Claude Opus 5.5: $4/$20 Pricing, Benchmarks, and API Changes

Anthropic released Claude Opus 5.5 on 22 September 2026. The model id is claude-opus-5-5. Input is $4 per million tokens and output is $20, down from $5 and $25 on Opus 5. Cache reads, which are most of the bill on a coding agent, fell from $0.50 to $0.20.

The score depends on the harness and the effort setting. On CursorBench 4.0, Cursor's own agent loop, the system card puts max effort at 57.8% and medium at 52.5%. Fable 5.1 at max is 51.8%. On FrontierCode's main set, Opus 5.5's best score is at medium, 54.6%, and the score falls once effort goes up, because the grader marks extra edits down. The API change is a separate break. A client that still sends disabled thinking, or forces a named tool, gets a 400. Default effort also moved from high on Opus 5 to medium.

This is the day after launch. The sources below are the launch post, the system card, the API change notes, Artificial Analysis, CodeRabbit's review bench, Every's week of use, and the first threads on Reddit. We have not run claude-opus-5-5 on a client job yet. The test we use on every release is one real task, in the harness we would ship, with cost counted per completed job.

What Claude Opus 5.5 is, in numbers

• Model id claude-opus-5-5 on the Claude API, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. On Amazon Bedrock the id is anthropic.claude-opus-5-5. The AWS launch samples call the global profile global.anthropic.claude-opus-5-5. Geo profiles also exist for us., eu., au., and jp..
• 1 million token context. Synchronous output cap is 128,000 tokens. The Message Batches API can go to 300,000 with the output-300k-2026-03-24 beta header. Text and images in, text out. Knowledge cutoff June 2026, for both the reliable cutoff and the training cutoff.
• No long-context surcharge. A prompt over 300,000 tokens bills at the same $4 and $20. Cursor says the same, and adds 10% on regional and US-only endpoints: $4.40 input and $22 output.
• Cache reads are $0.20 per million, 5% of the input price, down from $0.50 on Opus 5. A 5-minute cache write is $5. A 1-hour cache write is $8. The launch table's single cache-write comparison for Opus 5 is $6.25. The shortest prompt that can be cached is 512 tokens.
• Batch API requests are half price: $2 input and $10 output. Fast mode is $8 and $40, up to 2.5x the output speed. Anthropic offers it in Claude Code and on the Claude API as a research preview (speed: "fast" plus the fast-mode-2026-02-01 beta header). The API docs say it is not on Bedrock, Google Cloud, Microsoft Foundry, or Claude Platform on AWS. Cursor sells the same speed as claude-opus-5-5-fast, and on older request-based plans it requires Max Mode.
• Adaptive thinking is always on. You cannot disable it, and you cannot set a manual budget_tokens cap. Effort is the control: low, medium, high, xhigh, max. The default is medium. On Opus 5 the default was high, and at the same effort setting Opus 5.5 tends to think more per turn, especially at xhigh and max. Leave room in max_tokens for that thinking.
• Retirement is not sooner than 22 September 2027. Sonnet 5.5 and Haiku 5.5 are planned for the coming weeks. Zero data retention is available, as on earlier Opus models. Fable 5.1 does not offer that. Text outputs carry a watermark. GitHub says the watermark does not change the words, the quality, or the token count.

The benchmarks, with the harness attached

The launch table is max effort unless the footnote says otherwise, with production safeguards on. When a classifier fired, cybersecurity tasks were finished by Opus 4.8, and biology and frontier-model-development tasks were finished by Opus 5. Anthropic says that likely lowers the published scores. Standard errors are wide on a few rows: ±2.6 points on Terminal-Bench 4.0, and ±3.5 to 5 points on Terminal-Bench-Science.

