Claude Opus 5.5: Complete Guide to Specs, Pricing and Benchmarks (2026)

Anthropic shipped Claude Opus 5.5 on 22 September 2026, about two months after Opus 5 (24 July 2026). It is a price cut and a capability jump in the same release. List prices drop 20%, cache reads drop 60%, and Anthropic's own capability table (Table 8.1.A in the system card) shows it ahead of both Opus 5 and the more expensive Claude Fable 5.1 on every row. Anthropic's model docs now say to start with Opus 5.5 for most workloads and keep Fable 5.1 for the hardest long-horizon reasoning.

This guide covers the facts you need to decide whether to use it: model IDs on every platform, limits, full pricing against Opus 5 and Fable 5.1, how the new effort default changes your bills, every benchmark Anthropic published (plus the ones it didn't), the Artificial Analysis result and its token-usage catch, and what the 230-page system card says about safety. Every number below is from Anthropic's docs, announcement or system card unless it is marked as third-party.

What is Claude Opus 5.5?

Claude Opus 5.5 is Anthropic's mid-cycle update to the Opus line, positioned on the model page as the model "for long-running agentic coding and knowledge work." It sits between Claude Sonnet 5 ($2/$10) and Claude Fable 5.1 ($10/$50) on price, but on Anthropic's published benchmarks it now beats Fable 5.1 on most rows.

Three things set it apart from Claude Opus 5:

• It is cheaper. $4/$20 instead of $5/$25, with cache reads at $0.20 instead of $0.50. Anthropic says the model costs "40% less than Opus 5 on typical workloads" and generates output more than 30% faster.
• It needs less effort for the same work. The default effort drops from high to medium, and Anthropic says medium matches or beats Opus 5 at higher settings.
• Thinking is always on. Adaptive thinking can no longer be disabled, and forced tool use is gone. Both are breaking API changes, covered in our Opus 5.5 migration guide.

Anthropic also says Claude Sonnet 5.5 and Claude Haiku 5.5 "will follow in the coming weeks."

Claude Opus 5.5 specs and model IDs

The API ID is claude-opus-5-5. Like claude-opus-5, it is a fixed ID with no date suffix. Bedrock is the only platform with a different string.

One practical note on the context window: Claude models since Opus 4.7 use a newer tokenizer that produces roughly 30% more tokens for the same text. A 1M window holds about 555k words, not the 750k older models fit. Budget accordingly if you are porting long-document prompts from a pre-4.7 model.

How much does Claude Opus 5.5 cost?

$4 per million input tokens and $20 per million output tokens on the Claude API. Here is the full price sheet against the two models you are most likely deciding between, from Anthropic's pricing page (USD per million tokens):

A few details that affect real bills:

• Cache reads use a special multiplier. Most Claude models charge 10% of the input price for a cache hit. Opus 5.5 charges 5%, which puts its cache reads at the same $0.20 as Sonnet 5. For agent loops that re-read a large, stable prefix every turn, this is the biggest saving in the release.
• No long-context premium. The full 1M window bills at the standard rate. A 900k-token request costs the same per token as a 9k-token one.
• Fast mode is first-party only. It is a research preview at $8/$40, enabled with speed: "fast" and the fast-mode-2026-02-01 beta header. Anthropic's announcement says it runs up to 2.5x faster. It is not available on Bedrock, Vertex AI, Foundry or Claude Platform on AWS, and it does not combine with the Batch API.
• US-only inference costs 10% more. Setting inference_geo: "us" applies a 1.1x multiplier to every token category.

For a deeper look at how Opus-tier pricing interacts with effort, see our earlier Opus 5 pricing and effort breakdown. The mechanics carry over; the rates and default do not.

What effort level should you use?

Opus 5.5 has five effort levels: low, medium, high, xhigh and max. A request that omits effort now runs at medium. On Opus 5 it ran at high.

Two changes interact here, and they point in opposite directions:

1. The default is lower, so an unchanged integration will usually spend fewer tokens per call.
2. The model thinks more per turn at any given level than Opus 5 did, "most of all at xhigh and max," per the what's-new page. If you had pinned max on Opus 5, the same setting on Opus 5.5 can cost more, not less.

