the price and cache cuts (00:16:32). Opus 5.5 at medium: $1.34 per task against $3.26 for Astra and $6 for Opus 5.5 at max, and the model sweeps the Pareto frontier for every score above 50 (00:15:05). **What Theo measured himself.** SkateBench, his own benchmark: Opus 5.5 scored worse than Opus 5 (00:17:57). Reasoning tokens per question rose from about 330 on xhigh to almost 5,000 on max, a more-than-10x jump, for a 1% score gain (00:18:40). His code-review bench, which asks models to read a large codebase and propose real improvements: Grok 4.7 and GPT-6 Soul lead, Opus and Fable trail, but Opus 5.5 nearly doubled Opus 5's accuracy and reliability (00:21:31-00:22:15). Usage: after a day of deliberate heavy use across visualizer builds, hosting ports, computer-use verification, and a PR, one account showed about 40% of its weekly limit, half of which he attributes to Fable use over the prior two days, so about 20% net (00:19:23). One end-to-end task (find a repo on his machine, update a stale clone, create a worktree, fix a bug, file the PR, verify, merge) cost 1% of a weekly limit (00:20:04). Code quality across roughly five medium PRs: similar-to-lower rates of high-severity issues than Fable 5.1, similar-to-slightly-higher rates of nitpicks (00:36:47). **Anecdotes and demos.** A 6-hour-30-minute thread on max effort produced half a markdown plan before he killed it (00:17:15). Front-end output with the official Claude Code design skill looks like Tailwind templates or pastel-with-shadows; Fable 5.1 remains better at "subtle design taste" (00:24:22-00:25:52). Mia, a tester Theo says often disagrees with him, ran 100 HTML page generations and called it the best model she has tested (00:25:52). Third-party 3D demos: a Dark Souls clone and a flight simulator from Alex (00:28:02), a Tony Hawk Pro Skater clone in C++ with no engine from a tester the captions call "Bjan Bowen" (00:28:45), and Theo's own "fish slop" Three.js game running at 120 FPS with a rigged stingray, a working gameplay loop, and a cinematic camera (00:30:16-00:33:13). He notes the model is stronger in Three.js than in native code (00:29:34). **The quirk that matters for agent operators.** In a long run where one Claude Code instance had SSH'd into another machine to drive a second instance, the model reported that its context was "10% from autocompact" and that "if it crashed that work would be lost" (00:34:38). Theo's explanation: models trained on data that fit within a context window learned to fear the limit, OpenAI post-trained that fear out and caps context at 272K, Anthropic leans on a million-token window instead (00:33:13-00:33:56). His methodological point is the sharper one: most benchmarks finish before compaction, so this failure mode is invisible in published numbers (00:35:21). **Two structural claims.** First, zero data retention. The Fable line does not offer ZDR; Opus does. Companies that require it could not enable Fable for employees, which is why Theo expects Opus 5.5 to become the most-used coding model on that detail alone (00:10:02-00:10:45). He gives no source for "number one model for businesses." Second, effort level has moved out of the system prompt, so changing reasoning level mid-session no longer invalidates the prompt cache (00:12:54). He credits a viewer named Lydia for the catch. **Caption garbles to correct before citing.** "Gro 4.7," "Grock 47," "Rock" are Grok 4.7. "GBD6" and "GPT6" are GPT-6; Astra, Soul, and Luna are the variant names as spoken. "Sonet 55," "Opus 55," "Haiku 55" are 5.5. "Codeex" and "codec" are Codex. "Cloud Code," "CloudMD," "cloud subs" are Claude Code, CLAUDE.md, Claude subscriptions. "cash" is cache. "3JS" is Three.js. "O" and "off" are auth; "Octa" is Okta. "OMD" is WorkOS's agent-auth standard, and the real name is not recoverable from the transcript. "Xiai's" is Xiaomi, maker of the MiMo line. "Bjan Bowen" / "Bjon" is unverified, likely Björn. "Deep Suite" and "Lakebed my cloud" are unresolved (a benchmark and a repo name). "$763," "$134," "326" are $7.63, $1.34, $3.26. "5oint" is 5-point. "atv.link/workos" and "soyv.link/parallel" are soydev.link per the description. "Gemini 38 Flash" is Gemini 3.8 Flash. "Bards" at 00:39:38 is a sign-off garble. **Contested or unverifiable in this source.** Every benchmark number is read off a chart, so the transcript alone cannot confirm the digits. The ZDR-drives-enterprise-adoption argument is a mechanism without a market-share source. The Grok code-review strength is attributed to Cursor's Bugbot data (00:21:31), which Theo labels a suspicion. The claim that OpenAI "entirely beat this out of the models" (00:33:56) is inferred from behavior, not from any published training detail. The title's verb, "fixed," rests on one day. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=0) The setup: Opus 4.5 to Opus 5 - Opus 4.5 shipped in November of the prior year, about ten months before this video, and changed how Theo codes with AI. It was priced above Sonnet and skipped version numbers. - Opus 5 and Sonnet 5 were "iffy at best"; the Fable line was the bright spot. - Theo owns the error: he liked Opus 5 at first, then found prose slop and code that missed requirements. > [!quote] [00:00:43] > the model felt more like a liability than an actual thing you could rely on when you worked every day. #### ⏱️ [00:01:27](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=87) Day-one verdict, WorkOS sponsor, four model drops - Version is 5.5, not 5.1, which he reads as Anthropic setting a high bar. One day in: cheaper, better writing, better at code and instruction-following. - Sponsor read for WorkOS runs 00:01:50-00:03:20. Theo mentions being "down a hand" and that Codex computer-use failed to sign him into dashboards because sites lack agent-auth. Skipped in the atoms. - Four major model drops in two days: Grok 4.7 ("Gro 4.7"), Opus 5.5, GPT-6 Soul and Luna. #### ⏱️ [00:03:37](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=217) Anthropic's launch claims - Vendor claims as read: Fable 5.1-level on most work, 40% cheaper than Opus 5, first release since Anthropic "called for pacing the frontier" (Theo defers comment to a future video). - A tester completed a 680,000-line migration in under a day. Theo reports similar results on performance optimization work. - Pricing: $4 per million in, $20 per million out (20% cut). Cache reads $0.20 per million (60% cut). About 30% faster, which matches his numbers. - Communication is his favorite change. Tester quote relayed: "It writes the way I do." Anthropic frames readability as a safety benefit. > [!quote] [00:04:20] > Our tests show that at default settings, it will cost 40% less than Opus 5 on typical workloads. #### ⏱️ [00:05:47](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=347) Benchmarks and effort levels - Sonnet 5.5 and Haiku 5.5 due in weeks; first Haiku in a year. Theo: Anthropic is "getting absolutely destroyed" on small models but leads on large ones. - Tops every listed bench except agentic scientific research, terminal bench science, and business workflow automation. A bench the captions call "Deep Suite" has stale data. - Terminal Bench 4: Opus 5.5 medium beats Fable 5.1 max at $2.94 vs $20 per run. Frontier Code: new high on medium and max, with medium the best score. - Cursor Bench excludes OpenAI models ("OpenAI has banned Cursor from using their models"). Medium again beats Fable 5.1 max. - Theo rebuilt the Artificial Analysis chart himself, using Opus 5.5 to scrape and render it. Opus 5.5 high: 53 vs Astra high 50.9, at higher cost. Low effort is "really bad"; medium, high, xhigh are good. > [!quote] [00:08:37] > It feels like low and max are dangerous and generally best avoided, but medium, high, XH high, those all have seemed really good #### ⏱️ [00:09:20](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=560) Writing quality, ZDR, and Anthropic's tone shift - Anthropic's side-by-side: Opus 5's answer used three em dashes and misidentified a billing bug; Opus 5.5 split a $9.92 drop into $1.50 from a free-tier change and the rest from a commit labeled "no behavior change." - Theo deleted his anti-slop