Anthropic released Claude Opus 5.5 on September 22, 2026, positioning it as a daily driver for coding and knowledge work with Fable 5.1-level performance on most tasks and lower running cost than Opus 5.
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. That is 20% below Opus 5’s token prices, while Anthropic estimates typical workloads cost about 40% less because the new model also uses fewer tokens. Cache reads fall more sharply to $0.20 per million tokens, 60% below Opus 5.
What changed from Opus 5
- Standard input price falls from $5 to $4 per million tokens.
- Standard output price falls from $25 to $20 per million tokens.
- Cache reads cost $0.20 per million tokens instead of $0.50.
- Anthropic says typical task cost is about 40% lower because Opus 5.5 is more token-efficient.
- Fast mode offers up to 2.5 times higher speed at $8 input and $40 output per million tokens.
The distinction between price per token and price per task is important. A 20% rate cut does not automatically create a 40% task-cost reduction. Anthropic’s larger estimate depends on the model completing comparable work with fewer output tokens, turns and tool calls. Teams should verify that effect on their own tasks.
The intended position is below Fable, not below the old Opus
Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 for most work while costing less. That makes the launch a product-line compression: organizations may be able to reserve Fable for the tasks where its extra capability is measurable and move ordinary advanced work to Opus.
This is not a universal equivalence claim. Anthropic still describes Fable 5.1 as the option for demanding reasoning and long-horizon agentic work. Opus 5.5 is the recommended daily driver for production coding, complex documents, enterprise workflows and computer use. The correct routing boundary is where a team’s own evaluation shows Fable’s additional quality pays for its higher cost.
Where Anthropic says Opus 5.5 improved
| Area | Launch claim | What to test internally |
|---|---|---|
| Agentic coding | Stronger long-running work and root-cause analysis | Merge readiness, regressions and reviewer corrections |
| Multi-tool agents | Better planning, subagent coordination and memory use | Retries, permission errors and completion rate |
| Documents | Professional spreadsheets, slides and docs | Formula accuracy, citations and edit time |
| