Claude Opus 5.5

In the first day after release, well-known developers praised the generous limits, the character and clarity of the answers. They criticized the dense language, max mode and mistakes on tricky questions. Below are practitioners' verbatim words, translated, with a date on each line. All assessments are personal and have no shared methodology. Our analysis and measurements follow in later chapters.

You can connect Opus 5.5 the same day and compare it with Opus 5 or Fable 5.1 on your own work. It has no free tier, and its weights are closed.

Two of the six requests hit a no, and both are about money and hardware: Opus 5.5 is not free and cannot run locally. The rest is handled by a subscription or the API.

You run long tasks in Claude Code: auditing someone else's project, migrations, rebasing dozens of branches. In Anthropic's Terminal-Bench 4.0 results, Opus 5.5 scores 66.4% against 55.8% for Fable 5.1 and 52.3% for Opus 5.

You paid for Fable 5.1 for quality and counted every token. Opus 5.5 rates are 2.5 times lower, and cache reads cost $0.20 per million against $0.25.

You build 3D scenes and games: Blender, Unreal, Three.js. In the first day of practice it produced a city in Unreal, a house from a photo in Blender, and a lens lab in an hour and a half; examples with videos are in chapter 3.

Your work involves vulnerabilities, exploits or biology. Safety checks hand such requests to Opus 4.8 and Opus 5, and extended access goes through separate vetting programs.

Your code on Opus 5 uses disabled reasoning or forced tool calls. On Opus 5.5 both requests will return an error, so you will need to rewrite the calls.

Ways to run it are in chapter 4, prices in chapter 6, and the comparison with Fable 5.1 and Astra in chapter 8. If you are only getting to know AI-assisted work, start with an introduction to vibecoding.

Opus 5.5 does the same work as Opus 5 and Fable 5.1: it reads materials, writes and edits code, and drives a browser and a terminal through the app. In the first day of practice it was shown on 3D scenes, formal code verification and long tasks, while Anthropic measured the difference in time and cost for the same work.

A proven algorithm in 15 hours. The Vals AI evaluation lab gave ten agents on Opus 5.5 the task of devising a faster shortest-path algorithm and proving it in Lean. After 15 hours they produced C-HD, a formally verified improvement on published evaluations. The post has no preprint or code; this is the lab's claim (Sep 22, 2026).

A lens lab in an hour and a half. Ryan Sael asked the model to explain camera focus and build an interactive lens lab. It did it in one prompt, 1 hour 26 minutes and $25.66 via the API: the site lens.lab.sael.net is live, and the focus ring moves the sharp plane through the scene. Figures and terms are as the author reports them (Sep 23, 2026).

16 fixes through formal verification. Boris Cherny of Anthropic, who leads Claude Code, asked the model to formally verify the Claude Agent SDK in Lean. A couple of short prompts produced 16 pull requests with fixes for bugs and race conditions, and he says he knows Lean poorly himself. He works for the vendor, so read this case as a demo (Sep 22, 2026).

San Francisco in Unreal with Jev. Matthew Berman tasked Opus 5.5 with recreating the city in Unreal Engine and handed the behavior of people, cars and animals to Jev. In the video the camera flies over an intersection with tram tracks, traffic lights and pedestrians. The author doesn't say how long it took (Sep 22, 2026).

A house from a photo and plans. Higgsfield gave the model one photo of a house and its floor plans. Opus 5.5 built a full 3D model in Blender with a Python script, from marking out the plot to the second-floor slab, and added an offline in-browser viewer with an X-ray view of the walls. This is a demo from a partner platform (Sep 22, 2026).

Audit of a large project. Opus 5.5 reviewed and fixed a 200,000-line codebase in under three hours, while Opus 5 spent more than twenty hours on the same work. Token use and provenance are in the figure below: this is a case from an early tester, cited in Anthropic's announcement, with no independent check (Sep 22, 2026).

Our run. We gave both models one task via the API: a TypeScript date-parsing function with a test suite. Opus 5.5 answered in 22 seconds, Opus 5 in 37; both test suites passed without edits, and the new model's cost per answer came out 43% lower (Sep 22, 2026, default settings).

More. The model ported HAProxy from C to Rust in 9.5 hours versus 12 for Fable 5.1, at 51% lower cost; it closed a deal breakdown in Excel in 63 minutes versus 93 for Opus 5; at low effort it found 72% of the errors in a Deloitte review, while Opus 5 at high effort found 56%. A migration of 680,000 lines at one tester's company took under a day, and Cristian Rivera of Stripe handed it a rebase of forty pull requests. From practitioners: Mia got 100 HTML pages with none broken, Shikhar is making a game in Three.js at Medium effort, and Alex Albert of Anthropic built Market Street in 1906 in Blender. All stories come from the Sep 22, 2026 announcement and posts from the next day; none has independent verification.

