Claude Opus 5.5

In the first 24 hours after release, prominent developers praised the generous limits, personality, and clarity of responses; they criticized dense language, max mode, and mistakes on tricky questions. Below are practitioners' words in translation, with a date for each entry; all assessments are personal and use no common methodology, with our analysis and measurements in later sections.

Opus 5.5 can be connected the same day and compared on your own work with Opus 5 or Fable 5.1. It has no free tier, and its weights are closed.

Two of the six requests hit a hard no, both about money and hardware: Opus 5.5 is not available free or locally. Everything else is solved with a subscription or API.

You run long tasks in Claude Code: auditing someone else's project, a migration, rebasing dozens of branches. In Anthropic's Terminal-Bench 4.0 benchmarks, Opus 5.5 scores 66.4% versus 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 versus $0.25.

You make 3D scenes and games: Blender, Unreal, Three.js. In the first day, users showed a city in Unreal, a house from a photo in Blender, and a lens lab in an hour and a half; video examples are in chapter 3.

Your work involves vulnerabilities, exploits, or biology. Safety checks route such requests to Opus 4.8 and Opus 5, while extended access comes through separate review programs.

You have code for Opus 5 with reasoning disabled or a forced tool call. Both requests will return an error on Opus 5.5, so you will need to rewrite the calls.

Ways to get started are collected in chapter 4, pricing and access from Russia in chapter 6, and comparisons with Fable 5.1 and Astra in chapter 8. If you are just getting started with AI-assisted work, begin with an introduction to vibe coding.

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

An algorithm with a proof in 15 hours. Vals AI's 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 bounds. The post contains no preprint or code; this is the lab's claim (22.09.2026).

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 so in one prompt in 1 hour 26 minutes and $25,66 via API: lens.lab.sael.net is live, and the focus ring moves the focal plane through the scene. Figures and conditions are according to the creator (23.09.2026).

16 fixes through formal verification. Boris Cherny of Anthropic, who leads Claude Code, tasked the model with formally verifying Claude Agent SDK in Lean. A couple of short prompts produced 16 pull requests fixing bugs and race conditions, while he says he barely knows Lean. He is a vendor employee, so treat this case as a demonstration (22.09.2026).

San Francisco in Unreal with Jev. Matthew Berman tasked Opus 5.5 with recreating the city in Unreal Engine, while giving 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 creator does not say how long it took (22.09.2026).

A house from a photo and floor plans. Higgsfield gave the model one photo of a house and its floor plans. Opus 5.5 built a complete 3D model in Blender with a Python script, from site layout to the second-floor slab, and added an offline browser viewer with X-ray walls. This is a partner-platform demo (22.09.2026).

Auditing a large project. Opus 5.5 audited and fixed a 200-thousand-line codebase in under three hours, while Opus 5 took more than twenty on the same task. Token usage and provenance are in the figure below: this is an early tester's case from Anthropic's announcement, without independent verification (22.09.2026).

Our run. We gave both models the same task via API: a TypeScript function for parsing Russian dates with a test suite. Opus 5.5 answered in 22 seconds, Opus 5 in 37; both test suites passed for us without edits, while the new model's response bill was 43% lower (22.09.2026, default settings).

More. The model translated HAProxy from C to Rust in 9,5 hours versus 12 for Fable 5.1 and at 51% lower cost; it completed an Excel deal analysis in 63 minutes versus 93 for Opus 5; at low effort, it found 72% of bugs in a Deloitte review, while Opus 5 at high effort found 56%. A 680-thousand-line migration for one tester took under a day, and Christian Rivera of Stripe gave it a rebase of forty pull requests. From practitioners: Mia got 100 HTML pages with none broken, Shikhar builds a Three.js game at Medium effort, and Alex Albert of Anthropic built 1906 Market Street in Blender. All stories come from the 22.09.2026 announcement and posts from the following day; none has independent verification.

Where it stumbled. At maximum effort, Simon Willison says the model overthought a pelican breakage: it spent all 128 thousand output tokens without a result, at $2,56 and nearly 20 minutes per attempt, and he called max mode useless. Victor Taelin got either output cut off by the limit or a refusal under model-copying rules in the API while trying to build a Game Boy game. Yuchen Jin stayed with Astra: Opus 5.5 got his tricky question wrong, and he prefers Astra for code. Anthropic itself warns that the model thinks longer at the same effort and asks users to leave room in max_tokens (posts dated 22.09.2026 and 23.09.2026).

In the Claude app on paid plans, Opus 5.5 is the default and is listed simply as Opus. The free plan gets Sonnet and Haiku. Before a long task, check your remaining limit in settings: Anthropic raised five-hour limits for Pro, Max, Team, and seat-based Enterprise, and added one limit reset that can be saved for later.

In Claude Code. Update the client with claude update: Opus 5.5 requires version 2.1.280 or later. Then use /model opus in an open session or claude --model claude-opus-5-5 at launch. On Pro, Max, Team, and Enterprise, it is already the default model, so no changes are needed.

Effort. Opus 5.5 defaults to medium, while Opus 5 used high. At equal effort, the new model thinks longer, especially on xhigh and max, so Anthropic advises against carrying over the old setting and recommends trying the levels again. Effort is changed with the effort parameter in the API and a setting in Claude Code.

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

From Russia. Russia is not on Anthropic's official country list for either the app or the API. Installing the client and paying for a subscription do not change this on their own. Available intermediary routes and their terms are covered on the page about paying for AI from Russia; before paying, verify the exact model ID and data-retention rules.

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, while 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 its gap with Fable 5.1 is narrower than the scores suggest.

