Claude Opus 5.5 vs Claude Opus 5
Opus 5.5 lowers standard input and output token prices by 20%. Use Opus 5 as a baseline and check for regressions on your existing tasks.
Reviewed 2026-09-25 · Official API facts and attributed benchmark results
| Specification | Claude Opus 5.5 | Claude Opus 5 |
|---|---|---|
| Context window | 1M | 1M |
| Max output | 128K | 128K |
| Input / 1M tokens | $4 | $5 |
| Output / 1M tokens | $20 | $25 |
| Cache read / 1M tokens | $0.2 | $0.5 |
| Reasoning | Adaptive, always on · medium by default | Adaptive · high by default |
When to evaluate Claude Opus 5.5
A starting point for evaluating coding and agent workloads, with lower token rates than Opus 5 or Fable 5.1.
Benchmark settings differ from the default medium effort. Measure task completion and total tokens on your own workload.
Standard API rates. Batch input/output is 50% lower; fast mode is $8 / $40. Cache writes: $5 (5 min), $8 (1 hour) per million tokens.
Official specificationsWhen to evaluate Claude Opus 5
Useful as a regression baseline for applications already tuned to Opus 5.
Anthropic marks this as a legacy model and recommends considering Opus 5.5. Retest prompts, tool calls and costs before migrating.
Standard API rates. Batch input/output is 50% lower. Cache writes: $6.25 (5 min), $10 (1 hour) per million tokens.
Official specificationsReported benchmark results
| Benchmark | Claude Opus 5.5 | Claude Opus 5 |
|---|---|---|
| Terminal-Bench 4.0Opus 5.5: xhigh effort | 66.4% | 52.3% |
| FrontierCode v1.1Opus 5.5: max effort | 54.4% | 48% |
| CursorBench 4.0Opus 5.5: max effort | 57.8% | 46.6% |
| GDPval-AA v2.1Opus 5.5: max effort | 1846 Elo | 1708 Elo |
| AutomationBenchOpus 5.5: max effort | 40% | 26.9% |
| Humanity's Last ExamWith tools · Opus 5.5: max effort | 67.7% | 63.6% |
Vendor-reported launch results, September 22, 2026. Opus 5.5 uses max effort except Terminal-Bench (xhigh). Other models' effort, harness and tool settings can differ; see the original methodology. These are not AI IDE List tests and are separate from Anthropic's medium-effort cost charts.
Benchmark source and methodologyTry the work that matters to you
Use the same task, inputs, tools and success criteria. Record the model version, effort, total cost, latency and any human fixes. Repeat tasks before choosing a model; a single demo does not establish a general winner.