Anthropic
CurrentClaude Opus 5.5
Anthropic's model for long-running agentic coding and knowledge work, with a 1,000,000-token context window.
Anthropic describes it as
“For long-running agentic coding and knowledge work”
Last verified:
Key facts
- Context window
- 1,000,000
- Max output
- 128,000
- Input price
- $4.00
- Output price
- $20.00
- Extended output (Batch API (beta))
- 300,000
- Released
- Sep 22, 2026
- Knowledge cutoff
- Jun 2026
- API model ID
- claude-opus-5-5
tokens
tokens, default
per 1M tokens, standard
per 1M tokens, standard
tokens, special mode only
- Batch API (beta): Message Batches API only, with the output-300k-2026-03-24 beta header. Synchronous requests are limited to 128,000 output tokens.
- Anthropic's pages list a context window and a maximum output but do not state a separate maximum-input figure.
About Claude Opus 5.5
Claude Opus 5.5 is Anthropic's model for long-running agentic coding and knowledge work. It has a 1,000,000-token context window and up to 128,000 output tokens, takes text, image and PDF input, and always uses adaptive thinking. It costs $4 input and $20 output per million tokens, with the full 1M context at standard pricing.
API pricing
Per-token API prices from the provider's pricing page. Thresholds and processing modes are shown explicitly, not folded into one number.
| Rate | Price |
|---|---|
| Input | $4.00 |
| Cached input | $0.20 |
| Output | $20.00 |
| Cache write (5 minutes) | $5.00 |
| Cache write (1 hour) | $8.00 |
| Rate | Price |
|---|---|
| Input | $2.00 |
| Output | $10.00 |
| Rate | Price |
|---|---|
| Input | $8.00 |
| Output | $40.00 |
- Cache write prices are shown for 5-minute and 1-hour retention.
- Setting inference_geo to "us" applies a 1.1x multiplier on Claude 4.6 and later models.
- Fast mode is a premium-priced option at $8 input and $40 output; it is not available with the Batch API.
- Anthropic includes the full 1M-token context window at standard pricing for this model.
Modalities and capabilities
What the provider documents for this model. “API tool” means it needs a provider tool or endpoint rather than being native to the model.
- Input modalities
- Text, Image, PDF
- Output modalities
- Text
- ReasoningReasoning: ✓ Yes
Adaptive thinking, always on and can't be turned off.
- Tool / function callingTool / function calling: ✓ Yes
- Structured outputsStructured outputs: ✓ Yes
- Web searchWeb search: ✓ Yes · API tool
Web search tool.
- Code executionCode execution: ✓ Yes · API tool
Code execution tool.
- Computer useComputer use: ✓ Yes · API tool
On the Claude API, through computer_toolset_20260801.
- Image generationImage generation: ✕ No
Output is text only.
- Prompt cachingPrompt caching: ✓ Yes
Minimum cacheable prompt length is 512 tokens.
- Batch APIBatch API: ✓ Yes
Benchmarks
Individual named benchmarks only: there is no overall score. Vendor-reported results come from the model maker's own launch materials. Scores from different sources, versions or settings are not directly comparable.
| Benchmark | Score | Source type | Reported by |
|---|---|---|---|
| CursorBench (v4.0)Setting: max effort | 57.8% | Vendor-reported | AnthropicSep 22, 2026 |
| FrontierCode (v1.1, Main)Setting: max effort | 54.4% | Vendor-reported | AnthropicSep 22, 2026 |
| GDPval-AA (v2.1)Setting: max effort | 1846 | Vendor-reported | AnthropicSep 22, 2026 |
| Humanity's Last Exam (with tools)Setting: max effort | 67.7% | Vendor-reported | AnthropicSep 22, 2026 |
| OSWorld (v2.1, partial credit)Setting: max effort | 81.8% | Vendor-reported | AnthropicSep 22, 2026 |
| Terminal-Bench (v4.0)Setting: xhigh effortStandard error ±2.6 points, per Anthropic. | 66.4% ±2.6 | Vendor-reported | AnthropicSep 22, 2026 |
Availability and API identifiers
| Surface | Identifier |
|---|---|
| Claude APIEvery Claude model ID is a pinned snapshot; there is no separate dated ID. | claude-opus-5-5 |
| Amazon Bedrock | anthropic.claude-opus-5-5 |
| Google Cloud (Google-managed and Anthropic-managed) | claude-opus-5-5 |
| Microsoft Foundry | claude-opus-5-5 |
| Claude Platform on AWS | claude-opus-5-5 |
Limitations and caveats
API vs. apps. These limits and prices describe the Claude API model. The Claude apps can apply different limits and are billed through subscriptions, not per token.
- Anthropic lists retirement as not sooner than September 22, 2027.
- Anthropic says Claude 4.7 and later models produce approximately 30% more tokens for the same text than earlier Claude models.
- Adaptive thinking can't be turned off, and forced tool use returns an error.
- The computer_20251124 computer-use tool is not accepted on the Claude API; use computer_toolset_20260801.
- Anthropic's launch page says Opus 5.5 was evaluated with production safeguards on, which likely lowers its benchmark scores.
Sources
Last verified against these sources on Oct 31, 2026. Prices and limits change; follow the links for the live figure.
- Anthropic docs — Claude Opus 5.5
Supports: status, release date, knowledge cutoff, context window, max output, modalities, pricing
OfficialChecked Oct 31, 2026
- Anthropic docs — Models overview
Supports: status, context window, max output, modalities, capabilities
OfficialChecked Oct 31, 2026
- Anthropic docs — Pricing
Supports: pricing
OfficialChecked Oct 31, 2026
- Anthropic docs — Structured outputs
Supports: capabilities
OfficialChecked Oct 31, 2026
- Anthropic docs — Web search tool
Supports: capabilities
OfficialChecked Oct 31, 2026
- Anthropic docs — Computer use tool
Supports: capabilities
OfficialChecked Oct 31, 2026
- Anthropic docs — Code execution tool
Supports: capabilities
OfficialChecked Oct 31, 2026
- Anthropic docs — PDF support
Supports: modalities, capabilities
OfficialChecked Oct 31, 2026
- Anthropic docs — Prompt caching
Supports: capabilities
OfficialChecked Oct 31, 2026
- Anthropic docs — Batch processing
Supports: capabilities
OfficialChecked Oct 31, 2026
- Anthropic — Claude Opus 5.5 launch page
Supports: benchmarks
Anthropic launch page; scores are Anthropic-reported.
Vendor benchmarkChecked Oct 31, 2026