Independent index. Artificial Analysis measured the model before release and on announcement day ranked it first: 58 Intelligence Index points at maximum effort, versus 53 each for Fable 5.1 and Astra (22.09.2026). The model leads in six of the index's ten measurements, including Humanity's Last Exam and SciCode, and trails in CritPt, AA-LCR, and GDP.pdf.

In Artificial Analysis's Terminal-Bench 4.0 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 effort and run conditions differ. The high variant with fallback scored 54 points and cost $1,82 per index task.

Price of results. According to the same lab, Opus 5.5 at maximum effort uses about 119 thousand output tokens per index task, versus 73 thousand for Opus 5 and 27 thousand for Astra. Reduced rates kept task cost at Opus 5's level, and four of five effort levels sit on the “intelligence for money” frontier. Per index task at maximum effort, Opus 5.5 costs $5,98 and Astra $3,26. Anthropic calculates differently: at default effort, Terminal-Bench level with Astra costs it about 40% of Astra's task price.

Opus 5.5 has no RuBench run of its own yet: our fleet was running on Opus 5 on 22.09.2026, and the first paired measurement is described in chapter 3. The run result will appear on the profile after the client is updated.

API rates fell by a fifth from Opus 5: $4 per million input tokens and $20 for output. Cache reads dropped the most, from $0.50 to $0.20, and that is the main source of the promised 40% savings on long coding work.

Task price. Calculate by task, not by rate. Anthropic estimates 40% savings on typical work at default settings. Artificial Analysis calculated the same task price as Opus 5 because the new model writes 1.6 times more tokens at maximum effort. In our run at default effort, the response was 43% cheaper.

Subscription. There is no separate subscription for the model: Opus is included in Pro for $20 per month or $17 with annual billing, in Max from $100, and in Team for $20 per seat with annual billing. Free does not include Opus. The API is billed separately by token; the subscription does not cover its bill.

Limits in practice. Over five consecutive hours at high effort, Leighton Gott used 8% of the weekly limit, without naming the plan. On a $200 plan, Max Weinbach used 6% of his weekly allowance in six hours of model work. Freddy Vega from Platzi notes lower usage than Fable 5.1. These are observations from the first day; Anthropic does not publish its own usage figures (23.09.2026).

From Russia. Russia is not on Anthropic's list of supported countries. Paying through an intermediary does not change Anthropic's rules, and the intermediary has its own rates and limits. Restrictions and payment options are collected on a separate page.

Opus 5.5 launched with the same safeguards as Fable 5.1 because Anthropic places it alongside Mythos 5.1 in biology and cybersecurity. For ordinary coding work, this means one thing: another model may handle some requests, and the response will state who answered.

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

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

Maximum effort. The max mode does not speed up the response and may not produce one at all: Simon Willison's model spent its entire output limit on reasoning and returned nothing, while Viktor Taelin saw the same thing in the API even at medium effort. For work tasks, start with medium and high, and reserve max for cases where you are willing to wait twenty minutes and pay for an empty attempt (23.09.2026).

Other people's instructions on the page. Across all prompt injection tests, Opus 5.5 is no worse than Opus 5, and Anthropic calls it the most resilient Opus-class model. When cyber defense triggers, Opus 4.8 completes the response; its defenses against such attacks were strengthened after the release of Fable 5.1. The right to change data and send emails is still set by the application, so check what will require confirmation.

Data retention. Like earlier Opus models, this model is available with zero data retention by contract. For API accounts created from August 31, 2026, preserved thinking applies: the context around already received reasoning blocks cannot be changed, which is how Anthropic prevents distillation. Text is watermarked under EU AI Act requirements, and no-reasoning mode is disabled permanently.

Our experience. We build the site with agents, and as of 22.09.2026 its fleet runs on Opus 5: the installed Claude Code had not been updated to a version with Opus 5.5. We made our first test of the new model through the API that same evening; its result is in chapter 3. We will migrate the fleet after updating the client and running a paired test on our tasks.

Opus 5.5 or Fable 5.1. Anthropic itself recommends starting with Opus 5.5 and switching to Fable 5.1 when you need the most complex reasoning and very long autonomous work. Rates differ by 2.5 times, $4 and $20 versus $10 and $50, while cache reads are even cheaper with Opus 5.5. In the release table, Opus 5.5 leads on every row, and it is five points higher in the independent index. On his own tasks, Victor Taelin saw a clear gap over Fable 5.1, while noting that Anthropic trains models in his field.

Opus 5.5 or Astra. According to Artificial Analysis, the models are tied on Terminal-Bench 4.0; Opus 5.5 is higher in the overall index, while GPT-6 Astra costs 2.5 times more per token. Yet Astra is cheaper per Artificial Analysis index task, $3.26 versus $5.98 for Opus 5.5 at maximum effort, because it writes four times fewer tokens; Anthropic calculated the reverse at default effort, around 40% of Astra's price for the same Terminal-Bench. Astra leads on scientific Terminal-Bench-Science in Anthropic's own table. Which editor you use matters too: Claude lives in Claude Code, Astra in Codex. On release day, Simon Willison made both Sol and Opus 5.5 his default coding models; Yuchen Jin stayed with Astra.

Opus 5.5 or Opus 5. The new model is cheaper, faster, and ahead in every benchmark, so the only reason to stay on Opus 5 is code that expects it. Reasoning can no longer be disabled, forced tool calls are unsupported, the old computer tool is not accepted in the Claude API, and default effort has dropped to medium.

For a first comparison, give the models one familiar task and track time, number of edits, and the price of the accepted result. Our run with figures is in chapter 3; the Fable 5.1 profile collects its real-world work.

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