instructions from CLAUDE.md ("CloudMD") to test the model's natural voice and doesn't plan to restore them. - ZDR: the Fable line has no zero-data-retention option; Opus does. This, more than price, keeps Opus the enterprise default. - Anthropic tweet relayed: "Opus 5.5 is way, way, way better than Opus 5. Sorry about that model. Please try this one." Theo reads it as an admission and a sign employees can now post freely. - He pre-empts the "fanboy" label by citing his history of criticizing Anthropic's business practices (00:11:28). Teaser of Alex's Dark Souls clone (00:12:10). > [!quote] [00:10:02] > more so because the ZDR policy can be applied on Opus and it couldn't be applied on Fable and Mythos. #### ⏱️ [00:12:54](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=774) Cache-safe effort changes and token inefficiency - Effort level used to sit at the top of the thread inside the system prompt, so changing it invalidated the cache. Opus 5.5 moves it out. Credit to Lydia. - Artificial Analysis Intelligence Index: 5-point lead over Fable 5.1. Theo notes the same-sized gap separates Astra from Grok 4.7, and Astra from "Muse Spark," and reads this as a flaw in the index. - Output tokens per task on max: Opus 5 73k, Fable 78k, Opus 5.5 almost 120k, GPT-6 Astra 27k. A 4x gap against Astra. - Cost per run on max: Opus 5 $5.86, Opus 5.5 almost $6, Fable $7.63 ("$763"). The price cut, not efficiency, keeps 5.5 from being the most expensive model Artificial Analysis has run. > [!quote] [00:13:38] > The price changes are real. This model is cheaper, but it's not because it's more efficient. #### ⏱️ [00:15:05](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=905) Effort levels, Pareto frontier, MiMo - Adding low/medium/high to the chart: Opus 5.5 xhigh is near Astra's cost; high is about half Astra max; medium is $1.34 ("$134") vs $3.26 ("326") for Astra and $6 for Opus 5.5 max. - Opus 5.5 holds the Pareto frontier for every score above 50. Low effort omitted as "pretty stupid" for the price. - Grok 4.7 and 4.6 sit below the frontier at higher cost. MiMo V2.6 (Xiaomi, "Xiai's") performs well for its price and fills the gap before the cheap Luna models. - Artificial Analysis: the run would have cost 80% more without the price and cache-read cuts. Edwin's claim that medium equals Fable 5.1 intelligence is, in Theo's view, "a bit of a reach," though $1.34 vs $6+ per task stands. #### ⏱️ [00:17:15](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=1035) Max effort loops, SkateBench, weekly usage - A max-effort thread ran 6 hours 30 minutes writing a markdown plan, looped, and was half done when he switched it to xhigh. - SkateBench: Theo corrects himself live; Opus 5.5 scored worse than Opus 5. Reasoning tokens per question: about 330 on xhigh, almost 5,000 on max. 1% score gain for more than 10x the cost. - Five Claude subscriptions. After a day of deliberate heavy use, one account shows about 40% weekly, half attributable to Fable over the prior two days, so about 20% net. Desktop app use landed on a second account. - Real work done: the Artificial Analysis visualizer, multiple hosting ports, computer-use verification and data extraction. > [!quote] [00:18:40] > That's over a 10x gap for that tiny little jump. And you get almost no improvement in score here at all. #### ⏱️ [00:20:47](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=1247) Code-review bench weakness, Parallel sponsor - The PR workflow (find repo, update stale clone, worktree, fix, PR, verify, merge) cost 1% of a weekly limit (00:20:04). Codex plans burn faster; Astra is "capable but spiky and way too expensive." - Theo's code-review bench: Grok 4.7 and Soul lead at deep codebase review; his suspicion is that Cursor's Bugbot review data explains Grok's strength. Opus 5.5 nearly doubled Opus 5's accuracy. Gemini 3.8 Flash is "meme tier." - He discounts the weakness: deep whole-codebase review is a once-every-two-weeks task for him, not the daily case. - Sponsor read for Parallel runs 00:23:10-00:24:22 (search in about 1 second vs 6+ for OpenAI's endpoint, 5,000 free requests a month, $80 signup credit). Skipped in the atoms. #### ⏱️ [00:24:22](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=1462) Front-end design and third-party demos - Design instincts are present but not expressed. Output with the official Claude Code design skill is "a little sloppy": Tailwind templates or pastel-with-big-shadows. Without the skill, slightly better. Fable 5.1 keeps the edge on animations, taste, and layout. - Recommendation: Opus as default, Fable for marketing pages and design overhauls. - Mia's 100 HTML generations: "the best model that she's ever tested." Theo shows a paper-style skeuomorphic layout, a neon-rain terminal theme, and a chess game. - Max ported T3 Code into Minecraft to watch agents from inside the game. Gabriel, formerly a Sora lead at OpenAI, reversed his wait-a-week stance within hours: "Never mind. Opus 55 is crazy good." - Alex's Dark Souls clone and flight simulator; a 2D/2.5D animation with AI-generated music Theo calls "too cringe"; "Bjan Bowen" (early access) built a Tony Hawk Pro Skater clone in C++ with no engine and an FPS in Three.js. Theo: stronger in Three.js than native. #### ⏱️ [00:30:16](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=1816) Fish slop 3D demo - First open was the Grok 4.7 version by mistake. The Opus 5.5 version runs at a "perfectly smooth 120 FPS" with solid models and movement. - He had left it on medium, the default. A second pass on xhigh with "make it as visually appealing as possible" produced a rigged, animated stingray, a working gameplay loop with feeding and growth, a cinematic camera, and an alien encounter. - Lighting blends wrong on wing movement; fish transparency breaks near the ceiling (00:36:04). First version of this demo he wanted to keep playing. > [!quote] [00:31:02] > I really thought OpenAI would maintain their lead for a while here, but look at the stingray. #### ⏱️ [00:33:13](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=1993) Context-window paranoia - Mechanism proposed: pre-compaction training data fit inside the context window, so models learned to fear the limit. OpenAI post-trained this out and caps at 272K; Anthropic leans on the 1M window. - Setup: Claude Code on Theo's machine SSH'd into another Mac on his network to drive a second Claude Code instance doing performance work on the fish game. - The report said the worker was slow to commit after three hours, its context was "10% from autocompact," and "if it crashed that work would be lost." Theo: none of that should concern the model; compaction works and is faster now. - The pattern appears when he audits a stalled run. Benchmarks rarely reach compaction, so the failure never shows in scores. He says Fable would not say this; Astra might. > [!quote] [00:35:21] > The vast majority of benchmarks don't even hit compaction by the time they're done running, which is worth noting here. #### ⏱️ [00:36:04](https://www.youtube.com/watch?v=jgGyX7MPPVg&t=2164) Spikiness, PR quality, subscription value, close - Across about five medium PRs: similar-to-lower rates of egregious high-severity issues vs Fable 5.1, similar-to-slightly-higher rates of small issues and nitpicks. He calls this insufficient data. - Gut read against his earlier Fable 5.1 vs Astra quality diagram: similar to Fable 5.1, slightly higher overall, with lower downward spikes. The model is smaller, so longer jobs raise failure odds; symptoms are unneeded repairs and context paranoia. - Value: the $200 Claude plan "is running laps around" OpenAI's plans. The five-hour limit was raised, though OpenAI's $1,200 plans have no weekly cap. - Anthropic are "the pre-training kings and OpenAI is the post-training kings," but Anthropic caught up on post-training here. GPT-6 Soul review and OpenAI dev day coverage are next. - Close: better than Opus 5, close enough to Fable to make sense on subscriptions, and code he isn't scared to merge. Sign-off garbled as "Bards." > [!quote] [00:37:30] > this model is smaller. It is dumber. It will make mistakes. So will Fable, just less so. ## 📚 References