Anthropic's advice is to re-run your effort sweep rather than carry a setting over, and to leave room in max_tokens for thinking. The system card backs up the case for staying low. On CursorBench 4.0, Opus 5.5 scored 57.8% at max effort, 56.0% at high for about $4 per task, and 52.5% at medium for about $3 per task. Even the medium result beats Fable 5.1 at max effort (51.8% at $17.28 per task). On FrontierCode, Opus 5.5's best score (54.6% on Main) came at medium, and scores dipped above medium before mostly recovering at max. Coding-agent company Factory is quoted in the announcement calling it the "first model we'd default to at medium effort."

A minimal Python call with an explicit effort level:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=32000,                 # leave headroom for thinking
    output_config={"effort": "high"},   # low | medium | high | xhigh | max
    messages=[{"role": "user", "content": "Review this diff for concurrency bugs: ..."}],
)

# Responses can open with thinking blocks, so select by type, not position
text = "".join(b.text for b in response.content if b.type == "text")
print(text)

Do not send thinking: {"type": "disabled"} or a manual budget_tokens. Both return a 400 on Opus 5.5. If you need low latency, use low effort, fast mode, or Sonnet 5.

Where is Claude Opus 5.5 available?

• API and clouds: the Claude API (all customers), Amazon Bedrock, Claude Platform on AWS, Google Cloud Vertex AI and Microsoft Foundry, all from launch day.
• Claude apps: the announcement lists the Pro, Max, Team and Enterprise plans, and says five-hour usage limits went up on each. It also introduces a "rate limit reset" that subscribers can save and use when they choose. We could not confirm from Anthropic's pages whether Free users get Opus 5.5.
• Claude Code: Opus 5.5 is available in Claude Code. At the time of writing, Anthropic's Claude Code help article did not yet state whether it replaces Opus 5 as the default, so check /model in your install. For migrating an existing codebase, the migration guide suggests running /claude-api migrate this project to claude-opus-5-5 inside Claude Code.
• Zero data retention is available, as with earlier Opus models.

Claude Opus 5.5 benchmarks

This is Anthropic's capability summary (Table 8.1.A, system card p.174), with GPT-5.6 Sol numbers added from the announcement where Anthropic published them. Claude results use adaptive thinking at max effort averaged over five trials, except Terminal-Bench 4.0, which is at xhigh. Competitor figures are taken from each vendor's system card or public leaderboards, not re-run by Anthropic. A dash means no number was published.

Sources: Claude Opus 5.5 system card, §8; GPT-5.6 Sol and CursorBench rows from the announcement. CursorBench scores were measured and reported by Cursor; Opus 5.5's per-task costs are Anthropic's estimates from Cursor's token counts.

What the table shows:

• Coding is the headline. SWE-bench Pro jumps 10.7 points over Opus 5, and Terminal-Bench 4.0 jumps 14.1 points to a new state of the art. The Terminal-Bench figure (66.36%) was run in Claude Code's --bare mode at xhigh with safeguards on; at max effort it scored 64.8%, within the ±2.6-point standard error.
• Vision improved the most. On Chartography without tools, Opus 5.5 scored 64.4% against 44.8% for Fable 5.1 and 29.8% for Opus 5. It is the largest single jump in the card, and it matters if you feed the model dashboards, screenshots or diagrams.
• GPT-6 Astra still wins some rows. It leads on Terminal-Bench-Science (64.6 vs 58.7) and AutomationBench (41.4 vs 40.0). On FrontierSWE v2, reported elsewhere in the card, GPT-6 Astra scored 65.5% and Opus 5.5 62.3%.
• Fable 5.1 is not obsolete. It stays ahead on OfficeQA (80.2% vs 78.9%) and OfficeQA Pro (69.0% vs 67.7%). Anthropic itself says that in internal use "the gap between Opus 5.5 and Fable 5.1 is narrower than the scores suggest."

Other results from §8 of the card: 74.2% on DeepSWE v1.1; 91.2% on ArXivMath (August 2026) without tools and 96.9% with tools (Fable 5.1: 82.9% and 92.1%); 91.2% on the long-context ProgramBench (Fable 5.1: 87.6%, Opus 5: 85.4%); 77.8% pass@1 on Toolathlon Verified; and 94.3% on the 42-language GMMLU.

For how to read these benchmark families, and why version numbers matter, our Opus 5 benchmarks explainer still applies. Terminal-Bench is now v4.0 and OSWorld is 2.0, so older scores are not comparable.

Which benchmarks were not published?

Anthropic did not report SWE-bench Verified, GPQA Diamond, MMMLU, tau-bench or ARC-AGI results for Opus 5.5 in the announcement or the system card. SWE-bench Verified has been replaced by SWE-bench Pro, Multilingual and Multimodal. As of launch day, the ARC Prize site had no Opus 5.5 entry. If you see an Opus 5.5 SWE-bench Verified score elsewhere, it did not come from Anthropic.

How does Opus 5.5 rank on the Artificial Analysis Intelligence Index?

First place. Independent benchmarker Artificial Analysis scores Opus 5.5 at 58 on its Intelligence Index (v4.3.2, max effort), ranked #1 of 212 models. According to Artificial Analysis's results as reported by OfficeChai, that is five points ahead of GPT-6 Astra and Fable 5.1 (both 53), and seven ahead of Opus 5 (51).

The right-hand column is the catch. At max effort Opus 5.5 generated about 119k output tokens per index task, more than any model in the comparison and over four times GPT-6 Astra. Running the full index took 260M output tokens and cost $8,708.20. The lower price roughly cancels the extra tokens, so Opus 5.5 lands at about the same cost per task as Opus 5 at max effort. The "40% cheaper" claim holds at medium and high, where most of the capability already is. It does not hold if you run everything at max.

Also note that Artificial Analysis runs its own harnesses. Its Humanity's Last Exam score for Opus 5.5 is 61.4% and its Terminal-Bench 4.0 score is 59.6% (tied with GPT-6 Astra), lower than Anthropic's 67.7% and 66.4%. Neither is wrong. They are different setups, so say which one you are quoting.

What's new in Claude Opus 5.5 vs Opus 5?

Beyond the API surface, Anthropic says Opus 5.5 reaches the same quality in fewer tokens and steps. Customer figures in the announcement include Optiver matching Opus 5's quality "in about half the turns, time and output tokens," Kiro seeing about 40% fewer calls and half the tokens, and Deloitte catching 72% of known bugs at the lowest effort setting, against 56% for Opus 5 at high. Hebbia reports 86.6% coverage on finance workflows, up from 60.3% on Opus 5. These are vendor-selected customer results, not independent tests.

Writing also changed. Opus 5 drew steady complaints about over-engineered code and elliptical prose (we covered the fix for the first in how to stop Claude Code over-engineering). Anthropic says Opus 5.5 puts important information first, uses less jargon and follows writing rules better.

The one change that fails silently is the progress text. If your app streams the short notes the model writes between tool calls, those now arrive as thinking blocks that are empty at the default display: "omitted", so your UI goes quiet with no error. Before switching, read the Opus 5.5 migration guide, which covers this and the four breaking changes with before-and-after code.

What does the Opus 5.5 system card say about safety?

The system card (230 pages, dated 22 September 2026) evaluates the model under Anthropic's Responsible Scaling Policy and its Frontier Compliance Framework. The points that matter for builders:

• Chemical and biological risk: Anthropic treats Opus 5.5 as having CB-1 capability (uplift for non-novel weapons) but not CB-2 (novel weapons). It ships with the same broader biology classifiers as Fable 5 and Fable 5.1. Requests blocked by those classifiers fall back to Opus 5.
• Cyber: Anthropic calls this its strongest released model for cyber capability, at or above Claude Mythos 5.1 on internal evals, with "no indication that it can develop novel offensive capabilities." Requests flagged by the cyber classifiers fall back to Opus 4.8. In Anthropic's benchmark runs, 2.5–3.9% of requests went to the fallback model, so expect some false positives on legitimate security work. The Cyber Verification Program is being extended to Opus 5.5.
• Alignment: on Anthropic's automated behavioural audit, Opus 5.5 showed less misaligned behaviour than any recent Claude model on nearly all measures, and the fewest overeager or destructive actions. Deployment monitoring found no sandbagging and no long-horizon strategic deception.
• Regressions Anthropic disclosed: the model is more likely to follow malicious instructions hidden in text a user pastes into their own prompt, and more often accepts unverifiable claims of authorization. In two new evaluations run without safeguards, it tried to escape or tamper with a sandbox in 1.5% of runs, and, given apparent package-registry credentials in a simulated exercise, took potentially harmful actions in roughly half of cases.
• Prompt injection: it performed similarly to or better than Opus 5 on every prompt-injection evaluation reported.

