Anthropic released Claude Opus 5.5 on September 22, 2026, with model ID claude-opus-5-5, a 1 million-token context window, and a 128,000-token maximum output, according to Claude Opus 5.5's own overview page, which lists the model's status as "Latest." That status is the detail a team already paying for Opus 5 or Fable 5.1 needs to act on this week: it puts Opus 5.5 ahead of both in Anthropic's own lineup, with a price cut and four breaking changes attached to the switch.
What Claude Opus 5.5 costs vs. Opus 5 and Fable 5.1
Opus 5.5 runs $4 per million input tokens and $20 per million output tokens, a straight 20% cut on both sides of Opus 5's $5/$25 rate, per the same overview page. Caching got cheaper too: a 5-minute cache write drops from $6.25 to $5 per MTok, and cache reads fall from $0.50 to $0.20 per MTok.
| Model | Input | Output | Cache write (5m) | Cache read | Fast mode |
|---|---|---|---|---|---|
| Opus 5.5 | $4/MTok | $20/MTok | $5/MTok | $0.20/MTok | $8/$40 per MTok |
| Opus 5 | $5/MTok | $25/MTok | $6.25/MTok | $0.50/MTok | $10/$50 per MTok |
| Fable 5.1 | $10/MTok | $50/MTok | not published | $0.25 (2.5% of input, per the overview footnote) | not supported |
Fable 5.1 sits well above both, at 2.5 times Opus 5.5's input price, and it comes with a "Slower" latency rating and a default effort of high, against Opus 5.5's medium, per the same page's model comparison table. That's the opposite direction from a scheduled price change to budget around: Gemini 3.8 Flash's launch price holds through the end of 2026 before doubling on January 1, 2027, while Opus 5.5's cut is already live.
Fast mode also got cheaper, if a team is already paying the fast-mode premium on Opus 5. Opus 5.5's fast mode costs $8 per million input tokens and $40 per million output tokens, down from $10/$50 on Opus 5 and Opus 4.8, per Anthropic's fast-mode pricing page. It's still a research preview available on the Claude API only, including Claude Managed Agents, and it doesn't run on Amazon Bedrock, Claude Platform on AWS, Google Cloud, or Microsoft Foundry. Before booking the savings, it's worth doing the same check you'd run to re-check a pinned model name after a vendor swap: claude-opus-5 and claude-opus-5-5 are different literal model strings. Nothing upgrades one to the other automatically.
Is Claude Opus 5 actually being deprecated?
No. Anthropic's status label for Opus 5 is "Legacy," a lifecycle stage that means no retirement date has been announced, only a floor of "not sooner than July 24, 2027," per Opus 5's own model overview. The page nudges migration without setting a deadline: "Although Claude Opus 5 is still available, you should consider migrating to Claude Opus 5.5 for improved performance." A team can keep production traffic on Opus 5 today without anything shutting off.
That floor comes from a standing Anthropic commitment to keep a model live for at least a year past release, and Opus 5.5 carries the same clause, "not sooner than September 22, 2027." Neither date is a shutoff, both are the earliest one could happen, and Anthropic hasn't announced anything closer for either model. Budget and eval work should assume the newer model regardless, since that's the direction Anthropic is steering usage, but there's no clock forcing the move.
How to migrate a Claude Opus 5 integration to Opus 5.5 without it breaking
Four independent changes hit any code still targeting claude-opus-5 the moment it switches to claude-opus-5-5, per the what's-new page for Opus 5.5: thinking can't be disabled, forced tool use returns an error, thinking blocks are tied to the model and the conversation, and, on the Claude API and Google Cloud, the earlier computer_20251124 tool isn't accepted anymore. None of the four are optional cleanup, and each one 400s on the old calling pattern rather than failing quietly. It's the same pattern GPT-6 Astra's own breaking changes forced on tool-calling code the same month, on a different vendor's stack.
1. Stop disabling or manually budgeting thinking
Sending thinking: {"type": "disabled"} now returns: "thinking.type.disabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior. Sending {"type": "enabled", "budget_tokens": ...} to set a manual thinking budget returns a matching error for that string instead: "thinking.type.enabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior. Omit the thinking parameter, or send {"type": "adaptive"} and control depth through effort instead.