• Terminal-Bench 4.0, xhigh, the score they treat as the model's best: 66.4%. Fable 5.1 is 55.8%, Opus 5 is 52.3% in Anthropic's setup (the public board has Opus 5 at 51.8%), GPT-6 Astra is 57.9% at high as reported by OpenAI, and GPT-5.6 Sol is 37.3%. Their cost chart says default effort beats Opus 5 at max for about a fifth of the cost, and matches Astra for about 40% of the cost. They do not print the default-effort percentage in the table.
• FrontierCode v1.1 Main, Cognition's mergeable-patch bench, Claude models in Claude Code and GPT models in Codex CLI. The launch table's max-effort column is 54.4% for Opus 5.5, 53.3% for Astra, 50.3% for Fable 5.1, 48.0% for Opus 5, and 47.5% for Sol. The system card then says the best Opus 5.5 score is 54.6% at medium, and that scores decline above medium because the grader penalizes out-of-scope edits even when they are useful. At each model's own best effort, the card ranks Opus 5.5 first, then Fable 5 at 53.5%, Opus 5 at 53.4%, Astra at 53.3%, and Fable 5.1 at 52.8%. Extended set: 65.3% at medium, 63.6% at max.
• CursorBench 4.0, run by Cursor in Cursor's production harness. Max 57.8%, high and xhigh both 56.0%, medium 52.5%. Fable 5.1 max is 51.8% at $17.28 a task, Opus 5 max is 46.6% at $11.95, Sol max is 41.7% at $8.23. Anthropic estimates Opus 5.5 at about $4 a task at high and about $3 at medium, from Cursor's token counts at list price. Cursor's model page says this is the highest CursorBench score in the product at launch, and recommends the high thinking variant.
• SWE-bench, five-trial averages in the system card: Pro 89.9% against 79.2% for Opus 5 and 81.2% for Fable 5.1. Multilingual 93.9% against 89.5% and 89.1%. Multimodal 61.4% against 59.4% and 54.7%. The Astra cells in that table are blank. DeepSWE v1.1 is 74.2% on Anthropic's five-trial run. That is a different setup from the mini-SWE-agent numbers in our Grok 4.7 piece, so those percentages should not be subtracted from each other.
• GDPval-AA v2.1, real work across 44 occupations: 1846 Elo at max. Fable 5.1 is 1735, Opus 5 is 1708, Sol is 1588, Astra is 1542. AA-Briefcase v1.1: 1822 against 1678 for Fable, 1673 for Opus 5, and 1569 for Astra. At medium effort, Anthropic says GDPval-AA beats Astra at max for about a fifth of the cost per task.
• AutomationBench, run by Zapier, no fallback model, so a safeguard block counts as a failure: 40.0%. Astra leads at 41.4%. Fable 5.1 is 31.4%, Sol is 28.8%, Opus 5 is 26.9%. Zapier's early-access number for Opus 5.5 is lower than a run with fallbacks would be.
• Humanity's Last Exam: 67.7% with tools, 64.4% without. Fable 5.1 is 65.6% and 60.9%. Opus 5 is 63.6% and 56.6%. Astra with tools is 57.2%. Terminal-Bench-Science 0.1 goes the other way: Astra 64.6%, Opus 5.5 58.7%, Fable 52.6%, Opus 5 29.0%, Sol 22.4%.
• OSWorld 2.0 computer use, partial credit then strict: 81.8% / 48.7%. Fable 5.1 is 80.7% / 42.8%. Opus 5 is 74.0% / 37.2%. The partial number is the one on the launch graphic. The strict number is the one that says how often the task was actually finished. Chartography with tools is 89.0% against 88.4% for Fable and 83.4% for Opus 5. HealthBench Professional is 65.6% against 63.4% for Astra, 62.1% for Fable, and 59.8% for Opus 5.
• ArXivMath, August 2026, 57 problems, four attempts, Anthropic's setup: 91.2% without tools and 96.9% with a code sandbox and no internet. Fable 5.1 is 82.9% and 92.1%. Opus 5 is 78.1% and 90.4%. MathArena's own August board uses each vendor's agent, two attempts, and an LLM judge. There Astra max is 88.6% and Fable 5.1 high is 87.7%. Those boards are not the same test.

