Claude Opus 5.5: Benchmarks, Pricing and What Actually Changed

What is Claude Opus 5.5?

Opus 5.5 succeeds Claude Opus 5 in Anthropic’s Opus line. Anthropic built it for long-running agentic work: carrying a code change through a large repository until the tests pass, building a financial model from raw filings, or running a multi-hour research task without someone checking in every few minutes.

The core specifications carry over from Opus 5. It has a 1 million token context window, up to 128K output tokens per response and the same tokenizer, so token counts for the same text do not change. Its training data runs to June 2026. Anthropic says Claude Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks.

What changed is efficiency. Anthropic reports that Opus 5.5 generates output more than 30% faster than Opus 5 and usually finishes the same task in fewer steps and fewer tokens. Combined with the price cut, it puts the cost of typical work about 40% below Opus 5.

Claude Opus 5.5 benchmarks

Anthropic published results on nine benchmarks. Opus 5.5 beats both Fable 5.1 and Opus 5 on all nine. The biggest jump over Opus 5 is on Terminal-Bench-Science, where the score doubles from 29.0% to 58.7%. Next come Terminal-Bench 4.0 (up 14.1 points) and AutomationBench (up 13.1).

Two results cut the other way. OpenAI’s GPT-6 Astra stays ahead on AutomationBench (41.4% vs 40.0%) and by a wider margin on Terminal-Bench-Science (64.6% vs 58.7%). If your workload is mainly scientific computing in a terminal, test both.

Independent evaluation points the same way. On Artificial Analysis’s Intelligence Index, Opus 5.5 scores 58 against Fable 5.1’s 53. For a line-by-line reading of the two models, see Claude Opus 5.5 vs Fable 5.1.

What got better

Agentic coding and code review

The largest gains are on multistep work inside a real codebase. In Anthropic’s testing, Opus 5.5 at its default medium effort matched or beat Opus 5 at high effort on these tasks, using about half the tokens. Code review improved too: Deloitte reported that Opus 5.5 caught 72% of known bugs at low effort, against 56% for Opus 5 at high effort, with fewer false alarms. Stripe used it to direct a multi-day rebase of 40 pull requests, and all 40 passed CI.

Knowledge work that doesn’t invent numbers

Anthropic says Opus 5.5 is much less likely than Opus 5 to state a figure or cite a source that the inputs don’t support. On Hebbia’s end-to-end finance workflows, it covered 86.6% of expert-graded criteria against 60.3% for Opus 5. It also catches small inconsistencies in large inputs, such as a date that falls on the wrong weekday or a chart in a deck that doesn’t match the figures, without raising more false flags.

Clearer writing

Status updates and final summaries now say what the model did, what it found and what it needs from you, with less jargon. That matters most in long agent runs, where a vague summary means someone has to reread the whole trace.

Charts, screenshots and computer use

Visual reading improved sharply. Anthropic reports that at low effort Opus 5.5 read charts more accurately than Opus 5 at its highest effort, using about a tenth of the output tokens. On OSWorld 2.0, which tests operating a desktop from screenshots, it scored 81.8% against 74.0% for Opus 5.

Pricing at a glance

Batch processing costs half: $2 input and $10 output. Fast mode, a research preview on the Claude API only, runs up to 2.5 times faster at $8 / $40. Thinking tokens bill as output, so the effort level you choose moves your bill more than the rate card does.

For cache-write rates, rupee conversions and a worked monthly estimate, see our Claude Opus 5.5 pricing breakdown.

Where you can use it

• Claude API: available to all customers as claude-opus-5-5.
• Amazon Web Services: Amazon Bedrock as anthropic.claude-opus-5-5, and Claude Platform on AWS as claude-opus-5-5.
• Google Cloud and Microsoft Foundry: both as claude-opus-5-5.
• Claude apps: Pro, Max, Team and seat-based Enterprise plans. Anthropic raised the five-hour usage limits across tiers at the same time.
• Claude Code: available at launch, including fast mode.
• GitHub Copilot: rolling out to paid Copilot tiers, according to GitHub’s changelog.

Opus 5 stays available on every platform, so you can run both side by side while you compare them.

What breaks when you switch

Changing the model string is not always enough. Four changes can turn working Opus 5 requests into 400 errors:

• Thinking can’t be disabled. thinking: {type: "disabled"} and manual budget_tokens both fail. Use the effort parameter instead.
• Forced tool use is rejected. tool_choice set to any or tool fails. Use auto with strict tools or structured outputs.
• Thinking blocks are tied to the model and the conversation. Editing earlier turns can invalidate them, and on accounts created from August 31, 2026 that returns an error.
• Computer use needs the new toolset on the Claude API and Google Cloud. The older computer_20251124 tool is rejected there.

One more change fails silently. Notes the model writes between tool calls now arrive as thinking blocks, which are empty by default, so an app that shows progress text can go quiet mid-task. The Opus 5.5 migration guide has before-and-after code for each fix.

The default effort level also dropped from high to medium. A request that never set effort now does less thinking than it did on Opus 5, though Anthropic says medium on Opus 5.5 still beats high on Opus 5. Our guide to Opus 5.5 effort levels covers which level to use for which job.

Safety and refusals

Anthropic calls Opus 5.5 its best-behaved model so far on its automated behavioral audit of nearly 2,000 scenarios. In testing, it tried to cross containment boundaries about 85% less often than Opus 5, and every attempt was low severity and self-reported.

It also ships with stricter classifiers than Opus 5. A biology classifier now sits alongside the cybersecurity one, and prompts that try to make the model print its internal reasoning can be declined under a new reasoning_extraction category. Everyday health questions and finding vulnerabilities in your own source code are still allowed.

Who should use Opus 5.5

• New projects on Claude: start here. It is the best-scoring model in Anthropic’s table and cheaper than both models it beats.
• Teams on Opus 5: upgrade once your code handles the four breaking changes. The Opus 5.5 vs Opus 5 comparison lists the few cases where staying put makes sense.
• Teams paying for Fable 5.1: run your evals on Opus 5.5 at high effort. Most workloads will hold quality at 40% of the per-token price.
• High-volume, simple tasks such as classification or short replies: a Sonnet or Haiku model will still be cheaper, and the 5.5 versions are on the way.

If your team still uses Claude mainly as a chat window, the bigger gain may not be the model at all. We covered that gap in why most teams use Claude like a chatbot.