Where it tripped. At maximum effort, the model overthought its pelican task until it broke, by Simon Willison's account: it spent all 128,000 output tokens without a result, at $2.56 and nearly 20 minutes per attempt, and he called max mode useless. Victor Taelin, using the API, got output cut off by the limit once and a refusal under rules against copying models another time while trying to build a Game Boy game. Yuchen Jin stuck with Astra: Opus 5.5 got his tricky question wrong, and he prefers Astra's code. Anthropic itself warns that the model thinks longer at the same effort and asks to leave headroom in max_tokens (posts of Sep 22 and Sep 23, 2026).

In the Claude app on paid plans, Opus 5.5 is the default; in the model list it's labeled simply Opus. The free plan gets Sonnet and Haiku. Before a long task, check your remaining limit in settings. Anthropic raised the five-hour limits for Pro, Max, Team and Enterprise (seat-based billing) and added one limit reset you can save for later.

In Claude Code. Update the client with claude update: Opus 5.5 requires version 2.1.280 or newer. Then run /model opus in an open session or claude --model claude-opus-5-5 at launch. On Pro, Max, Team and Enterprise the default model is already this one, so nothing needs changing.

Effort. By default Opus 5.5 runs at medium; Opus 5 used high. At the same effort level, the new model thinks longer, especially at xhigh and max, so Anthropic advises not to carry over the old setting but to go through the levels again. Effort is changed with the effort parameter in the API and with the setting in Claude Code.

Via API. Exact ID claude-opus-5-5, on Bedrock anthropic.claude-opus-5-5. Reasoning is always on: a request with thinking disabled or a manual budget returns a 400 error, as does tool_choice with the values any and tool. The old computer tool computer_20251124 is not accepted on the Claude API and Google Cloud; use the computer_toolset_20260801 set instead.

In Anthropic's release table, Opus 5.5 leads Fable 5.1 and Opus 5 on every row. The terminal result was taken at xhigh effort, and the Astra row comes from OpenAI's publication at high effort, so these are each model's best results rather than one shared run. Anthropic itself adds that in its own use the gap with Fable 5.1 is narrower than the scores show.

Independent index. Artificial Analysis measured the model before release and ranked it first on announcement day: 58 points on the Intelligence Index at maximum effort, with Fable 5.1 and Astra at 53 each (Sep 22, 2026). The model leads in six of the ten index evaluations, including Humanity's Last Exam and SciCode, and trails in CritPt, AA-LCR and GDP.pdf.

On Terminal-Bench 4.0 in Artificial Analysis's run, Opus 5.5 scored 59.6%, level with Astra and 11 points above Opus 5; this is below the 66.4% in Anthropic's table, because the effort level and run conditions differ. The high-effort variant with fallback scored 54 points and cost $1.82 per index task.

Price of the result. Per the same lab, Opus 5.5 at maximum effort spends about 119,000 output tokens per index task, versus 73,000 for Opus 5 and 27,000 for Astra. Thanks to lower rates, the cost per task stayed at Opus 5's level, and four of the five effort levels landed on the “intelligence per dollar” frontier. At maximum effort, Opus 5.5 costs $5.98 per index task and Astra $3.26. Anthropic counts it differently: at default effort, Terminal-Bench level with Astra costs it about 40% of Astra's cost per task.

API rates fell by a fifth compared with Opus 5: $4 per million input tokens and $20 per million output. Cache reads got the biggest cut, from $0.50 to $0.20, and this is the main source of the promised 40% savings on long coding work.

Cost per task. Count by task, not by rate. Anthropic estimates savings on typical work at default settings at 40%. Artificial Analysis got the same cost per task as Opus 5, because the new model writes 1.6 times more tokens at maximum effort. In our run at default effort, the answer came out 43% cheaper.

Subscription. There is no separate subscription for the model: Opus is included in Pro for $20 a month or $17 billed annually, in Max from $100, and in Team for $20 per seat billed annually. Free has no Opus. The API is billed separately by tokens, and a subscription does not cover its bill.

Limits in practice. Layton Gott used 8% of the weekly limit over five hours in a row at high effort, without naming his plan. Max Weinbach spent 6% of the weekly allowance over six hours of work on the $200 plan. Freddie Vega from Platzi notes lower usage than Fable 5.1. These are observations from the first day; Anthropic does not publish its usage figures (Sep 23, 2026).