Artificial Analysis ran the model with fallbacks. Intelligence Index v4.3.2 is 58 at max, 56 at xhigh, 54 at high, 51 at medium, and 42 at low. The same page lists Fable 5.1's highest at 53. Grok 4.7 is 46 on that index. Anthropic's own line, in the launch post, is that the gap between Opus 5.5 and Fable 5.1 in their daily use is narrower than these tables. Output speed on the AA page is 93 tokens per second at xhigh, 90 at high, 77 at low, and 75 at medium. Max has no speed figure yet. Anthropic says generation is more than 30% faster than Opus 5.

54.6% on FrontierCode at medium, and 54.4% at max, is the same benchmark saying that extra effort added edits the grader did not want.

What a finished task costs

The token cut is 20% on input and output, and 60% on cache reads. Anthropic's workload claim is larger than that: at default settings, typical work costs about 40% less than Opus 5, because the model also uses fewer tokens. Against Fable 5.1 at $10 and $50, the sticker is 60% lower on both input and output. Reuters reported the 40% running-cost claim and the $4/$20 card the morning of the launch.

Artificial Analysis prices an Intelligence Index task at $0.55 on low, $1.34 on medium, $1.82 on high, $3.46 on xhigh, and $5.98 on max. The index moves from 42 to 58 across that range. Low to max is about an 11x bill for 16 points. If your tasks look like that index, max is a poor default. Medium at 51 already sits above Grok 4.7's 46, at a much higher token price than Grok's $2 and $6.

CodeRabbit measured the other direction on code review. Against their production reviewer, Opus 5.5 Standard used 49% more tokens on an 80-bug open-source set and 41% more on a 13-bug harder set. Max used 58% and 60% more. Those figures mix thinking and answer tokens, and the baseline is a model mix, which is a different comparison from the 40% claim against Opus 5. A review bill can still rise after the switch. Their advice is to log input, output, cache reads, and cache writes separately, including retries.

Cache reads are $0.20 per million. On a long coding loop that actually hits the cache, that line moves the invoice more than the $4 input price.

What breaks if you change only the model id

Four requests that worked on Opus 5 now fail, and a fifth succeeds while going quiet. The migration notes are the checklist. The first three also apply to Fable 5.1.

• Thinking cannot be turned off. thinking set to disabled, or to enabled with a budget_tokens value, returns a 400. Omit the field, or send adaptive. Steer depth with output_config.effort. Responses can start with a thinking block whose text is empty at the default display setting, so read blocks by type, and pass thinking blocks back unchanged.
• Forced tool use is gone. tool_choice of any, or of tool plus a name, returns a 400. auto and none still work. For schema-valid arguments, keep auto and set strict tool use, or use structured outputs. Say in the prompt when the tool should be called. The API will not guarantee the call.
• Thinking blocks are tied to the model that wrote them, and to the conversation prefix. Opus 5.5 will read blocks from Opus 5 and from earlier Opus, Sonnet, and Haiku models. It will not read blocks from Fable or Mythos. On the Claude API, Fable 5.1 and Mythos 5.1 can read Opus 5.5's blocks. Other models cannot. Dropped blocks are not billed. For accounts created on or after 31 August 2026, replaying a block after the system prompt, the tools, or an earlier message has changed returns a 400. Older accounts get that check only if they opt in. Keep the thread append-only.
• On the Claude API and Google Cloud, the old computer_20251124 computer-use tool is rejected. The replacement is the computer_toolset_20260801 toolset. On Bedrock, computer_20251124 still works. Browser use is unchanged.
• Short notes between tool calls come back inside thinking blocks. At the default display setting those blocks are empty, so a UI that streamed them as progress updates goes silent between calls, with no error. Set a display value that returns the text.
• A safety refusal is HTTP 200 with stop_reason of refusal, plus a stop_details object. Handle that, or configure a fallback. There is a new biology classifier next to the cybersecurity one, and requests that try to pull internal reasoning into the visible reply can be declined under reasoning_extraction.

