Anthropic launched Claude Opus 5.5 on September 22, 2026, framing it as the first model in a new Claude 5.5 family and the stronger default for long-running coding, agentic work, and professional knowledge tasks.
The headline is not just a higher benchmark table. Anthropic says Opus 5.5 performs around the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 on typical workloads. The API model ID is claude-opus-5-5, with a 1 million-token context window, 128K max output, adaptive thinking always on, and a default effort setting of medium.
That makes Opus 5.5 a practical release: faster and cheaper than the previous Opus flagship, but still pitched at sprawling jobs like codebase migrations, audits, and multi-step software tasks.
The benchmark embed
Anthropic's own launch table puts Opus 5.5 ahead of Opus 5, Fable 5.1, and major rivals across several coding and knowledge-work tests. The important caveat: Anthropic also says benchmark margins have become a less reliable guide to real-world differences at this level, so teams should still run private evaluations on their own code and workflows.
| Benchmark | Claude Opus 5.5 | Claude Fable 5.1 | Claude Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| FrontierCode v1.1 | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | - | 41.7% |
| GDPval-AA v2.1 | 1,846 Elo | 1,735 Elo | 1,708 Elo | 1,542 Elo | 1,588 Elo |
| AutomationBench | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
| Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% | 22.4% |
The table is useful because it shows where the model is strongest: agentic terminal coding, ambiguous multi-file coding sessions, and professional knowledge work. It also shows where it is not automatically first. GPT-6 Astra leads Anthropic's reported AutomationBench and Terminal-Bench-Science rows, which is a reminder that model selection should depend on workload, not brand momentum.
Independent benchmark pages also support the broad story, with caveats. Vals AI says Opus 5.5 ranked first on its overall Vals Index at launch and led several individual boards, including Terminal-Bench 2.1 and Terminal-Bench 4.0. Ridge also listed it at or near the top across tracked coding and arena-style metrics. Those leaderboards can move, but they are useful outside checks against the vendor launch post.
A real project-style example
The most concrete sample project in Anthropic's announcement is not a toy app. Anthropic says an internal test asked Opus 5.5 and Fable 5.1 to translate HAProxy from C to Rust. Both rewrites reportedly passed nearly all of HAProxy's own regression tests, but Opus 5.5 finished in 9.5 hours versus 12 hours for Fable 5.1 and cost 51% less.
That is the right kind of sample to pay attention to. It tests more than "can the model make a slick demo?" It checks whether a model can preserve behavior across a mature codebase, survive regression tests, and finish without turning the output into an expensive review burden.
Anthropic also says an early tester completed a 680,000-line code migration in under a day, and another used Opus 5.5 to audit and fix a 200,000-line codebase in under three hours. Those claims are impressive, but they are still reported anecdotes unless the underlying repos, prompts, diffs, tests, and review notes are shared publicly.
What changed for developers
The pricing change is straightforward:
| Price per 1M tokens | Claude Opus 5.5 | Claude Opus 5 |
|---|---|---|
| Input tokens | $4 | $5 |
| Output tokens | $20 | $25 |
| Cache reads | $0.20 | $0.50 |
| Cache writes | $5 | $6.25 |
The cache-read cut matters more than it looks. Agentic coding sessions often reuse large repo context, tool results, plans, and prior edits. If cache reads dominate the bill, reducing that line item from $0.50 to $0.20 per million tokens can change the economics of long sessions.
Sonar's independent evaluation points in the same direction. In its Java benchmark, Opus 5.5 produced roughly the same tested pass rate as Opus 5 while writing 27.5% less code and using 40% fewer output tokens. Sonar also found fewer total findings and lower code-smell and vulnerability density, while noting that bug density per line and concurrency findings rose.
That is the trade-off developers should remember: Opus 5.5 may be cleaner and more efficient overall, but it still needs tests, static analysis, and human review.
Migration gotchas
Anthropic's platform docs list four breaking changes for teams moving from Opus 5 to Opus 5.5:
- Thinking cannot be disabled.
- Forced tool use returns an error.
- Thinking blocks are tied to the model and conversation.
- On the Claude API and Google Cloud, the older
computer_20251124computer-use tool is not accepted.
The migration path starts with changing the model ID from claude-opus-5 to claude-opus-5-5, then checking any code that assumes thinking can be turned off, forces a specific tool call, or reuses thinking blocks across model changes.
Why it matters
Opus 5.5 looks less like a pure "bigger brain" release and more like a cost-performance correction. The useful story is not only that it scores higher. It appears to finish large coding jobs with fewer turns, fewer tokens, less generated code, and better review economics.
For teams already using Claude Code, GitHub Copilot with Claude, or the Claude API for software work, Opus 5.5 is worth a fast benchmark against real tasks. The best test is simple: pick three recent engineering tickets, run Opus 5 and Opus 5.5 under the same harness, then compare pass rate, edit size, test failures, review comments, wall-clock time, and total token cost.
If the public numbers transfer to your codebase, Opus 5.5 could become the new default for serious agentic coding. If they do not, the launch still offers a useful warning: frontier model releases should be judged by cost per accepted change, not just leaderboard rank.