Claude Opus 5.5: stronger, cheaper, and the default in Claude Code
Read in : Français

Contents
Claude Opus 5.5, launched by Anthropic on September 22, 2026, is the first model in the Claude 5.5 family, and it became Claude Code's default Opus model the same day. Anthropic says it performs at the level of Claude Fable 5.1 on most tasks and costs 40% less to run than Opus 5. For you, the news is mostly practical: a stronger, cheaper model in your agent, with an effort setting worth checking.
What did Anthropic announce on September 22, 2026?
Anthropic released Claude Opus 5.5 on the Claude API, on AWS, Google Cloud and Microsoft Azure, and in the Pro, Max, Team and Enterprise plans. The key line in the announcement: it "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5."
Anthropic also raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans, and said Claude Sonnet 5.5 and Claude Haiku 5.5 will follow "in the coming weeks."
Pricing
| Item | Claude Opus 5.5 | Opus 5 for reference |
|---|---|---|
| Input (per million tokens) | $4 | $5 |
| Output (per million tokens) | $20 | $25 |
| Cache reads | $0.20 | — |
| Batch processing | $2 / $10 | — |
| Fast mode (up to 2.5x faster) | $8 / $40 | — |
Prices as announced on September 22, 2026 on Anthropic's page and in the Claude Platform docs. The 20% cut applies to per-token prices; the 40% saving Anthropic quotes applies to typical workloads at default settings.
Is it actually better at coding?
On Anthropic's published numbers, yes, and clearly so on terminal work. These are vendor results; GPT-6 Astra scores are as reported by OpenAI.
| Benchmark (Anthropic figures) | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% |
| FrontierCode v1.1 (Main) | 54.4% | 50.3% | 48.0% | 53.3% |
Anthropic notes that its Terminal-Bench results were run at "xhigh" effort, and itself warns that benchmark margins are "a less reliable guide than they used to be." On CursorBench, the announcement text gives 52.5% at default effort, against 41.7% for GPT-5.6 Sol.
What changes in Claude Code?
Claude Code 2.1.280, released September 22, 2026, adds Opus 5.5 and makes it the default Opus model. The official changelog lists a 1 million token context, $4 / $20 per million tokens and $0.20 cache reads. If you use the opus alias, you move to Opus 5.5 after updating, with nothing else to do.
Two changes deserve your attention, as reported by Classmethod from the Claude Code docs:
- The default model on Pro and Team Standard plans moves from Sonnet to Opus. If you would rather stay on Sonnet (to stretch your allowance, for instance), you can set
"model": "sonnet"in~/.claude/settings.json. - Effort is now set per model. A global
effortLevelyou set earlier no longer applies to Opus 5.5: the model starts at its default, "medium," until you pick a level.
What changes if your app calls the API
If your app calls Claude directly, the docs list four breaking changes compared with Opus 5. The two most likely to bite:
- Thinking can no longer be turned off. A request with
thinking: {"type": "disabled"}returns a 400 error. Remove that parameter and set an effort level instead. - Forced tool use is no longer supported.
tool_choiceset toanyortoolreturns a 400 error; switch toautowith strict tool use.
Here is a prompt to give your agent so the migration does not break anything:
My app calls the Claude API with the claude-opus-5 model. I want to move to claude-opus-5-5.
On a new branch:
1) list every API call and its thinking, tool_choice and effort parameters;
2) remove thinking "disabled" or "enabled" with a budget, and set an explicit effort;
3) replace tool_choice "any" or "tool" with "auto";
4) run the tests and give me the list of changed files.
Do not deploy anything.
Before you merge, click through your app's main flow by hand. Our guide to testing AI-generated code when you are not a developer shows you how.
What about safety safeguards?
Anthropic says Opus 5.5 is the first Opus model to launch with safeguards in the same class as Fable 5.1, covering biology and cybersecurity. Most cybersecurity tasks are rerouted to Opus 4.8, and organizations that need more can go through the Cyber Verification Program. For vibe coders, the likely effect is a refusal or an answer from another model on a highly offensive request; a security review of your own code remains a legitimate ask.
What to do now
- Update Claude Code and check with
/modelthat you are on Opus 5.5. - **Pick your effort with
/effort**: "medium" for small fixes, higher for a refactor or a multi-file feature. - Watch your allowance if you are on Pro: defaulting to Opus instead of Sonnet uses more, even with raised limits.
- Update your context file (
CLAUDE.md): it is what carries your instructions from one model to the next. See our guide to CLAUDE.md, AGENTS.md and rules files. - Recalculate your API budget if your app uses Opus: at $4 / $20, that line drops by at least 20%. The full breakdown is in what a vibe-coded app really costs.
To get started with the tool, read our Claude Code beginner's guide. To see how it compares with Cursor and Codex, check the coding agents comparison, and to pick a model for your project, our guide to the best AI models for coding.
Frequently asked questions
Frequently asked questions
How much does Claude Opus 5.5 cost?
Anthropic lists $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5 ($5 and $25). Cache reads cost $0.20 per million, batch processing is half price ($2 and $10), and a fast mode up to 2.5x quicker is billed at $8 and $40. The model is also available in the Pro, Max, Team and Enterprise plans.
Is Claude Opus 5.5 the default model in Claude Code?
Yes for the Opus alias: Claude Code 2.1.280, released September 22, 2026, makes Opus 5.5 the default Opus model, with a 1 million token context. According to Classmethod, citing the docs, Pro and Team Standard plans also switch their default from Sonnet to Opus. You can switch back to Sonnet in your settings.
Why does Opus 5.5 seem less thorough than Opus 5 on some tasks?
Most likely because of effort. On the API, a request without an effort setting ran at "high" on Opus 5 and runs at "medium" on Opus 5.5. In Claude Code, an old global effort setting no longer applies to this model. Pick an explicit level with /effort before judging quality.
Do I need to change my app's code to move to Opus 5.5?
Often, yes. Anthropic's docs list breaking changes: thinking can no longer be disabled, forced tool use returns an error, and some older computer use tool versions are no longer accepted. Change the model ID on a test branch, fix those parameters, and test before you deploy.
When are Claude Sonnet 5.5 and Haiku 5.5 coming out?
As of September 28, 2026, Anthropic had only said they would follow "in the coming weeks," with many of the same performance, efficiency and safety improvements. No specific date was given. In the meantime, if your allowance runs out too fast, you can set Sonnet back as Claude Code's default model.
Sources
- Introducing Claude Opus 5.5 — Anthropic
- What's new in Claude Opus 5.5 — Claude Platform documentation
- Claude Code CHANGELOG (version 2.1.280) — Anthropic, GitHub
- Claude Code v2.1.280 Major Updates — Classmethod DevelopersIO
- Claude Opus 5.5: Anthropic cuts price 20%, tests safety — BetaNews
- Anthropic launches Opus 5.5, its first model since CEO Amodei called for AI slowdown — Yahoo Finance


