Claude Opus 5.5: What Changed, What It Costs, and What Breaks
Published 23 September 2026 by Jake Hissitt, Founder of Stob.AI.
Anthropic released Claude Opus 5.5 on 22 September 2026 as model ID claude-opus-5-5. It is cheaper than the model it replaces, keeps the million-token context window, and changes enough in its default behaviour that existing integrations need checking before you switch.
The specifications
| Attribute | Claude Opus 5.5 |
|---|---|
| Model ID | claude-opus-5-5 |
| Input price | $4.00 per 1M tokens |
| Output price | $20.00 per 1M tokens |
| Cache reads | $0.20 per 1M tokens |
| Cache writes (5 minute) | $5.00 per 1M tokens |
| Batch discount | 50% |
| Context window | 1,000,000 tokens |
| Max synchronous output | 128,000 tokens |
| Knowledge cutoff | June 2026 |
| Availability | Generally available, 22 September 2026 |
Against Claude Opus 5 at $5 / $25, that is a 20% price cut at the same context size. Against Claude Fable 5.1 at $10 / $50, it is 60% cheaper and, for most production work, close enough on quality that Fable becomes the escalation model rather than the default.
Adaptive thinking is always on
The biggest behavioural change is that extended thinking is no longer something you switch on. Adaptive thinking runs by default at medium effort, and the model decides how much deliberation each step deserves.
That is good for quality and bad for anyone who budgeted output tokens assuming no thinking. Thinking tokens bill as output. Re-measure your per-request cost before assuming the headline price cut flows straight to your invoice.
Migration checklist
- Disabled thinking. Integrations that explicitly disabled extended thinking need reviewing against the new default behaviour.
- Forced tool use. Patterns that force a specific tool call behave differently under adaptive thinking; test each one.
- Thinking blocks. Code that parses or passes back thinking blocks from earlier Claude versions needs checking for compatibility.
- Older tool versions. The model rejects the older
computer_20251124tool version. Update the tool definition before switching. - Output budgets. Raise your max-tokens ceiling if it was tuned tightly around a non-thinking model.
- Prompt caching. Cache writes are $5 per million against $0.20 reads, so caching only pays when a prefix is genuinely reused. Check your hit rate.
Run the switch on a copy of your evaluation set before it touches production. The cheapest migration bug is the one you find on twenty test cases.
Where it sits against everything else
| Model | Input / 1M | Output / 1M | Context |
|---|---|---|---|
| GPT-6 Sol | $2.00 | $10.00 | 1,050,000 |
| Claude Opus 5.5 | $4.00 | $20.00 | 1,000,000 |
| GPT-6 Astra | $10.00 | $50.00 | 1,050,000 |
| Claude Fable 5.1 | $10.00 | $50.00 | 1,000,000 |
| Claude Opus 5 (superseded) | $5.00 | $25.00 | 1,000,000 |
Opus 5.5 is now our default Claude recommendation for agentic coding, research and professional knowledge work. Fable 5.1 stays for the hardest long-horizon reasoning. The full head-to-head against OpenAI's equivalent is in Claude Opus 5.5 vs GPT-6 Sol.
Frequently asked questions
How much does Claude Opus 5.5 cost?
$4 per million input tokens and $20 per million output tokens, with cache reads at $0.20, five-minute cache writes at $5, and a 50% Batch API discount.
Is Claude Opus 5.5 better than Opus 5?
Yes, and it is 20% cheaper. Opus 5 is superseded; there is no reason to start a new build on it.
Should I move from Claude Fable 5.1 to Opus 5.5?
For most production work, yes — it is 60% cheaper at comparable quality. Keep Fable 5.1 for the hardest reasoning tasks where you can show it wins on your own evaluations.
What breaks when I upgrade?
Integrations that disable thinking, force tool use, replay old thinking blocks, or use the older computer-use tool version. The model rejects computer_20251124 outright.
What is the context window?
One million input tokens with up to 128,000 tokens of synchronous output.
What to do next
Update your model ID on a branch, run your evaluation set, and compare cost per accepted outcome rather than cost per token. Then check the full field in our AI model comparison tool and the provider guide, Best Claude Model 2026.