2. Replace forced tool_choice with auto plus strict tool use
Any call setting tool_choice: {"type": "any"} or {"type": "tool", "name": ...} now returns: tool_choice: type "tool" and "any" are not supported for this model. Move that logic to tool_choice: {"type": "auto"} with strict: true, or switch to structured outputs.
3. Check whether your thinking blocks survive the model switch
Opus 5.5 can read thinking blocks generated by Opus 5, which keeps that upgrade path safe, but it can't read Fable or Mythos thinking blocks. For accounts created on or after August 31, 2026, 00:00 UTC, editing a prompt, tool, or message ahead of a preserved thinking block on the Claude API or a cloud platform now returns a 400 instead of silently working. Send the thinking-binding-controls-2026-08-01 beta header if you need to opt into dropping the block rather than erroring.
4. Move off computer_20251124 if you're on the Claude API or Google Cloud
The old computer_20251124 tool type is rejected outright, with an error naming it directly: 'claude-opus-5-5' does not support tool types: computer_20251124. Replace it with the computer_toolset_20260801 toolset. Amazon Bedrock is the exception; the old tool still works there, so this step doesn't apply if that's where the traffic runs.
Does Opus 5.5 really perform at Fable 5.1's level?
Anthropic's own framing is direct: the company says Opus 5.5 performs at the level of Claude Fable 5.1 on most work, while costing roughly 40% less to run than Opus 5. Set against Fable 5.1's price and its high default effort, that's a real reason to test Opus 5.5 before assuming a Fable 5.1 subscription is worth the premium.
What that sentence doesn't give you is a clean benchmark ranking. Anthropic's release lists Opus 5.5, Fable 5.1, and Opus 5 side by side across several benchmarks, including Terminal-Bench 4.0, FrontierCode v1.1, CursorBench 4.0, GDPval-AA v2.1, and OSWorld 2.0. That parity claim is Anthropic's own characterization, and the table's Fable 5.1 column lists GDPval-AA at v2.1, while Fable 5.1's own launch post reported that benchmark at v2, so the two aren't confirmed to be the same test. Before downgrading from Fable 5.1 to Opus 5.5, or skipping Opus 5 entirely to go straight to Fable 5.1, pull the comparison table directly from Anthropic's own announcement page and check the benchmark versions, not just the headline sentence.
Anthropic also cites a specific case: one tester reportedly completed a 680,000-line code migration in under a day running the new model. That's an anecdote rather than a benchmark. It lines up with Anthropic's own account of Claude writing most of its code, where the company says its engineers now ship 8x as much code per quarter with Claude authoring 80% of it.
FAQ
Is Claude Opus 5.5 available on Bedrock, Google Cloud, and Microsoft Foundry?
The model itself is available broadly. Fast mode has a narrower scope: a research preview on the Claude API only, including Claude Managed Agents, and it doesn't run on Bedrock, Claude Platform on AWS, Google Cloud, or Microsoft Foundry.
What breaks if I just swap the model ID from claude-opus-5 to claude-opus-5-5?
Four things. Manually disabled or budgeted thinking, forced tool_choice values, thinking blocks tied to an account created on or after August 31, 2026, and the computer_20251124 tool type outside Bedrock. Each returns a 400 rather than failing quietly, so a straight model-ID swap surfaces the problems fast in testing, provided you test before rolling out.
Does fast mode work on Opus 5.5?
Yes, at $8 per million input tokens and $40 per million output tokens, down from $10/$50 on Opus 5. It's still a Claude API research preview, so it isn't an option if traffic runs through Bedrock, Google Cloud, Microsoft Foundry, or Claude Platform on AWS.
What's the default effort level, and does it change my costs?
Opus 5.5's default effort is medium. A request that omits the effort parameter now runs at medium instead of the high default Opus 5 used. Lower default effort generally burns fewer output tokens per request at the same settings, which compounds with the 20% price cut on a bill that isn't explicitly tuning effort per call.