Claude Opus 5.5
Language Available Full comparison ↗ anthropic/claude-opus-5-5 · by Anthropic
· decoder-only-transformer
Pricing — 1 offering(s)
About this offering: Announced 2026-09-22 as the first Claude 5.5 model; API id claude-opus-5-5 (a pinned snapshot). Opus 5 remains available. Rates not shown as tiers: 1-hour cache writes $8 per 1M tokens, and US-only inference (inference_geo: "us") at 1.1× on all categories. 1M-token context at standard pricing and 128K max output (300K on the Batches API with the output-300k-2026-03-24 beta header). Default effort is medium (Opus 5's is high), and thinking is adaptive and always on.
Input tokens
- $4.00 / 1M tokens (input)
Current
2026-09-22 → present
$1 below Opus 5's $5. The same figure appears in the models-overview table and the launch post.
Output tokens
- $20.00 / 1M tokens (output)
Current
2026-09-22 → present
$5 below Opus 5's $25. The same figure appears in the models-overview table and the launch post.
Cache hits and refreshes
- $0.20 / 1M tokens (cached input)
Current
2026-09-22 → present
Cache hits and refreshes: $0.20 per 1M tokens, 0.05× the base input rate (most models use 0.1×).
5-minute cache writes
- $5.00 / 1M tokens (cache write)
Current
2026-09-22 → present
5-minute cache writes ($5 per 1M tokens, the standard 1.25× input). 1-hour cache writes cost $8 per 1M (2×).
Fast mode input tokens
- $8.00 / 1M tokens (input)
Current
2026-09-22 → present
Fast mode: up to 2.5× faster output at a premium. Available on the Claude API (first-party) and on Claude in…
Fast mode: up to 2.5× faster output at a premium. Available on the Claude API (first-party) and on Claude in Microsoft Foundry for Azure-hosted deployments; Anthropic's pricing page says it "is not available on Claude Platform on AWS or partner-operated cloud platforms" and "is not available with the Batch API". No longer a research preview for Opus 5.5 (changed in the 2026-10-16 pricing page update, observed 2026-10-20).
Fast mode output tokens
- $40.00 / 1M tokens (output)
Current
2026-09-22 → present
Fast mode: up to 2.5× faster output at a premium. Available on the Claude API (first-party) and on Claude in…
Fast mode: up to 2.5× faster output at a premium. Available on the Claude API (first-party) and on Claude in Microsoft Foundry for Azure-hosted deployments; Anthropic's pricing page says it "is not available on Claude Platform on AWS or partner-operated cloud platforms" and "is not available with the Batch API". No longer a research preview for Opus 5.5 (changed in the 2026-10-16 pricing page update, observed 2026-10-20).
Batch
Batch input tokens
- $2.00 / 1M tokens (input)
Current
2026-09-22 → present
Message Batches API rate: 50% of the standard price.
Batch output tokens
- $10.00 / 1M tokens (output)
Current
2026-09-22 → present
Message Batches API rate: 50% of the standard price.
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Claude Opus 5.5 fits into a cost-aware routing setup
See how →About Claude Opus 5.5
Claude Opus 5.5 is the first model of Anthropic's Claude 5.5 family, announced on 2026-09-22 and published with a system card the same day. It is a decoder-only transformer trained with large-scale pretraining, instruction tuning and preference optimisation. Thinking is adaptive and always on, steered by a five-level `effort` parameter that defaults to medium.
On release it became Anthropic's recommended starting model for most workloads, and Anthropic moved Claude Opus 5 and Claude Fable 5 to its "legacy models (still available)" list. API pricing dropped to $4 / $20 per million input / output tokens, down from Opus 5's $5 / $25, with cache reads at 5% of the base input price. Anthropic says the model needs less compute to serve than Opus 5.
Four API changes break Opus 5 code: thinking can't be disabled; forced tool choice is rejected; thinking blocks are readable only by Opus 5.5, Fable 5.1 and Mythos 5.1; and the older `computer_20251124` computer-use tool is rejected on the Claude API and Google Cloud in favour of `computer_toolset_20260801`.
Capability profile
How Claude Opus 5.5 rates across core capability dimensions, with the task-level evidence behind each rating.
Humanity's Last Exam (2,500 expert-written questions): 64.4% with no tools and 67.7% with web search, web fetch, programmatic tool calling and code execution, both at max effort averaged over five trials. Vendor-reported (system card Table 8.1.A).
