Opus 5.5 field notes

Released September 22, 2026

Fable-class work at a lower Opus price.

Claude Opus 5.5 is Anthropic's new model for long-running agentic coding and knowledge work. Per-token prices drop 20% from Opus 5, cache reads drop 60%, and Anthropic estimates typical workloads cost about 40% less.

  • $4 / $20 per M input / output
  • $0.20 per M cache reads
  • 30%+ faster output
  • Model ID claude-opus-5-5
See the full rate card Plan the migration

Price one agent run

Move the sliders to match your workload. Figures are millions of tokens.

Opus 5

Opus 5.5

The rate card, side by side

Every line got cheaper, but not by the same amount. Cache reads fell the most, and those make up the bulk of a coding agent's bill.

Per million tokensOpus 5Opus 5.5Change
Input$5.00$4.00−20%
Output$25.00$20.00−20%
Cache write, 5 minutes$6.25$5.00−20%
Cache write, 1 hourn/a$8.00
Cache read$0.50$0.200.05× base input−60%
Batch input / outputn/a$2 / $10half price
Fast mode input / outputn/a$8 / $40up to 2.5× faster

Thinking tokens bill as output. US-only inference (inference_geo: "us") is reported to add a 1.1× multiplier. Fast mode is a research preview on the Claude API only. Check Anthropic's pricing page before you budget.

What's new beyond the price

Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work and leads on several agentic coding, computer use, and knowledge work benchmarks.

Speed

30%+ faster output
Generates tokens faster than Opus 5 at standard pricing.
Fast mode preview
Up to 2.5× the output speed for 2× the price. Set speed: "fast" with the fast-mode-2026-02-01 header.

Thinking and effort

Adaptive thinking, always on
Control depth, latency, and cost with the effort parameter instead of a token budget.
Default effort is now medium
Opus 5 defaulted to high. Set effort explicitly and re-run your evals.
Per-message effort beta
Change effort mid-conversation for cheap turns and hard turns.

Long-running agents

Compact on demand beta
Request a signed summary of the whole conversation when you choose, even in the background.
Tools defined in a message beta
Add or change a tool mid-conversation without losing the prompt cache.
Task budgets and mid-conversation system messages
Steer a session without editing earlier turns.
512-token cache minimum
Smaller prompts qualify for caching.

Vision and computer use

Sharper chart and screenshot reading
Reads values off dense charts and layouts more precisely without tools.
Computer use toolset
On the Claude API and Google Cloud, use computer_toolset_20260801.
Files, PDFs, batch, server and client tools
All supported.

Moving from Opus 5

Four changes can fail requests that work today. Do these in order.

  1. Swap the model ID

    Point your client at the new model.

    - model: "claude-opus-5"
+ model: "claude-opus-5-5"

  2. Remove thinking overrides

    thinking: {"type": "disabled"} and manual budget_tokens now return a 400. Omit the field or send {"type": "adaptive"}, then tune effort.

  3. Replace forced tool use

    tool_choice of any or tool returns a 400. Use auto with strict tool use, or structured outputs for JSON.

  4. Update computer use

    On the Claude API and Google Cloud, the older computer_20251124 tool is rejected. Move to the toolset. Bedrock still accepts the old tool.

  5. Read thinking blocks for progress text

    Notes between tool calls now arrive as thinking blocks, empty by default. Set thinking.display if your UI streams them. Select blocks by type, not position, and handle stop_reason: "refusal" with a fallback.

Where you can use it

Available now in the Claude apps, Claude Code, and these platforms.