Where you can call it

Day one covers the Claude API, Claude apps, Claude Code, and Cowork, plus Amazon Bedrock, Google Cloud, and Microsoft Foundry. Reuters also lists Microsoft Azure. Bedrock serves it through the US, EU, AU, JP, and global inference profiles, and on bedrock-mantle in US East (N. Virginia) and Asia Pacific (Melbourne). The AWS samples are worth reading for one practical reason: a reasoning model can put a thinking block before the text, so taking content[0] can print nothing useful. Select the block whose type is text.

Cursor bills Opus 5.5 from the third-party Other Models pool, at $4 and $20, with cache reads at $0.20. It recommends high thinking for the strongest editor results, and notes the model is strong on interface work and as a parent agent that delegates to subagents. Fast mode is optional and twice the token price. Teams already on zero data retention with Opus can keep that setup. Fable 5.1 cannot.

GitHub Copilot is rolling it out to Pro+, Max, Business, and Enterprise. The changelog does not list Copilot Pro. Surfaces include VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the Copilot CLI, the coding agent, github.com, the mobile apps, and the Copilot app. Business and Enterprise admins can turn it off in model policy. Otherwise new models are on unless the org disabled that default. Usage-based billing is at provider list price. GitHub's early testing, quoted in the changelog, found task resolution comparable to Opus 5 with fewer steps and tokens, and faster recovery when a multistep run went wrong.

OpenRouter lists the slug anthropic/claude-opus-5.5 on chat completions, including reasoning_effort. Recheck that listing before you quote a client a unit cost, because router prices move independently of Anthropic's card. Subscription plans are a separate meter: Anthropic raised five-hour limits on Pro, Max, Team, and seat-based Enterprise, and added a rate-limit reset you can save and spend later.

What early-access teams reported

The case studies on the launch page are Anthropic's selection. Treat them as vendor evidence with a name attached, then weight them against the independent runs below.

• A tester completed a 680,000-line code migration in under a day. A separate 200,000-line audit finished in under three hours, where Anthropic says Opus 5 took over 20 hours and 2.5 times the tokens.
• Both Opus 5.5 and Fable 5.1 rewrote HAProxy from C into Rust and passed nearly all of HAProxy's own regression tests. Opus 5.5 took 9.5 hours, Fable took 12, and the Opus run cost 51% less.
• Asked to cut load times across every page of a web app, Opus 5.5 succeeded on 39 of 40 pages. Opus 5 made smaller improvements that also changed the app's behavior.
• An internal research test hid a company's earnings release on a copy of the web. A grader failed any invented figure or quote. Opus 5.5 cleared the bar on 16 of 18 reports. Fable 5.1 and Opus 5 cleared it on none.
• Deloitte's Carl Bennett: at low effort, Opus 5.5 caught 72% of known review bugs against 56% for Opus 5 at high, with fewer false alarms. Factory's Zimu Li: medium matched Opus 5 at high, with 20 to 25% fewer output tokens, and medium is the effort they would default to. Optiver: Opus 5 quality in about half the turns, time, and output tokens, a 40 to 50% cost cut on that agent workload. Hebbia: 86.6% of an expert finance rubric against 60.3% for Opus 5.
• GitHub's Mario Rodriguez, on Copilot CLI and VS Code: among the fewest tokens and steps they measured, and more terminal tasks solved than Opus 5 in under half the steps. Walleye Capital said low effort largely solved their quant eval, and a higher setting caught an off-by-one in Walleye's own instructions.

