Anthropic ships Opus 5.5 on 22 September

Anthropic introduced Claude Opus 5.5 on 22 September 2026. The post calls it the first model in a Claude 5.5 family and Anthropic's first release since the lab asked peers to pace the frontier.

Anthropic says Opus 5.5 performs at Claude Fable 5.1 level on most work and costs 40% less to run than Opus 5 on typical workloads at default settings. API list prices are $4 input and $20 output per million tokens, with cache reads at $0.20 per million. The model id on the Claude Platform is claude-opus-5-5. The company also published a system card dated 22 September 2026.

Anthropic Claude Opus 5.5 announcement social card Caption: Anthropic social card for the Claude Opus 5.5 launch · Source: Anthropic, Introducing Claude Opus 5.5, 22 September 2026 · link

Claude Platform, three clouds, gated bio and cyber

Developers on the Claude Platform can call claude-opus-5-5 now. Anthropic says the model is available on Amazon Web Services, Google Cloud, and Microsoft Azure as well.

Fast mode is named for Claude Code and the Claude Platform. Pro, Max, Team, and seat-based Enterprise plans get higher five-hour usage limits, plus a banked rate-limit reset users can spend later.

Biology and cyber access stay gated. Vetted groups can apply to the Life Sciences Verification Program for biology research use. Anthropic says Cyber Verification Program access for Opus 5.5 expands in the coming weeks for verified cybersecurity practitioners.

External pre-release testers named in the post include Frontier Design and METR.

Lower list prices and three operator-facing changes

Relative to Opus 5, Anthropic's post stresses three operator-facing changes:

  1. Price and speed. List prices cut 20% on input and output versus Opus 5. Cache reads drop 60% to $0.20 / MTok. Anthropic reports more than 30% faster output generation and 40% lower cost on typical workloads at default settings.
  2. Capability claims on agentic coding and knowledge work. The post publishes lab tables (Terminal-Bench 4.0, FrontierCode v1.1 Main, CursorBench 4.0, GDPval-AA v2.1, and others) that put Opus 5.5 ahead of Opus 5 and competitive with Claude Fable 5.1 and GPT-6 Astra on several rows. Anthropic also says real-world gaps versus Fable 5.1 look narrower than those scores suggest.
  3. Safeguards and external review. The system card covers RSP chemical/biological, cyber, agentic safety, and alignment audits. Anthropic says Opus 5.5 is the strongest model it has tested on its automated behavioral audit. The post compares biology and cyber capability to Claude Mythos 5.1 and says safeguards match the Fable 5.1 class, with verification programs for vetted use.

Sonnet 5.5 and Haiku 5.5 are promised in coming weeks. This announcement does not ship those model ids.

No third-party board and no cloud SLOs

The post does not publish a single independent third-party leaderboard that replaces Anthropic's own numbers. External evaluators (Frontier Design, METR) are named; their full public write-ups are not attached to the launch page.

Cloud and regional price sheets, latency SLOs, and exact rate limits on AWS, GCP, and Azure are not spelled out in the Anthropic post. Operators still need vendor console pricing for those paths.

Biology and cyber use remain behind verification programs. The system card's CB and cyber conclusions are Anthropic's risk determinations without a regulator sign-off.

Anecdotes in the post (for example a 680,000-line migration completed in less than a day) are early-tester stories without a controlled baseline you can bill against.

Canary Opus 5.5 before a fleet cutover

  • API and Claude Code teams: Point a canary at claude-opus-5-5. Compare token spend and cache-hit rate against Opus 5 on the same agent traces before a fleet cutover.
  • Buyers on cloud marketplaces: Confirm the model id and list price in your AWS / GCP / Azure console; do not assume Anthropic Platform pricing maps 1:1.
  • Life-sciences and cyber shops: If you need those workloads, start the verification applications Anthropic names. Do not treat general API access as clearance for those domains.
  • Ignore for now: Sonnet 5.5 / Haiku 5.5 marketing until those model ids ship with their own cards.