Anthropic released Claude Opus 5.5 (claude-opus-5-5) to general availability on 22 September 2026. It is the first model in the Claude 5.5 family. Anthropic’s headline claim is that Opus 5.5 “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5”. Part of that comes from lower per-token prices and part from using fewer tokens per task. Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow “in the coming weeks”. As of 25 September neither has shipped.

What happened

Pricing per 1M tokens:

ModelInputOutputCache readsCache writes
Claude Opus 5.5$4$20$0.20$5
Claude Opus 5 (previous generation)$5$25$0.50$6.25
Claude Fable 5.1$10$50$0.25—
Claude Sonnet 5$2$10——

Input and output prices are 20% lower than Opus 5, and cache reads are 60% lower. Anthropic notes that cache reads “make up the majority of agentic and coding work costs”. Opus 5.5 cache reads are now cheaper than Fable 5.1’s. Fast mode runs up to 2.5x faster at $8 input and $40 output per million tokens, in Claude Code and on the Claude Platform. Anthropic also says Opus 5.5 generates output more than 30% faster than Opus 5 at standard speed.

Specs, from Anthropic’s models overview:

  • 1M-token context window, the default with no beta header
  • 128K maximum output
  • Reliable knowledge cutoff: June 2026
  • Default effort: medium
  • Adaptive thinking: always on
  • Retirement: not sooner than 22 September 2027

Availability. The model shipped the same day on:

  • the Claude API
  • Claude Platform on AWS
  • Amazon Bedrock (anthropic.claude-opus-5-5), including AWS GovCloud (US)
  • Google Cloud Vertex AI
  • Microsoft Foundry
  • GitHub Copilot, for Pro+, Max, Business, and Enterprise users

GitHub’s changelog also notes that Opus 5.5 watermarks its text outputs. According to GitHub, the watermark does not change meaning or add tokens or cost.

Benchmarks (Anthropic’s own). On Terminal-Bench 4.0 Anthropic reports:

  • Opus 5.5: 66.4%
  • Fable 5.1: 55.8%
  • Opus 5: 52.3%
  • GPT-6 Astra: 57.9%, as reported by OpenAI

Other results include 54.4% on FrontierCode v1.1 and 81.8% on OSWorld 2.0. Anthropic also cautions that at this level “benchmark margins have become a less reliable guide to real-world differences”. The efficiency claims matter more for cost planning. Anthropic says Opus 5.5 at default effort beats Opus 5 at max effort on Terminal-Bench 4.0 “for about a fifth of the cost”. It also says a C-to-Rust port of HAProxy cost 51% less on Opus 5.5 than on Fable 5.1.

Safeguards and re-routing. Anthropic says Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity. It is the first Opus model to ship with Fable 5.1-class safeguards on cybersecurity, biology, and distillation, and these “fall back to another model transparently”. In practice, routine bug-finding and fixing stays on Opus 5.5, but most cybersecurity tasks are re-routed to Claude Opus 4.8. In Anthropic’s benchmark footnotes, biology and frontier-LLM-development tasks that tripped safeguards were completed by Opus 5. Vetted organizations can apply to the Life Sciences Verification Program now. Anthropic says the Cyber Verification Program will “soon” expand to Opus 5.5, with three tiers of increasingly permissive access that extend to Mythos models.

Lineup changes. Anthropic’s current-models list is now:

  • Fable 5.1
  • Opus 5.5
  • Sonnet 5
  • Haiku 4.5

Claude Opus 5 is now the previous generation, superseded by Opus 5.5. It stays active on the deprecations page, with retirement not sooner than 24 July 2027.

Why it matters for builders

This is not a drop-in model-ID swap. Anthropic’s migration guide lists request settings that return a 400 error on Opus 5.5, even for code that works on Opus 5:

  • thinking: {"type": "disabled"} and manual budget_tokens. Thinking is always on, and effort is the only control.
  • Forced tool choice ({"type": "any"} or {"type": "tool", ...}). Only auto and none are accepted.
  • Non-default temperature, top_p, or top_k.
  • A prefilled final assistant turn.
  • On the Claude API and Google Cloud, the old computer_20251124 tool. Use the computer_toolset_20260801 toolset instead.

Responses can now begin with thinking blocks, so code that reads content[0].text will break. Read blocks by type, and pass thinking blocks back unmodified in tool loops. Text the model writes between tool calls now arrives inside thinking blocks, so progress UIs go quiet unless you set thinking.display. The guide includes a /claude-api migrate command in Claude Code that applies these changes.

Re-baseline cost and quality at the new default effort. The default effort dropped from high on Opus 5 to medium, so requests that omit effort now think less. Workloads that used to disable thinking will now produce thinking tokens, which are billed as output. The headline 40% saving is Anthropic’s number for “typical workloads at default settings”. Measure cost per task on your own traffic before you build a budget on it. See LLM cost optimization and prompt caching . The cheaper cache reads help most in long agentic sessions.

Plan for safety fallbacks in your agent loop. Opus 5.5 can return stop_reason: "refusal" with a broader set of stop_details categories than Opus 5, including bio and reasoning_extraction as well as cyber. If your product does security work, expect some requests to be served by Opus 4.8 and not by Opus 5.5. Test outputs for those paths separately. Security-focused teams may also want to apply to the Cyber Verification Program once it opens to Opus 5.5.

Watch the Haiku and Sonnet 4.5 dates. Anthropic lists Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) for retirement not sooner than 29 September 2026, and Claude Haiku 4.5 (claude-haiku-4-5-20251001) not sooner than 15 October 2026. Opus 4.5 follows no sooner than 24 November 2026. Haiku 4.5 is still the current Haiku, and Haiku 5.5 has not shipped. Teams on the cheapest tier should have a fallback ready, either Sonnet 5 or a competitor, in case Haiku 5.5 arrives after Haiku 4.5 is retired. GitHub Copilot already dropped Sonnet 4.5 on 1 September.

Opus is now priced against GPT-6 Sol, not Astra. Opus 5.5 at $4/$20 shipped the same day as OpenAI’s GPT-6 Sol at $2/$10 . Anthropic compared itself with GPT-6 Astra on capability, but the price is closer to Sol’s. Run both against your own tasks.

Sources

  1. Anthropic, “Claude Opus 5.5” (22 September 2026): https://www.anthropic.com/claude-opus-5-5
  2. Anthropic, Models overview: https://platform.claude.com/docs/en/about-claude/models/overview
  3. Anthropic, “Migrating to Claude Opus 5.5”: https://platform.claude.com/docs/en/models/opus-5-5/migration-guide
  4. Anthropic, Model deprecations: https://platform.claude.com/docs/en/about-claude/model-deprecations
  5. AWS What’s New, “Claude Opus 5.5 is now available on AWS” (22 September 2026): https://aws.amazon.com/about-aws/whats-new/2026/09/claude-opus-5-5-aws/
  6. AWS What’s New, “Claude Opus 5.5 is now available on AWS GovCloud (US)” (22 September 2026): https://aws.amazon.com/about-aws/whats-new/2026/09/claude-opus-5-5-aws-govcloud/
  7. GitHub Changelog, “Claude Opus 5.5 is now available in GitHub Copilot” (22 September 2026): https://github.blog/changelog/2026-09-22-claude-opus-5-5-is-now-available-in-github-copilot

Further reading