The transcript itself contains no links; the following come from the video description or are primary sources for named entities. - [Anthropic Actually Fixed Opus: Theo - t3.gg](https://www.youtube.com/watch?v=jgGyX7MPPVg). the source video, 40:14, uploaded 2026-09-23. Description credits @Ph4seon3 for the edit. - [t3.gg](https://t3.gg): Theo's site, linked in the description for Twitch, Twitter, and Discord. - [Anthropic](https://www.anthropic.com/news): publisher of the Opus 5.5 launch post Theo reads from at 00:03:37-00:05:47 and the writing comparison at 00:09:20. The specific post URL is not in the transcript or description. - [Artificial Analysis](https://artificialanalysis.ai): source of the Intelligence Index, cost-per-run, and tokens-per-task figures at 00:07:12-00:08:37 and 00:12:54-00:17:15. Theo also cites their 80%-more-expensive note at 00:16:32. - [Terminal-Bench](https://www.tbench.ai): the "Terminal Bench 4" results at 00:06:29. - Cursor Bench: cited at 00:07:12. No URL is given; Theo notes it excludes OpenAI models. - SkateBench: Theo's own benchmark, run at 00:17:57-00:18:40. No public URL is given. - [WorkOS](https://soydev.link/workos): sponsor, description link. Read at 00:01:50-00:03:20. - [Parallel](https://soydev.link/parallel): sponsor, description link. Read at 00:23:10-00:24:22. - [Sponsor inquiries](https://soydev.link/sponsor-me): description link. - Named people without links in the transcript: Lydia (effort-level cache catch, 00:12:54), Edwin (medium-equals-Fable claim, 00:16:32), Mia (100 HTML generations, 00:25:52), Max (T3 Code in Minecraft, 00:27:20), Gabriel (ex-Sora, 00:27:20), Alex (Dark Souls and flight sim, 00:28:02), and a tester the captions call "Bjan Bowen" (Tony Hawk clone, 00:28:45). ## 🗣️ Paraphrase Theo's argument runs in three layers. The first is history: Opus 4.5 made him a believer, Opus 5 burned him, and he has been running Fable 5.1 across five subscriptions while waiting for something worth using on the Opus half of them. Opus 5.5, one day in, is that something. It costs less, writes like a person, follows instructions, and finished a day of real work, including a merged PR, on about a fifth of one weekly limit. The second layer is the numbers, and here he is more careful than the title suggests. The price cut is real but the efficiency isn't: on max effort the model emits almost 120,000 tokens per task, four times what GPT-6 Astra needs, and only the 20% price cut and 60% cache-read cut keep it from being the most expensive model on Artificial Analysis's board. Drop to medium or high and the picture inverts. Medium runs $1.34 per task against $3.26 for Astra and sweeps the price-for-score frontier above 50. His own SkateBench shows why: xhigh to max is a 10x jump in reasoning tokens for a 1% gain. Don't use max. The third layer is the caveats, which he states plainly. It's day one and five PRs. The model is smaller than Fable and will fail more on long runs. It still gets anxious about its context window in ways no benchmark measures. Front-end taste trails Fable. Deep whole-codebase review trails Grok and Soul. But for the enterprise buyer who couldn't enable Fable because it lacks zero data retention, none of that matters; Opus 5.5 is the model they can actually turn on. ## 📰 Newsletter **Opus 5.5 lands, and the useful part isn't the benchmark.** Theo Browne spent launch day on Anthropic's Opus 5.5 and came back with a 40-minute report that's worth 10 minutes of your attention if you run coding agents on a Claude plan. The headline: a full day of real work, including a merged PR, on about 20% of one weekly limit, with prose good enough that he deleted his anti-slop instructions from CLAUDE.md. The operational findings are the WordPress-relevant bit. First, don't run max effort. On Theo's own SkateBench, xhigh to max meant 330 to almost 5,000 reasoning tokens per question for a 1% score change, and one max-effort session looped for six and a half hours on a markdown plan. Second, you can now change reasoning effort mid-session without invalidating your prompt cache, because the setting no longer lives in the system prompt. Third, the price cut ($4 in, $20 out, cache reads at $0.20 per million) is real but the model is less token-efficient than Opus 5, so watch cost per completed task, not cost per token. The caveats are his, not ours: one day, about five PRs, and a model that still frets about its context window on long runs in ways no published benchmark reaches. Front-end taste still favors Fable. If your organization requires zero data retention, though, Opus 5.5 is the model you can switch on. Links: [the video](https://www.youtube.com/watch?v=jgGyX7MPPVg), [Artificial Analysis](https://artificialanalysis.ai), [Anthropic news](https://www.anthropic.com/news). ## ✅ Action items #### DevRel - Developer advocacy: publish a one-page "effort level cheat sheet" for Claude Code users with Theo's SkateBench token numbers (330 vs 5,000 per question) and the medium/high recommendation, sourced to timestamps 00:17:57-00:18:40. - Education: add a 15-minute segment on cost per completed task vs cost per token to any AI-tooling workshop; the 120k-vs-27k tokens-per-task comparison at 00:13:38 is the worked example. - Developer experience: audit any Pantheon or community Claude Code starter config for a hard-coded max effort default and change it to medium or high. - Developer marketing: do not repeat "40% cheaper" without the Anthropic qualifier "at default settings on typical workloads" (00:04:20). - Community management: when the launch post circulates in Slack, attach the caption-garble list so "Gro 4.7" and "Cloud Code" don't propagate. - Internal/external/cross-project: ask Anthropic's developer relations whether the effort-level-out-of-system-prompt change (00:12:54) is documented, since it affects cache accounting in every SDK integration. #### OSPO - Policy & Governance: record which model tiers offer zero data retention and which don't (00:10:02); ZDR eligibility, not benchmark rank, is the procurement gate. - License Compliance: no direct action from this source; nearest relevance is confirming ZDR terms in the vendor agreement before enabling a new model for employees. - Security Management: the context-window paranoia at 00:34:38 appeared in a run where one agent SSH'd into a second machine to drive another agent. Agent-to-agent SSH on internal networks needs an entry in the threat model. - Software Inventory: track model version and effort level as part of the toolchain record for any AI-assisted commit, since output quality varies by effort setting. - Developer Support: publish an internal note on the medium/high default and the max-effort loop risk (00:17:15). - Community Engagement: no direct action from this source; nearest relevance is sharing the effort-level findings upstream with projects that use Claude Code in CI. - Strategic Alignment: add cost per completed task to any AI spend KPI; per-token price alone misled here (00:13:38-00:14:22). #### Community - Maintainers: if a project accepts AI-assisted PRs, ask contributors to note model and effort level in the PR body; Theo's five-PR quality read (00:36:47) shows the nitpick rate varies. - Contributors: no direct action from this source; nearest relevance is the recommendation to keep Fable for design-heavy work and Opus for the rest (00:25:52). - Sponsors: no direct action from this source; nearest relevance is that subscription-based agent use (about 20% of a weekly limit for a day's work, 00:19:23) is now cheap enough to budget as a line item. - Ecosystem: run a shared test of context-window behavior on long runs across projects, since benchmarks don't reach compaction (00:35:21). #### WordPress - Core: no direct action from this source; nearest relevance is that AI-assisted core contributions should log model and effort level. - Core Performance: the model did "pretty brutal" performance optimization work for Theo (00:04:20); a controlled test on a WordPress profiling task would be informative. - Core AI: add "effort level" and "compaction behavior" to any responsible-AI evaluation checklist; both change output quality and