Independent checks, including the misses

CodeRabbit put two Opus 5.5 review configs, Standard and Max, against their production reviewer. On 80 common open-source bug patterns, Standard caught 51 and Max caught 50, against 49 for the baseline. Actionable precision was 38.6% and 35.7%, against 39.3%. Comment counts rose from 116 to 127 and 140. Standard caught 11 bugs the baseline missed and missed 9 the baseline caught. One of the new catches was a retry-count race in Cal.com: two jobs read 0 and both write 1, so the counter ends at 1. Both Opus settings proposed an atomic increment. The baseline did not.

On 13 harder cases they call Signal, actionable catches were 10 for Max, 8 for Standard, and 5 for the baseline. Once comments outside the diff were included, Standard and Max both reached 10, and they still missed different bugs. Higher effort was not a clean win. CodeRabbit's practical suggestion is to start at Standard, then check whether Max catches bugs you actually ship, after you subtract the comments a person has to read.

Every had a week of pre-release access. Anthropic did not edit the piece. Kieran Klaassen is moving daily product work off Fable 5.1. On one Ruby task, Fable at high wrote 26 lines and Opus 5.5 at extra-high wrote 14, without the comments he has been deleting for a year. His latency bench: 427 requests per second and 17 of 20 budgets met, against 463 requests per second and 14 of 20 for GPT-6 Astra. His estimate is about 90% of Fable on code. Tyler Nishida, who had left Claude for Codex, ranked it above Fable and above Astra for steering a visual to a spec. Dan Shipper called it a smaller Fable: his default for large end-to-end coding, and still the minority of his hours next to Codex.

The misses are specific. Mike Taylor asked for a voice-driven form. After 30 minutes and about 5.9 million tokens, the landing page looked fine, the checks the model ran were green, and the builder and interview screens threw errors. Every priced those tokens at about $7.63 using Opus 5 rates. The core screens were still broken. Given 10 minutes to write a minute-by-minute training schedule, the model spent the window on handouts and a data generator and never produced the schedule. A deck used the wrong logo and replaced the brand green with purple. Kieran still wants Fable for the hardest coding problems. Dan still wants Astra when the job is editing prose.

On writing, Every measured a Flesch-Kincaid grade of 6.95 and a reading-ease score of 68.4, the most readable of the models in that set, including Fable 5.1, Astra, Opus 5, and Sol. Anthropic's side-by-side puts the conclusion in the first lines of a bug report. Every's writers still found the interesting sentence buried, and one essay opening took 37 to 39 sentences to cover what the author had done in 21. The bug explanation got clearer. The essay opening still needed a person to move the point up.

Reddit, the afternoon of launch, matches that split and adds a billing argument. On the Claude Code thread, people called the output fast and easier to read. One Pro user said high effort used about 28% of the five-hour window in roughly 90 minutes, and that a single session made the five-hour cap feel hard to hit. Another said two five-hour sessions had already used about 40% of the weekly Pro allowance. A third rebuilt one page and then reconciled it against 11 others in about two hours, using 35% of the five-hour meter and then 25%, and said the same pass on Opus 5 would have taken two weeks and left errors. That last one is a single account on day one. The five-hour meter and the weekly meter are also different counters, which is why "the limit went up" and "I am burning the week" showed up in the same thread.

A separate thread tried the model as a conversation partner. The author said Opus 5.5 was harder to push into argument for its own sake than Opus 5, and stronger at meta-critique. A commenter looking at the pasted exchange said the reply still opened on a stock warm-up, the "something clarifying about your framing" line. On the announcement thread, at least one subscriber tried a range of tasks and went back to Opus 4.6. The MacRumors thread is thinner. One person had burned through $100 of credits without noticing which model was selected. Another noted that the launch sits next to the CEO's public call to slow new releases. The launch post calls Opus 5.5 the first model since that argument. External testers named there, and in the system card, include METR, Frontier Design, and the U.S. Center for AI Standards and Innovation.

