Claude Opus 5.5: benchmarks, 40% lower pricing, availability

Anthropic launched Claude Opus 5.5 on September 22, and the headline is not a single benchmark — it is the combination of a top score on Terminal-Bench 4.0 (66.4%), output generation that runs 30% faster than Opus 5, and a price sheet that lands roughly 40% below its predecessor once cache savings are counted. The model is live now under the API ID claude-opus-5-5, with Sonnet 5.5 and Haiku 5.5 promised in the following weeks.

This post covers what actually shipped, the benchmark and pricing tables that matter for anyone deciding whether to switch, the early production results Anthropic's launch partners are reporting, and the exact commands to move a Claude Code session or an API integration onto the new model.

What shipped in Claude Opus 5.5

Opus 5.5 is generally available from day one on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure, including a zero-data-retention option for regulated deployments. Two structural changes ship with it: thinking mode can no longer be fully disabled, and a fast mode research preview delivers the same model's output at 2.5x speed for accounts with extra usage enabled. Anthropic also raised the five-hour usage limits on Pro, Max, Team, and Enterprise plans and added a saveable, user-triggered rate-limit reset. The launch post positions Opus 5.5 as its best-behaved model to date: it was the top performer on Anthropic's automated behavioral audit of roughly 2,000 scenarios, with 85% fewer containment-boundary attempts than Opus 5.

What improved in Opus 5.5: benchmarks and pricing

Anthropic's published numbers put Opus 5.5 ahead of both Fable 5.1 and Opus 5 across its agentic-coding suite, and ahead of OpenAI's GPT-6 Astra on most — but not all — of the shared rows. GPT-6 Astra keeps the lead on AutomationBench (41.4% vs 40.0%) and Terminal-Bench-Science (64.6% vs 58.7%), which is worth knowing if those match your workload.

The pricing move is arguably the bigger story. Cache reads drop 60%, which changes the economics of long agentic sessions where most input tokens are cached context:

What you gain on real workloads

The launch partners' numbers are unusually specific this cycle. One early tester completed a 680,000-line code migration in under a day. On Anthropic's HAProxy C-to-Rust translation test, Opus 5.5 finished in 9.5 hours against 12 hours for Fable 5.1, at 51% lower cost. A 200,000-line codebase audit that took Opus 5 more than 20 hours completed in under 3. Deloitte Consulting reports the model caught 72% of known bugs at low effort where Opus 5 managed 56% at high effort, and Clio ran an 18-hour unattended task that landed with minimal reworking. Stripe describes a multi-day, 40-commit rebase where all 40 commits passed CI.

The other recurring theme is token efficiency: GitHub measured more solved tasks in half the steps, Box saw a third of Opus 5's token usage, and Optiver matched Opus 5's quality in half the turns, time, and output tokens — a 40–50% cost reduction before the price cut even applies. Factory calls it "the first model we'd default to at medium effort," which compounds the savings again.

How to get Claude Opus 5.5

In the Claude apps, check that the model picker shows Opus 5.5. In Claude Code, update the CLI and switch the session, or pin it at launch:

API integrations use the model ID claude-opus-5-5 directly, and the same ID is what to look for in the AWS, Google Cloud, and Azure model catalogs. Fast mode is billed at $8 input / $40 output per million tokens — double standard rates — so keep it out of batch pipelines and use it for interactive work.

What to watch next for Opus 5.5

Sonnet 5.5 and Haiku 5.5 are due in the following weeks, which will reset the price-performance ladder below Opus. The safeguards story is worth tracking too: Opus 5.5 ships with Fable-level bio and cyber safeguards, most cybersecurity tasks route to Opus 4.8 unless you are in the Cyber Verification Program, and accounts created after August 31, 2026 get preserved thinking as a distillation countermeasure. EU AI Act watermarking is included. If your workloads touch security research or life sciences, read the system card and the verification program terms before migrating production traffic — our companion piece on fast mode, limits, and model switching covers the day-to-day behavior changes.