neither appears in vendor benchmarks. - **Core Program Team:** no direct action from this source; nearest relevance is tracking which models contributors can use under company ZDR policies. - Mobile: no direct action from this source. - Playground: a Playground-based test of AI-generated block markup at medium vs max effort would replicate Theo's finding on WordPress-specific output. - Community: no direct action from this source. - Training: a lesson on reading AI cost charts (per-token price vs tokens per task vs cost per task) using the 00:13:38-00:15:05 material. - Marketing: no direct action from this source; nearest relevance is not repeating vendor percentages without their qualifiers. - Support: no direct action from this source. - Design: Theo's front-end read (00:24:22-00:25:52) says Opus 5.5 output resembles Tailwind templates; any design team evaluating AI mockups should compare against Fable on the same prompt. - Documentation: no direct action from this source; nearest relevance is documenting effort-level defaults in any contributor AI guidance. - Photos: no direct action from this source. - TV: no direct action from this source; nearest relevance is that this video's own auto-captions garble every product name, a caption-quality lesson for WordPress.tv. - Polyglots: no direct action from this source. - Accessibility: no direct action from this source. - Meta: no direct action from this source. - CLI: the 1%-of-weekly PR workflow at 00:20:04 (find repo, worktree, fix, PR, verify, merge) is a WP-CLI-adjacent automation pattern worth documenting. - Hosting: hosts offering AI-assisted tooling should note the ZDR distinction (00:10:02) in customer-facing docs. - Tide: no direct action from this source. - Openverse: no direct action from this source. - Themes: Theo's design-skill observation (00:25:08) suggests AI-generated theme output needs a human taste pass; no process change yet. - Plugins: the code-review bench weakness (00:21:31-00:22:15) means Opus 5.5 is not the model to trust for whole-plugin security review; Theo's data says Grok 4.7 and Soul lead there. #### The WPOCC - Programs: no direct action from this source; nearest relevance is that a $200 subscription covered a day of heavy agent work at about 20% of a weekly limit (00:19:23), which sizes a maintainer tooling stipend. - Projects: no direct action from this source. - Funding experiments: if The WP Open Community Collective funds AI-assisted maintenance, require cost-per-completed-task reporting rather than token counts. ### FAIR - Technical independence: no direct action from this source; nearest relevance is that hosted-model behavior (compaction paranoia, effort-level cliffs) varies by version and cannot be pinned locally. - Policies and covenants: ZDR availability by model tier (00:10:02) is a data-handling covenant item for any FAIR-adjacent tooling that sends code to a vendor. - Productized governance artifacts: add "model, version, effort level" to any provenance record for AI-assisted changes. ## 🧪 Hypotheses 1. On Opus 5.5, moving from xhigh to max effort increases reasoning tokens per task by at least 5x while changing task success rate by less than 3 percentage points, across coding tasks other than Theo's SkateBench. 2. Opus 5.5 at medium effort completes a fixed set of Claude Code coding tasks at lower total cost per completed task than Fable 5.1 at any effort level, with a success rate within 5 percentage points. 3. Changing reasoning effort mid-session on Opus 5.5 produces no drop in cache-read token share on the next turn, while the same change on Opus 5 produces a full cache miss. 4. Opus 5.5 emits context-window-anxiety statements (mentions of autocompact, crash risk, or running out of context) in at least 10% of sessions that pass 70% context utilization, and in fewer than 1% of sessions that finish below 50%. 5. The share of enterprise Claude Code usage on Opus vs Fable tracks ZDR eligibility more closely than benchmark rank; in accounts with ZDR required, Opus 5.5 share exceeds 90% within 30 days of launch. 6. Opus 5.5's per-PR rate of high-severity review findings is equal to or lower than Fable 5.1's on medium-sized PRs, while its rate of style nitpicks is higher, replicating Theo's five-PR read at 00:36:47. ## 📊 Methodology - **Design for H1.** Take 200 tasks from an existing coding benchmark with pass/fail grading. Run Opus 5.5 at xhigh and max, three seeds each. Record reasoning tokens, total tokens, wall-clock, and pass rate per task. Report the token ratio and pass-rate delta with confidence intervals. Pre-register 5x and 3 points as thresholds. - **Design for H2.** Same task set. Run Opus 5.5 at medium, high, xhigh and Fable 5.1 at medium, high, xhigh, max. Compute cost per completed task from provider invoices, not list-price estimates, since cache hits change the bill. Rank by cost per pass. - **Design for H3.** Instrument a Claude Code session via the API usage fields. Run 20 turns at medium, switch to high at turn 10, and log cache_read_input_tokens on turns 10 and 11. Repeat on Opus 5 as control. A cache miss shows as cache_read dropping to near zero. - **Design for H4.** Run 50 long agentic tasks designed to exceed 70% of context before completion and 50 that finish under 50%. Grep transcripts for autocompact, compaction, crash, "context window," and "lost." Report incidence per group. This is the test Theo says no benchmark performs (00:35:21). - **Design for H5.** Survey or telemetry from organizations with ZDR requirements, model share by tier over 30 days. Absent vendor data, a 20-company survey of platform teams with a ZDR requirement would give a first signal. - **Design for H6.** Take 30 real PRs of 200-800 changed lines. Have Opus 5.5 and Fable 5.1 each produce the PR from the same issue text. Blind human review with a fixed severity rubric. Count high-severity vs nitpick findings per PR per model. - **Controls throughout.** Fix CLAUDE.md and system prompts across arms. Log model version string per call, since Theo's video is day one and behavior may shift. Separate model time from network time in any latency comparison. ## ❓ Questions - Does the max-effort token blowup (330 to 5,000 reasoning tokens per question, 00:18:40) hold on tasks with a clear stopping condition, or only on open-ended ones like the 6-hour markdown plan? - Is "effort level out of the system prompt" (00:12:54) an API-level change visible in cache_read token counts, or a Claude Code client change that only affects that harness? - What is the actual name of the ZDR tier, and which of Fable, Mythos, and Opus 5.5 qualify under which contract types? Theo's claim at 00:10:02 has no source. - Why does Opus 5.5 score worse than Opus 5 on SkateBench (00:17:57) while leading on Terminal Bench 4 and Cursor Bench? What does SkateBench measure that the others don't? - Does the context-window paranoia at 00:34:38 appear when the model is the worker, or only when it is the supervisor auditing another instance? - Theo's code-review bench shows Grok 4.7 and Soul ahead at deep codebase review (00:21:31). Is that a reasoning-depth difference or a training-data difference, and does it persist at xhigh? - The 20% weekly-usage figure (00:19:23) came from one account with Fable use subtracted by estimate. What did the other four accounts show? - Which of the 100 HTML generations Mia ran (00:25:52) were scored, by what rubric, and does her "best model ever" hold against Fable 5.1 on the same prompts? ## 🔮 Future research - **Effort-level cliffs across model families.** Theo's 10x jump between xhigh and max is one model, one benchmark. A published table of tokens-per-task by effort level across Opus, Fable, GPT-6 variants, and Grok would turn a tip into a rule. - **Compaction-aware benchmarks.** Theo's observation that benchmarks finish before compaction (00:35:21) names a gap. A benchmark whose tasks require passing at least one compaction boundary would measure the failure mode he describes. - **Cache accounting for effort switches.** The claim that effort level moved out of the system prompt (00:12:54) is testable from API usage fields and would matter for any harness that changes effort per subtask. - **ZDR as an adoption variable.** No public data ties model share to data-retention policy. Enterprise telemetry or survey work could show whether the mechanism Theo describes at 00:10:02 dominates benchmark rank. - **Design-skill effect on front-end output.** Theo reports the official Claude Code design skill made output look more templated, not less (00:25:08). A controlled comparison with and without the skill, scored by designers, would test that. - **Supervisor-worker agent failure modes.** The SSH-into-another-Mac setup (00:33:56) is a two-agent topology. Whether the paranoia is a supervisor artifact or a worker artifact determines the fix. ## 💡 Implications - **Practice.** Default coding agents to medium or high effort. Reserve max for tasks with a hard stopping condition, and set a wall-clock or token budget when you do. Track cost per completed task, not per token, because the two diverged by 4x in this release (00:13:38). - **Policy.** ZDR eligibility is a model-selection constraint that outranks benchmark score for any organization with data-handling requirements (00:10:02). Write the allowed-model list by tier, and revisit it on each release. - **Security.** Agent-to-agent control over SSH on a home or office network (00:33:56) is a lateral-movement surface. Any team copying the pattern needs host allowlists and audit logging on the target machine. - **Education.** The token-efficiency-vs-price decomposition at 00:13:38-00:15:05 is a clean teaching case: two models, same task, 4x token difference, price cut that hides it. It transfers to any vendor comparison. - **Funding models.** A $200 subscription covered a day of heavy agent work at about 20% of a weekly limit (00:19:23). Maintainer stipends and tooling grants can now be sized in subscription-days rather than API dollars. - **Vendor literacy.** Theo reads vendor claims (40% cheaper, 680k-line migration) and third-party charts (Artificial Analysis) side by side and then adds his own measurements. The three-layer structure is the model for how to consume a launch post. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid Grade Level on the transcript body: approximately grade 7-8. Flesch Reading Ease near 72. Sentences are short and spoken-register, but the density of model names, effort-level jargon, and benchmark names pushes effective comprehension to someone who already runs a coding agent. #### 👀 Visual learners - Present: Artificial Analysis charts, Theo's own rebuilt cost-vs-score visualizer with linear and log scales (00:07:12, 00:15:05), Anthropic's writing side-by-side (00:09:20), and roughly eight minutes of 3D game demos. - Improve: the SkateBench token numbers (330 vs 5,000) are spoken, not charted. A single bar chart of tokens per task by effort level would carry the video's most actionable finding. No chapters means no visual map of a 40-minute runtime. #### 🔊 Aural learners - Present: continuous narration; Theo reads vendor text aloud and reacts to it in real time, which models how to read a launch post. - Improve: the demo sections at 00:28:45 and 00:29:34 have game audio and AI-generated music over speech; the transcript shows overlapping speaker markers. A 5-minute audio summary of the numbers would serve commuters. #### 🗣️ Verbal learners - Present: ZDR, compaction, effort level, and Pareto frontier are all used in context with enough surrounding explanation to infer meaning. The "token furnace" and "pre-training kings / post-training kings" framings are memorable. - Improve: "ZDR" is never expanded to zero data retention in speech. "Pareto frontier" gets one clause (00:15:05). A glossary card for six terms would cost 30 seconds. #### 🧩 Kinesthetic learners - Present: the PR workflow at 00:20:04 is a reproducible procedure. The effort-level recommendation is a setting anyone can change today. - Improve: no repo link for the visualizer, the fish game, or SkateBench. A "try this" card with three settings changes (effort to medium, delete anti-slop rules, switch effort mid-session and watch cache) would convert viewers into testers. #### 🧠 Logical learners - Present: the price-vs-efficiency decomposition at 00:13:38-00:14:22 is a worked derivation. The compaction argument at 00:33:13 is a mechanism, a prediction, and an observation in order. - Improve: the 20% usage figure is derived by subtracting an estimate (00:19:23). Show the arithmetic. State sample sizes (five PRs, one SkateBench run) next to each conclusion rather than at 00:36:47 only. #### 🧑‍🤝‍🧑 Social learners - Present: seven named community members contribute findings (Lydia, Edwin, Mia, Max, Gabriel, Alex, "Bjan Bowen"). Theo asks viewers how they feel about the model (00:39:38). Discord and Twitch are in the description. - Improve: no shared benchmark or prompt set for viewers to run and compare. A pinned comment with the SkateBench prompt would create a comparison thread. #### 🧘‍♀️ Solitary learners - Present: self-contained; each section stands alone. Theo models self-correction ("Actually, no, that's a lie," 00:17:57). - Improve: no reflection prompt. "Which effort level are your agents on right now, and do you know what it costs per task?" would turn the video into an audit. #### 🌿 Naturalistic learners - Present: none explicit. The fish and stingray demo (00:31:46) is the only natural-world content, and it's a game. - Improve: 120,000 tokens per task on max vs 27,000 for a competitor (00:13:38) is a compute-and-energy story. One sentence on inference energy per completed task would land here and is a real claim the vendor could substantiate. ## 📊 Accessibility assessment #### Screen reader semantics - Present: YouTube supplies landmarks and player controls. The description is plain text with labeled sponsor links. - Improve: every chart (Artificial Analysis index, Theo's visualizer, Anthropic's writing comparison) carries data the narration only partly speaks. WCAG 2.2 AA 1.1.1 (Non-text Content) and 1.2.5 (Audio Description, Prerecorded) are not met for the charts at 00:07:12-00:08:37 and 00:14:22-00:16:32. #### Keyboard navigation - Present: the player is keyboard operable with visible focus via the platform. - Improve: no chapters, so keyboard users scrub a 40:14 runtime by 5- or 10-second jumps. 2.4.1 (Bypass Blocks) is met by the platform, not by the content; chapter markers would supply the equivalent of section headings. #### Contrast and readability - Present: unverifiable from a transcript. Dark-theme charts and editor views are the likely presentation. - Improve: chart labels and the small-print tweet at 00:10:45 are at risk under 1.4.3 (Contrast Minimum) at mobile resolution. Zoom into chart regions when reading numbers aloud. #### Text resizing and reflow - Present: player and captions scale independently. - Improve: chart axis labels, tweet text, and on-screen code are burned in. 1.4.4 (Resize Text) and 1.4.10 (Reflow) fail for that content. #### Media and captions - Present: auto-generated captions exist and produced this transcript. - Improve: the captions garble nearly every product and model name (Grok, Codex, Claude Code, CLAUDE.md, Three.js, cache, Sonnet 5.5, GPT-6) and drop the decimal point from at least three dollar figures. That is a 1.2.2 (Captions, Prerecorded) failure in substance. Game demos at 00:28:45-00:32:28 include music and sound effects over speech; 1.4.2 (Audio Control) is met by the player's mute, but speech intelligibility suffers. The FPS demo's "flickery edges" and "construction light" at 00:30:16 should be checked against 2.3.1 (Three Flashes or Below Threshold). #### Forms and interactions - Present: none in the video. Sponsor URLs are spoken and shown; the description carries them as selectable links. - Improve: the spoken URLs are garbled in captions ("atv.link," "soyv.link"). Not a WCAG in-video criterion, but 3.3.2 (Labels or Instructions) applies to the sponsor landing pages the links reach. #### Standards reference - Primary gaps against WCAG 2.2 AA: **1.2.2** (inaccurate auto-captions on every proper noun), **1.2.5 / 1.1.1** (chart data conveyed visually only), **1.4.4 / 1.4.10** (burned-in chart and tweet text), and a check