Opus 5.5 ships with the same safeguards as Fable 5.1, because in biology and cybersecurity Anthropic puts it on par with Mythos 5.1. For everyday coding work, this means one thing: some requests will be handled by a different model, and the response will show which model answered.

Behavior audit. On an automated audit of nearly 2,000 scenarios, the model showed less unwanted behavior than any recent Claude. In a new test of pushing past set boundaries, it tried to get around them about 85% less often than Opus 5 and Mythos 5.1. Anthropic calls all those attempts minor and self-reported. This is the developer's own assessment; external evaluators METR and Frontier Design reviewed the model before release.

A catch from the system card. The model follows harmful instructions hidden in text the user pasted into the request more often than its predecessors. In a training scenario with access to a package registry, it took potentially harmful actions in about half of the runs, and in conversations about surveillance it gave an appropriate answer 65% of the time versus 88% for Opus 5. Anthropic calls this an improvement over earlier models but still a worrying result.

Maximum effort. Max mode does not speed up the answer and may not produce one at all: Simon Willison's model spent the entire output limit on reasoning and returned nothing, and Victor Taelin saw the same on the API even at medium effort. For real work, start with medium and high, and keep max for cases where you are ready to wait twenty minutes and pay for an empty attempt (Sep 23, 2026).

Hidden instructions on a page. On all prompt injection tests, Opus 5.5 is no worse than Opus 5, and Anthropic calls it the most robust model in the Opus class. When the cyber safeguard triggers, the response is completed by Opus 4.8, whose protection against such attacks was strengthened after the Fable 5.1 release. The right to change data and send emails is still set in the app; check what will require confirmation.

Data retention. Like earlier Opus models, it is available with zero data retention under contract. For API accounts created from Aug 31, 2026, preserved thinking applies: you cannot change the context around reasoning blocks already received, which is how Anthropic makes distillation harder. Text is watermarked to meet EU AI Act requirements, and the no-reasoning mode is turned off permanently.

Our experience. We build the site with agents, and as of Sep 22, 2026 its fleet runs on Opus 5: the installed Claude Code lacked the update to a version with Opus 5.5. We ran the first test of the new model through the API that same evening; its result is in chapter 3. We will move the fleet over after updating the client and running a paired test on our own tasks.

Opus 5.5 or Fable 5.1. Anthropic itself advises starting with Opus 5.5 and moving to Fable 5.1 when you need the most demanding reasoning and very long autonomous work. Rates differ by 2.5 times, $4 and $20 against $10 and $50, and cache reads on Opus 5.5 are even cheaper. In the launch table, Opus 5.5 leads on every row, and in the independent index it is five points higher. Victor Taelin saw a clear gap over Fable 5.1 on his tasks and notes himself that Anthropic trains models on his domain.

Opus 5.5 or Astra. On Artificial Analysis' Terminal-Bench 4.0 the models are neck and neck, the overall index favors Opus 5.5, and per token GPT-6 Astra costs 2.5 times more. Per task on the Artificial Analysis index, though, Astra is cheaper: $3.26 against $5.98 for Opus 5.5 at maximum effort, because it writes four times fewer tokens. Anthropic, at default effort, counted the reverse: about 40% of Astra's price for the same Terminal-Bench. Astra leads on the science benchmark Terminal-Bench-Science in Anthropic's own table. Which editor you use also decides it: Claude lives in Claude Code, Astra in Codex. On launch day Simon Willison made both Sol and Opus 5.5 the default models for code; Yuchen Jin stayed with Astra.

Opus 5.5 or Opus 5. The new model is cheaper, faster and ahead in every benchmark, so staying on Opus 5 only makes sense for code that depends on it. Reasoning can no longer be turned off, forced tool calls are not supported, the old computer-use tool is not accepted on the Claude API, and the default effort has dropped to medium.

Opus 5.5 or Sonnet 5.5. Since Sep 28, 2026, a mid-tier model has joined Opus: Sonnet 5.5 costs half as much per token and trails by two points on the Artificial Analysis index. Anthropic recommends it for well-defined tasks, and Opus 5.5 for complex work that requires judgment.

For a first comparison, give the models one familiar task and count the time, the number of edits and the price of the accepted result. Our run with the numbers is in chapter 3, and the Fable 5.1 card collects its real work samples.

Two months passed between Opus 5 and Opus 5.5. In between, Anthropic changed its tone: ten days before the release, Dario Amodei called for slowing the race in capabilities, and Opus 5.5 became the first model after that call. Hidden tests in the Claude web app were spotted a day before the announcement.