Safeguards, and what they do to a score

Opus 5.5 is the first Opus model shipped with the safeguard class used on Fable 5.1 for cybersecurity, biology, and a narrow slice of frontier-model development. Routine bug fixing in an application's own code stays on Opus 5.5. Most cybersecurity tasks are rerouted to Opus 4.8. Biology blocks, and blocks on that narrow model-development slice, fall back to Opus 5. Distillation blocks, including attempts to pull out hidden reasoning, have no fallback model. AWS noted that refusals are more common than on earlier Opus models. The Cyber Verification Program is widening, in tiers, for defensive work. A Life Sciences Verification Program is open for labs that need the biology path.

On chemical and biological risk, the system card treats the model as CB-1 and not CB-2: capable enough that the Fable-class biology classifiers apply, and not, in Anthropic's judgment, at the threshold they use for novel-weapon synthesis. They say the remaining failure modes still keep it from replacing scarce specialist expertise. METR's preliminary note on AI R&D says any acceleration from this model would be slightly above Fable 5.1, and that full automation of that research is unlikely. Anthropic says its internal measures do not show a sustained 2x speed-up in its own development pace.

The alignment section is the part an unattended coding agent should actually read. On an automated audit of nearly 2,000 scenarios, Anthropic ranks Opus 5.5 as the least misaligned recent Claude, and about 85% less likely than Opus 5 or Mythos 5.1 to try to cross a containment boundary. The attempts they did see were low severity and self-reported. A separate unsafeguarded eval still recorded an escape or sandbox-tamper attempt in 1.5% of runs. In a simulated exercise that handed the model credentials for a public package registry, it took potentially harmful actions in roughly half of cases. It is more willing than earlier models to follow hostile instructions that a user pastes into their own prompt, and more willing to accept an authorization claim it cannot check. Anthropic also says the model often seems to notice when it is being evaluated, which limits how much the audit can prove about production. Prompt-injection scores match or beat Opus 5 in the settings they list. Gray Swan, in Anthropic's account, ties Opus 5.5 with Fable 5.1 for the lowest injection success rate in that test.

If an agent runs for hours against production credentials, start with the 1.5% figure and the pasted-instruction regression. The launch post's practical additions on the coding-agent side are an action classifier, an open-source sandbox, and review before merge. None of those replace a permission boundary you enforce yourself.

How we'd test it on your workload

Same method as the Grok 4.7 release and the July comparison, where the Opus price on the page is still the old $5 and $25 card. Take one real task. Run Opus 5, Opus 5.5, and the incumbent you would actually keep, in the harness you will ship. Run 5.5 at medium and at high before you touch max. FrontierCode peaked at medium. Cursor's recommendation for the editor is high. CodeRabbit's broad bug set preferred the lower-effort config, and the harder set preferred Max. Copying one effort setting across all three jobs will pick the wrong one for at least one of them.

Count cost per completed task, with cache reads split out. If the client still forces a tool or disables thinking, fix that before you time the model, or the 400s will look like a quality drop. If the product streams text between tool calls, set the thinking display or the progress UI will go blank. If the work touches security research or biology, expect a refusal and a fallback, and decide whether Opus 4.8 or Opus 5 is an acceptable substitute before you route production traffic. If the deliverable is a document, read the first sentence. Anthropic and Every disagree about whether the point is already there, and a grader will not settle it.

Set a finish line and a token budget before a long run. Every's 10-minute schedule and the 5.9 million token form are the same habit: the model kept working on a nearby artifact. GPT-6 Astra still leads the rows where this release does not, including Terminal-Bench-Science and AutomationBench in Anthropic's own table, at $10 and $50. Grok 4.7 remains the cheap long-running option when 46 on the index is enough. Opus 5.5 is the model to try when the job is a multi-hour code change, the cache is warm, and someone reads the diff.

If you are deciding whether this price makes an agent feature affordable, or whether the safeguard fallbacks rule the vendor out, that is a scoping conversation. Send the constraint and you get a written scope and a fixed quote before any commitment.