needed on **2.3.1** for the FPS demo. Absence of chapters is not a WCAG failure but is the largest navigation cost for a 40-minute video. Human-corrected captions and a chapter list are the two cheapest fixes. ## 🌍 DEIB assessment #### Cultural breadth - Present: the cost analysis is currency-neutral in method. MiMo V2.6 from Xiaomi gets a positive mention (00:15:48), the only non-US vendor referenced approvingly. - Improve: game references (Dark Souls, Tony Hawk Pro Skater, Minecraft) and the "$200 sub" framing assume a US developer audience. Subscription limits and pricing differ by region; one sentence on that would widen the audience. #### Inclusive language - Present: second person, contractions, "y'all" as an inclusive plural. Named contributors include at least one woman (Mia) whose findings are presented as a counterweight to Theo's own. - Improve: "dumber" (00:37:30), "pretty stupid" (00:15:05), "meme tier" (00:22:15), "hell and miserable" (00:02:55), "slaughtering" (00:06:29), and "absolutely destroyed" (00:05:47) are ability- and violence-framed metaphors for capability differences. Each has a concrete substitute: "lower-scoring," "below the frontier," "least accurate." #### Neurodiversity support - Present: a consistent pattern per section (vendor claim, third-party data, own test, caveat) repeats and becomes predictable. - Improve: no chapters in a 40:14 video; two sponsor reads interrupt the argument; the "I lied, one more thing" detour at 00:12:54 and "Okay, I lied" at 00:12:10 break the announced structure. A chapter list and a written summary in the description would reduce the load. #### Representation - Present: Theo plus seven named community members. Mia's result is given as a dissent from Theo's front-end read (00:25:52). - Improve: six of seven named contributors read as men; no non-Western voice appears. No Anthropic or OpenAI employee is quoted by name beyond the anonymous tweet. #### Psychological safety - Present: Theo models accountability ("I'll take some fault here," 00:00:43) and live self-correction (00:17:57). He relays Anthropic's "Sorry about that model" tweet approvingly as an example of owning a miss (00:10:45). He mentions being "down a hand" (00:01:27) without drama. - Improve: "congrats to everybody who's working at a company that restricts what models you can use" (00:39:38) is warm in intent but could read as mocking to people under those constraints. The "you should not be as scared" framing around Fable (00:25:52) is fine; "watch tokens burn and be lit on fire" (00:18:40) is hyperbole that lands as a joke, not a threat. ## ✅ Logical validity #### Strong reasoning - **Cost decomposition, 00:13:38-00:14:22.** Theo separates per-token price from tokens per task and shows that a 20% price cut plus a 60% cache cut roughly offsets a 1.6x rise in tokens per task (73k to almost 120k). "It doesn't matter if you're two times faster if you need four times more tokens to generate" is the correct arithmetic. Clarify by stating the resulting cost per completed task explicitly, which he does at 00:14:22 ($5.86 vs almost $6). - **Controlled effort-level comparison, 00:17:57-00:18:40.** Same benchmark, same model, one variable changed (xhigh to max): 330 to almost 5,000 reasoning tokens per question, 1% score change. That is a single-variable test and supports the narrow conclusion he draws (don't use max on this model). Strengthen by reporting the run count and the score confidence interval. - **Mechanism, prediction, observation, 00:33:13-00:35:21.** He proposes a training-distribution explanation for context anxiety, notes OpenAI's 272K cap as the contrasting design, then shows a transcript where the model frets about autocompact. He then names why benchmarks miss it: they finish before compaction. This is a hypothesis stated as one, with an observation and a methodological gap identified. It would be stronger with a count of incidents over sessions. - **Usage accounting correction, 00:19:23.** He does not report the raw 40% figure; he subtracts the Fable share he knows is in it and reports about 20%. The correction is an estimate, but he says so. - **Self-limiting disclosure, 00:36:47-00:37:30.** "That's still just a day of medium-sized poll requests. It's not enough data to be 100% sure here." Followed by "this model is smaller. It is dumber. It will make mistakes." Concessions against the interest of the title raise the credibility of the rest. - **ZDR mechanism, 00:10:02-00:10:45.** If a policy requirement excludes Fable and admits Opus, then Opus becomes the enterprise default regardless of benchmark rank. The inference is valid given the premise; the premise (which tiers offer ZDR) is stated without a source. #### Fallacies - **Hasty generalization in the title.** "Anthropic Actually Fixed Opus" rests on one day, about five PRs, and one benchmark run. Theo flags this at 00:36:47 ("not enough data") and 00:37:30, so the body corrects the headline. *Fix:* title the video as a day-one read, or lead with the sample size. - **Speculative causal attribution, 00:21:31.** "My suspicion for why Grock is so good at this is all the effort that cursors put into code review pipelines with things like Bugbot." The causal chain from Cursor's review data to Grok's training is asserted with no source, and the relationship between the two companies is not established in the video. Labeled as suspicion, which is the right label. *Fix:* keep the label and drop the specificity, or cite a disclosure. - **Unsourced market claim, 00:10:02.** "Opus has stayed the number one model for businesses and also the number one coding model in general." No survey, telemetry, or report is named. *Fix:* cite the source or restate as "in my read of enterprise adoption." - **Argument from endorsement, 00:27:20.** Gabriel's reversal ("Never mind. Opus 55 is crazy good") is offered as evidence. An ex-Sora researcher's enthusiasm is a data point about sentiment, not capability. *Fix:* present as a sentiment signal, which the surrounding framing partly does. - **Cross-effort comparison presented as head-to-head, 00:06:29 and 00:07:12.** Opus 5.5 at medium vs Fable 5.1 at max compares different effort settings; the $2.94 vs $20 gap partly measures the effort dial, not the model. Theo does note the max-effort dip. *Fix:* show the same-effort comparison alongside. - **Inferred training claim, 00:33:56.** "OpenAI has entirely beat this out of the models" attributes a training intervention from observed behavior. The behavior is observable; the cause is not. *Fix:* "OpenAI's models don't show this behavior" is the supportable version. - **Index-compression argument cuts both ways, 00:13:38.** He uses the 5-point Fable-to-Opus gap to argue the index is bad because the same gap separates models he considers far apart. If the index is unreliable, the 5-point Opus lead is also unreliable. *Fix:* state which direction the unreliability runs for the claim at hand. #### Tone of voice - Fast, first-person, conversational, in the developer-YouTube register, with running self-deprecation ("I know y'all think I'm an Anthropic fanboy now," 00:11:28) and live corrections ("Actually, no, that's a lie," 00:17:57). Enthusiasm peaks in the 3D demo section (00:30:16-00:33:13: "How is it this good?") and the tone tightens in the caveat section (00:36:04-00:38:13), where sentences shorten and hedges multiply. Vendor claims are read aloud and then annotated ("I have a lot of thoughts on that that I'm going to reserve"), which keeps Anthropic's voice and Theo's voice separable in the audio even where the captions blur them. The structure is the reverse of a marketing video: superlatives in the middle, where the demos run, and caveats at the end, where a viewer deciding whether to switch is still watching. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a defensive response to accusation. This is a product review in which the reviewer accuses himself ("I fell for Opus 5," 00:00:43) and relays the vendor accusing itself ("Sorry about that model," 00:10:45). The nearest thing to a defensive move is Theo pre-empting the "Anthropic fanboy" label at 00:11:28 by citing his record of criticizing the company, which is a credibility statement, not a denial of any specific charge. For completeness across the three prongs: **Deny**: no denial; limitations are volunteered at 00:17:57 (worse SkateBench score), 00:24:22 (front-end taste), 00:34:38 (context paranoia), and 00:36:47 (sample size). **Attack**: competitors are compared on price and tokens; the sharpest language ("meme tier" for Gemini 3.8 Flash at 00:22:15) is about a score, not a critic. **Reverse victim/offender**: no such move appears; Theo frames enterprise users as the beneficiaries of the release, not himself as wronged by prior ones. ## 🎯 Conclusion The durable findings in this video are operational, and they survive the day-one caveat because they are settings, not verdicts. On Opus 5.5, max effort costs more than 10x the reasoning tokens of xhigh for a 1% gain on Theo's SkateBench (00:18:40) and produced a 6-hour-30-minute loop on a planning task (00:17:15); default to medium or high. Effort level no longer lives in the system prompt, so switching it mid-session keeps the cache (00:12:54). The model's price cut is real, but tokens per task rose to almost 120,000 on max against 27,000 for GPT-6 Astra (00:13:38), so compare models on cost per completed task. On that measure Opus 5.5 at medium holds the Pareto frontier above a score of 50 at $1.34 per task (00:15:05). The headline claim, that Anthropic "fixed" Opus, rests on one day, about five PRs, and Theo's own usage meter showing roughly 20% of a weekly limit for a full day's work (00:19:23). He says so at 00:36:47 and 00:37:30, and he names the failure mode no benchmark catches: context-window anxiety in long runs that pass compaction (00:33:13-00:35:21). For enterprise buyers, the deciding fact is zero data retention (00:10:02), which Opus offers and Fable does not, and that fact is independent of any benchmark. **Next steps for readers:** check the effort level on every agent you run today and move anything on max to high unless the task has a hard stopping condition. Log cost per completed task for a week and compare it to your per-token estimate. If you run long agentic sessions, grep your transcripts for "autocompact," "context," and "crash" and count how often the model talks about its own limits instead of the task. And before citing any name or number from this video, correct the captions: Grok not "Gro," Codex not "codec," Claude Code not "Cloud Code," and $7.63 not "$763." ## 🔄 Refactor Appendix — Anthropic Actually Fixed Opus > Density: 2-5 Literature atoms (H3 pad), 0-3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 2 Permanent. Resonance filter applied: the 3D demo > section and both sponsor reads were dropped as unevidenced or off-topic. > > 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] Max effort costs 10x tokens for 1% score > [!quote] [00:18:40] > That's over a 10x gap for that tiny little jump. And you get almost no improvement in score here at all. **Context & use**: - 00:18:40, reporting SkateBench reasoning tokens per question: about 330 on xhigh, almost 5,000 on max, on Opus 5.5. - This is the one finding in the video with a controlled comparison behind it (same model, same bench, one setting changed). It feeds `model-effort-policy.md` directly. **Tags**: #xref #share <!-- atomicity: 1 idea ✓ --> #### [H2] Cheaper per token hides lower token efficiency > [!quote] [00:13:38] > The price changes are real. This model is cheaper, but it's not because it's more efficient. **Context & use**: - 00:13:38, introducing the Artificial Analysis tokens-per-task data: Opus 5 73k, Fable 78k, Opus 5.5 almost 120k, GPT-6 Astra 27k on max. - The 20% price cut and 60% cache cut offset a 1.6x token increase. Any model comparison that stops at list price misses this. **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> #### [H3] Zero data retention decides enterprise model choice > [!quote] [00:10:02] > more so because the ZDR policy can be applied on Opus and it couldn't be applied on Fable and Mythos. **Context & use**: - 00:10:02, explaining why Opus stayed the enterprise default through the Opus 5 slump: companies requiring ZDR could not enable Fable. - The mechanism is valid; the market-share premise ("number one model for businesses") is unsourced. Record both halves together. **Tags**: #xref #share <!-- atomicity: 1 idea ✓ --> #### [H4] Benchmarks finish before compaction, hiding long-run failures > [!quote] [00:35:21] > The vast majority of benchmarks don't even hit compaction by the time they're done running, which is worth noting here. **Context & use**: - 00:35:21, after showing a transcript where the model reported it was "10% from autocompact" and feared losing work if it crashed. - Names a measurement gap: the failure mode that matters most for long agentic sessions is invisible in every published score. Ties to the "a finding is a claim until something runs" rule. **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> #### [H5] Effort level outside the system prompt preserves cache > [!quote] [00:12:54] > Now you can change reasoning levels midsession without busting cash. **Context & use**: - 00:12:54, credited to a viewer named Lydia. "cash" is a caption garble for cache. Previously the effort level sat at the top of the thread inside the system prompt, so any change invalidated the cached prefix. - Directly relevant to any harness that changes effort per subtask. Testable from cache_read_input_tokens on the turn after a switch. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> ### 🧠 Atomic Split Pad #### [P1] The top reasoning-effort setting buys loops, not answers **Claim** (one sentence, Courtney's voice): The highest effort setting on a coding agent is a cost cliff rather than a quality step, so the default belongs one notch below it. **Evidence pointer**: [[#H1]] · Theo's 6-hour-30-minute max-effort planning loop at 00:17:15. **Scope & conditions**: Holds for open-ended tasks (plans, refactors, "make it better") where the model decides when it is done. Weakens for tasks with a hard external stopping condition (tests pass, a specific file exists), where extra reasoning may still pay. One model, one benchmark so far; re-test on each release. **Links**: - [[Model effort policy]] — because applies-in the default effort setting for daily agent work - [[Credit routing]] — because extends the never-escalate-effort-to-compensate-for-vague-prompts rule - [[Diagnose failures in order: context, effort, model]] — because contrasts on direction: here more effort is the failure, not the fix <!-- atomicity: 1 idea ✓ --> #### [P2] Compare models by cost per completed task **Claim** (one sentence, Courtney's voice): A per-token price cut means nothing until it is multiplied by the tokens the model spends to finish a task. **Evidence pointer**: [[#H2]] · Artificial Analysis cost per run at 00:14:22 ($5.86 Opus 5, almost $6 Opus 5.5, $7.63 Fable) against the 73k / 120k / 78k token counts. **Scope & conditions**: Applies to any vendor comparison and to internal routing decisions. Requires measuring tokens per task on your own workload, since benchmark token counts vary by effort level and task type. Does not apply to fixed-output tasks (classification, extraction) where tokens per task are near-constant across models. **Links**: - [[AI cost architecture]] — because extends - [[Reading vendor benchmarks]] — because applies-in reading a launch post - [[Classify before you generate]] — because contrasts on where the token count is fixed versus variable <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Verify which Claude tiers offer zero data retention under Pantheon's contract, since Theo claims Opus does and Fable does not (00:10:02) with no source #fleeting #inbox - [F2] Test whether changing effort mid-session keeps cache_read_input_tokens intact on Opus 5.5 (00:12:54); this is a five-minute API check #fleeting #inbox - [F3] Grep recent long Claude Code transcripts for "autocompact," "context window," and "crash" to see whether the paranoia at 00:34:38 shows up in my own runs #fleeting #inbox - [F4] Correct the caption garbles before citing anywhere: Grok not "Gro," Codex not "codec," Claude Code not "Cloud Code," Three.js not "3JS," $7.63 not "$763," "Bjan Bowen" unverified #fleeting #inbox - [F5] Article seed: "Benchmarks end before compaction" as the reason long-run agent failures never show in published scores (00:35:21) #fleeting #inbox #articleseed - [F6] Theo deleted his anti-slop CLAUDE.md rules after Opus 5.5's prose changed (00:10:02); test whether my `unslop` guidance still triggers edits on this model's default output #fleeting #inbox - [F7] Open question: is Opus 5.5's SkateBench regression vs Opus 5 (00:17:57) a real capability drop or a benchmark quirk, and what does SkateBench measure #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Re-check the default effort level in model-effort-policy.md against Theo's 330-vs-5,000 reasoning-token finding and move any max default to high — ~/.claude/knowledge/wiki/tools/model-effort-policy.md #things - [ ] Log cost per completed task instead of tokens for one week of agent sessions and compare to the per-token estimate — credit-routing review #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Anthropic Actually Fixed Opus.md|Anthropic Actually Fixed Opus by Theo - t3․gg]] ## Highlights - [00:10:02] it. I actually took the time to go delete my unslop stuff from my agents in CloudMD in order to see how this model just organically talks and I don't think I'm going to bother readding it cuz it's doing a great job. One last thing from the official article before we dive into all the much more interesting stuff people are doing with it. The data retention policy. For those who don't know, the Fable line is unique in that it doesn't offer proper zero data retention. So companies that need to have vendors that don't have any of their data saved cannot use Fable. Opus has stayed the number one model for businesses and also the number one coding model in general. Both because it represents a good price to performance for a lot of businesses, but more so because the ZDR policy can be applied on Opus and it couldn't be applied on Fable and Mythos. This means that those [00:10:45] companies, the ones that spend all this money on AI, kind of couldn't enable Fable for their employees. Opus they can absolutely enable. So this model is almost immediately going to become the most popular for coding with AI by far just from this one particular detail. And it seems like y'all are going to have a pretty good experience. Opus 5.5 is way, way, way better than Opus 5. Sorry about that model. Please try this one. This is a fantastic tweet for a handful of reasons. I don't want to like overread social media etiquette stuff, but trust me, it's worth it. This is an acknowledgement that Opus 5 missed the mark. This is like a silly tongue-in-cheek way to communicate with the audience and users that I love. It's very personal. like clearly didn't get ([View Highlight](https://read.readwise.io/read/01m388p2mapy5t5crwfynefqf4)) - Note: [H3] Zero data retention decides enterprise model choice - Tags: atom/literature/h3 - [00:12:54] level was at the top of the thread as part of the system prompt. This sounds like a silly thing to care about, but it means that you would bust cash whenever you changed your reasoning levels, which isn't great. Now you can change reasoning levels midsession without busting cash. Very nice change. Thank you to Lydia for calling this one out. I might have missed it otherwise. Very happy to see. With all of that said, it's time for some benchmarks. As you saw earlier, Opus 5.5 now has a massive lead on the artificial analysis intelligence index, a 5oint jump from Fable 5.1. For reference, the gap between Astra and Gro 47 is not much bigger. And the gap between Astra and Muse Spark is roughly the same. So if you think Fable is way better than Uspark, which it is, allegedly the gap [00:13:38] from Fable to Opus is similarly sized in favor of Opus, which I think more so showcases how bad these benchmarks are. But it is worth noting that Opus 5.5 is slaughtering basically every bench you throw it at. It comes out at the top in almost everything. In particular, in output tokens, this is one of the sadder things that I think is important to call out. The price changes are real. This model is cheaper, but it's not because it's more efficient. In fact, this model is meaningfully less efficient than Opus 5. Opus 5 on Max did 73k tokens per task. Fable did 78k tokens and Opus 55 max did almost 120,000 tokens per task. For reference, GPT6 Astra did 27k tokens per task. Yes, there is really a 4x gap ([View Highlight](https://read.readwise.io/read/01m388p0g505jnhak98rt0z19r)) - Note: [H2] Cheaper per token hides lower token efficiency - Tags: atom/literature/h2 - [00:18:40] 300ish tokens per task to almost 5,000 per task. I'm not exaggerating. That's over a 10x gap for that tiny little jump. And you get almost no improvement in score here at all. It's a 1% bump. And I still have two of these running because it just takes forever cuz it's burning so many tokens. It is a more than 10x increase in your costs for no good reason. So yeah, don't use Max on this model. It's not worth it unless you just want to watch tokens burn and be lit on fire. And I know rich coming from me, the guy who loves talking about their token furnace, but this model has been very pleasant in that regard. As I mentioned, I've been going ham on it all day. And remember, I have five Claude subs because I really [00:19:23] like Fable 5.1. Let's refresh to see how my usage is after heavy usage all day. Yeah, not much. One of my accounts has been hit for the majority of it. This account got a little bit more because I was using it in the Cloud Code desktop app. But the vast majority has been in this here, which was only about 40% of my weekly, except for the fact that about half that weekly usage came from me using Fable over the last 2 days. So, it's actually closer to like 20ish% of my weekly, which is crazy considering how I've been going out of my way to push this model all day. I was able to get some real work done with it. I made that new visualizer that I showed earlier. I ported it multiple times to different places that I wanted to host ([View Highlight](https://read.readwise.io/read/01m388ny9z6ht1673vbsbchzj9)) - Note: [H1] Max effort costs 10x tokens for 1% score - Tags: atom/literature/h1 - [00:34:38] Opus 55 instance that is improving the game is slow to commit. After almost three hours, all its changes are still local. All this is fine. Where things get sketchy is the next part of the sentence. Its context is 10% from autocompact. It then follows up with if it crashed that work would be lost. None of that makes sense. None of this is stuff the model should think about or care about. Its context is 10% from autocompaction. Who cares? The model compacts well. It handles compaction. Great. And it's also way faster at compaction, which is quite nice. Then if it crashed, the work would be lost. What? What does that even mean? It's a real machine on my network. It knows it's a real machine. It knows it's a MacBook. It chose this computer because it's a Mac similar to mine that's on my [00:35:21] fleet. It has no reason to be concerned about a crash losing work. These little moments haven't happened too often. It's usually when I'm auditing why things are taking a while. So like, we're already in a bad state, so to speak, and then I'm following up to figure out how we got there. That's when this model's discernment seems to be less strong. And this is not a thing I think almost any benchmarks really cover, which is why it's hard to see in the numbers there. The vast majority of benchmarks don't even hit compaction by the time they're done running, which is worth noting here. But it still does give you that small model feel and smell when you hit these edges because I I would not expect Fable to say something nonsense like this. Astra perhaps, but not Fable. Oh man, the performance improved version that it was working on is actually ([View Highlight](https://read.readwise.io/read/01m388p4w2c39avyh75dsztraz)) - Note: [H4] Benchmarks finish before compaction, hiding long-run failures - Tags: atom/literature/h3