Terminal-Bench 2.1 (89 real terminal tasks; Terminus 2 harness): completes 87.6%, so it fails roughly 1 task in 8. Terminal-Bench 4.0 (66 long-horizon multi-domain tasks; mini-swe-agent): 65.2%, the top score on the board. Independent (Vals AI; both runs used Opus 5 or Opus 4.8 as refusal fallbacks, and counting those as failures gives 79.8% and 58.1%). Anthropic's own vendor-reported figures: Terminal-Bench 4.0 66.4% (xhigh effort) and SWE-bench Pro 89.9% (max effort, no harness named).
Positioned by Anthropic 'for long-running agentic coding and knowledge work'. AutomationBench 40.0% (run and reported by Zapier, no fallback models) and OSWorld 2.0 81.8% partial / 48.7% strict (vendor-reported). Forced tool choice (`any`/`tool`) is not supported: Anthropic recommends strict tool use or structured outputs instead.
1M-token context window and 128K max output (models-overview table; the migration guide says Opus 5.5 'keeps Claude Opus 5's 1M token context window and 128k max output tokens'). The system card (Section 8.10) also describes episodes 'up to the full 1M token window.'
Comparative latency is listed as 'Moderate' in Anthropic's models-overview table. Fast mode is up to 2.5x faster at $8/$40 per MTok. It's available on the Claude API and Claude in Microsoft Foundry, not with the Batch API, and is no longer a research preview since the 2026-10-16 pricing page update.
$4/$20 per MTok, 20% below Opus 5, with cache hits at 0.05x base input ($0.20/MTok) rather than the usual 0.1x. Anthropic says it also uses fewer tokens per task, which 'nets out to a 40% drop in costs' against Opus 5 (vendor claim). Launch post: at default (medium) effort it beats GPT-6 Astra's top FrontierCode score for about a fifth of the cost per task.
Coding benchmarks
View the full coding benchmark leaderboard →| Benchmark | Score | Harness | Source |
|---|---|---|---|
| terminal-bench-2-1 | 87.6% 2026-09 | terminus-2 | 📊 leaderboard ↗ |
| terminal-bench-4-0 | 65.2% 2026-09 | mini-swe-agent | 📊 leaderboard ↗ |
Operator guidance
Anthropic's recommended starting model "for most workloads" (models-overview). Route here in place of Opus 5 for demanding coding, agentic and knowledge work; Opus 5 stays available but is now listed as legacy. Escalate to Claude Fable 5.2 only when evals at higher Opus 5.5 effort still fall short (that's Anthropic's own guidance). Set `effort` explicitly: the default is `medium`, where Opus 5's is `high`. Don't route requests that need forced tool choice or thinking disabled (both return an error). Drop to Sonnet 5 for the usual price/performance balance, or to Haiku 4.5 for high-volume, latency-sensitive work.
Use cases
- Long-running agentic coding and knowledge work (Anthropic's own stated positioning)
- Codebase-wide migrations and audits (launch-post examples: a 680,000-line migration in under a day; a 200,000-line audit in under three hours)
- Workloads previously routed to Opus 5 where cost matters: it's cheaper per token and uses fewer tokens per task
Limitations
- Every benchmark figure here is vendor-reported (launch post and system card), apart from AutomationBench, which Zapier ran and reported, and GDPval-AA v2.1, which is Artificial Analysis's independent leaderboard (see below).
- The system card reports Opus 5 at 56.6% (no tools) / 63.6% (with tools) on HLE, versus 56.3% / 64.7% in Opus 5's own July system card. Anthropic's re-runs shift by roughly 1 point, so treat small cross-model HLE gaps with care.
- GDPval-AA v2.1: 1,866 Elo at max effort, from Artificial Analysis's independent leaderboard (verified 2026-10-10). Anthropic's launch table quotes 1,846, which was AA's figure when checked on 2026-09-23. v2.1 is anchored to DeepSeek V4.1 Flash (max) = 1,600 with no human baseline, so don't compare it with GDPval-AA v2 figures, which used a different scale.
- Anthropic's Terminal-Bench 4.0 figure (66.4%, xhigh effort) is vendor-reported; the independent single-harness run (Vals AI, 65.2%) is the structured coding score. SWE-bench Pro (89.9%) was published without a harness name, so it isn't recorded as a structured score.