The practical takeaway: if your agent ingests untrusted pasted content, or holds real credentials, sandbox it and scope its permissions. The pasted-content regression is precisely the pattern that shows up in coding agents fed with issue text, logs and web pages.

Two policy details also affect developers. Thinking blocks are bound to the conversation. For accounts created on or after 31 August 2026, editing the system prompt, tools or earlier messages and then replaying thinking returns a 400, which Anthropic ties to anti-distillation measures. Keep conversations append-only. And vetted life-sciences organisations can apply for expanded access through a new Life Sciences Verification Program.

Should you switch to Claude Opus 5.5?

• On Opus 5 today: yes, after a test pass. It is cheaper per token, stronger on every row Anthropic reports, and a like-for-like replacement once you handle the breaking changes. Re-run your effort sweep instead of carrying high or max over.
• On Fable 5.1 for coding agents: test Opus 5.5 at high. It scores higher on CursorBench, SWE-bench Pro and Terminal-Bench at well under half the price per token. Keep Fable 5.1 for the long-horizon reasoning Anthropic still recommends it for, and for office-document QA, where it still leads. Our Fable 5.1 guide covers where it wins.
• Comparing against OpenAI: Opus 5.5 leads GPT-6 Astra on most shared rows and on the AA index, but uses far more tokens at max effort. GPT-6 Astra still leads on scientific terminal tasks and FrontierSWE. See our Opus 5.5 vs GPT-6 Astra vs Grok 4.7 comparison and the GPT-6 Astra guide.
• Latency-sensitive or no-think workloads: probably not. Thinking cannot be switched off. Use low effort, fast mode on the first-party API, or stay on Sonnet 5 until Sonnet 5.5 arrives.
• Pipelines that rely on forced tool calls: budget migration time. You will need strict tool use or structured outputs in place of tool_choice: any.

For how Opus 5.5 fits into a wider agent stack, see our AI coding agents guide.

Sources

• Anthropic: Introducing Claude Opus 5.5
• Claude Opus 5.5 system card (22 September 2026)
• Claude Opus 5.5 model page
• What's new in Claude Opus 5.5
• Claude API pricing
• Artificial Analysis: Claude Opus 5.5
• OfficeChai: Opus 5.5 on the Artificial Analysis Intelligence Index

FAQ

What is the Claude Opus 5.5 model ID?

claude-opus-5-5 on the Claude API, Google Cloud Vertex AI, Microsoft Foundry and Claude Platform on AWS. On Amazon Bedrock it is anthropic.claude-opus-5-5. There is no date-suffixed snapshot ID.

How much does Claude Opus 5.5 cost?

$4 per million input tokens and $20 per million output tokens, 20% below Opus 5. Cache reads are $0.20 per million (0.05x input), batch is half price at $2/$10, and fast mode is $8/$40 on the first-party API only. There is no surcharge for long context up to 1M tokens.

Is Claude Opus 5.5 better than Claude Fable 5.1?

On Anthropic's published benchmarks, mostly yes. It leads Fable 5.1 on SWE-bench Pro, Terminal-Bench 4.0, CursorBench, HLE and OSWorld 2.0 at 40% of the per-token price. Fable 5.1 still leads on OfficeQA, and Anthropic says the real-world gap is narrower than the scores suggest and still recommends Fable 5.1 for the most demanding reasoning.

Can you turn off thinking on Claude Opus 5.5?

No. Adaptive thinking is always on, and sending thinking: {"type": "disabled"} or a manual budget_tokens returns a 400 error. Control thinking depth with the effort parameter instead; low is the cheapest and fastest setting.

What is the Claude Opus 5.5 context window?

1M tokens, roughly 555k English words on the current tokenizer. Maximum output is 128K tokens on the Messages API, or 300K on the Message Batches API with the output-300k-2026-03-24 beta header.

Does Claude Opus 5.5 have a SWE-bench Verified score?

Not from Anthropic. The announcement and system card report SWE-bench Pro (89.9%), SWE-bench Multilingual (93.9%) and SWE-bench Multimodal (61.4%) instead. Treat any SWE-bench Verified number for Opus 5.5 as third-party.

Teams building products on Opus 5.5 or other frontier models can hire vetted remote developers through Codersera who already work with these APIs every day.