New · Released September 22, 2026
Claude Opus 5.5: Specs, Pricing & Fast Mode
Anthropic's new flagship for long-running agentic work launched on September 22, 2026 at $4/$20 per MTok — 20% below Opus 5 — with a 1M-token context window by default, 128K max output and always-on adaptive thinking. Every figure below is read from Anthropic's pricing page via our model registry.
The short answer
Claude Opus 5.5 (claude-opus-5-5) is Anthropic's newest flagship, launched September 22, 2026. It costs $4 per million input tokens and $20 per million output tokens — 20% below Opus 5 on both — with a 1M-token context window by default, 128K max output tokens and always-on adaptive thinking: the thinking blocks can't be disabled, and depth is controlled with the effort parameter instead. Fast mode (research preview) bills 2x at $8/$40 on the first-party Claude API only. Anthropic positions it as performing "at the level of Claude Fable 5.1 on most work" while costing "40% less to run than Opus 5" — the sticker rates are 20% lower; the rest of that claim rests on adaptive thinking spending fewer tokens, which only your own workload can confirm.
Claude Opus 5.5 specs and API pricing
All rates are USD per million tokens, taken from our model registry entry for claude-opus-5-5 — never hard-coded into this page.
| Specification | Claude Opus 5.5 |
|---|---|
| Released | 2026-09-22 |
| API model ID | claude-opus-5-5 |
| Input / MTok | $4.00 |
| Output / MTok | $20.00 |
| Cache read (hit) / MTok | $0.20 |
| Cache write, 5-minute TTL / MTok | $5.00 |
| Cache write, 1-hour TTL / MTok | $8.00 |
| Batch API (in / out) | $2.00 / $10.00 |
| Context window | 1M tokens by default |
| Max output tokens | 128K |
| Thinking | Always-on adaptive thinking — cannot be disabled |
| Thinking depth control | Controlled with the effort parameter |
| Available on | Claude API, Bedrock, AWS, Google Cloud, Foundry |
| Data retention | Not stated on the pricing page — not confirmed |
Figures verified against Anthropic's pricing page on 2026-09-25; the shared Anthropic baseline in our registry was last checked on 2026-09-25.
Fast mode: $8 in / $40 out, exactly 2x
Fast mode entered research preview for Opus 5.5 on launch day. The premium is a clean 2x on the model's own base rate — derived here from the registry, matching Anthropic's published $8/$40.
| Rate | Standard | Fast mode (2x, derived) |
|---|---|---|
| Input / MTok | $4.00 | $8.00 |
| Output / MTok | $20.00 | $40.00 |
| Cache read / MTok | $0.20 | $0.40 |
| Batch API (in / out) | $2.00 / $10.00 | cell_no_batch |
The Fast column is our arithmetic: Anthropic's pricing page publishes $8 in / $40 out for Opus 5.5 Fast mode, exactly 2x the $4/$20 base row — the same 2x rule that covers Opus 5 and Opus 4.8 at $10/$50. Fast mode is available on the first-party Claude API only (not Bedrock, AWS, Google Cloud or Foundry), does not combine with the Batch API, and prompt-caching and data-residency multipliers apply on top of it. Full rules on our Claude Fast mode pricing page.
Opus 5.5 vs Claude Opus 5
The direct predecessor, still on the pricing page under Additional models. Rates from the registry.
| Rate | Opus 5.5 | Opus 5 |
|---|---|---|
| Input / MTok | $4.00 | $5.00 |
| Output / MTok | $20.00 | $25.00 |
| Cache read / MTok | $0.20 | $0.50 |
| Cache write, 5-minute / MTok | $5.00 | $6.25 |
| Cache write, 1-hour / MTok | $8.00 | $10.00 |
| Batch API (in / out) | $2.00 / $10.00 | $2.50 / $12.50 |
Per token, Opus 5.5 is 20% cheaper on input and 20% cheaper on output than Opus 5 — both rows sit in our registry, so the comparison is arithmetic, not marketing. Anthropic's own launch wording goes further, claiming Opus 5.5 "costs 40% less to run than Opus 5"; that larger figure is about real workloads — always-on adaptive thinking spending fewer output tokens — and is a vendor claim we can't re-derive from published rates. The behavioural change is just as significant: on Opus 5.5 you cannot disable thinking at all, so migrate only after you have measured how the effort parameter changes your token spend.
Opus 5.5 vs Claude Fable 5.1
Anthropic's other current flagship. The comparison is the launch headline — here are the numbers behind it.
| Rate | Opus 5.5 | Fable 5.1 |
|---|---|---|
| Input / MTok | $4.00 | $10.00 |
| Output / MTok | $20.00 | $50.00 |
| Cache read / MTok | $0.20 | $0.25 |
| Cache write, 5-minute / MTok | $5.00 | $12.50 |
| Cache write, 1-hour / MTok | $8.00 | $20.00 |
| Batch API (in / out) | $2.00 / $10.00 | $5.00 / $25.00 |
Anthropic's announcement states Opus 5.5 "performs at the level of Claude Fable 5.1 on most work" — vendor wording, not our benchmark. On rates the gap is wide: 60% less on input and 60% less on output per million tokens, both rows verified in the registry. One wrinkle favours Fable 5.1 on cache-heavy workloads: its cache reads bill at 0.025x base input versus 0.05x on Opus 5.5 (and the standard 0.1x on Opus 5), so a heavily cached Fable 5.1 deployment narrows the gap faster than the sticker rates suggest.
Where Opus 5.5 runs, and what changes
Platform availability is broad — broader than Fast mode — but the thinking and tool rules differ from every earlier Opus.
Every major surface at launch
Anthropic's release note lists the Claude API, Claude in Amazon Bedrock, Claude Platform on AWS, Claude on Google Cloud and Claude in Microsoft Foundry — day-one availability everywhere, unlike Fast mode which stays first-party only.
Partner clouds included
Unlike Fast mode, there is no first-party restriction: the pricing page's main model table carries Opus 5.5 as a standard row, and the release note names Bedrock, AWS, Google Cloud and Foundry explicitly. Computer use differs by surface — on the Claude API and Google Cloud it requires the computer_toolset_20260801 toolset and the earlier computer_20251124 tool returns a 400 error, while on Amazon Bedrock computer_20251124 keeps working.
Thinking cannot be switched off
On Opus 5.5, a request that passes a thinking field — whether type is disabled or enabled with a budget — returns a 400 error. Adaptive thinking is always on; you control depth with the effort parameter instead. The same release note applies the Fable 5.1 rule to tools: tool_choice types any and tool return 400 — use auto with strict tool use.
Mid-conversation tools, in beta
Tools can now be defined inside a mid-conversation system message, in beta on the Claude API with the inline-tools-2026-09-15 beta header — a change that shipped alongside Opus 5.5 and applies across the API generally.
Opus 5.5 launched days before this page was last verified on 2026-09-25, and Anthropic moved Opus 5 and Opus 4.8 to the "Additional models" section of its pricing table the same day. Availability, retention terms and tool behaviour for this model may still shift — treat this page as current, not contractual.
Price Opus 5.5 on your own workload
Put your token counts next to the rates above with our cost calculator, or compare every model we track side by side.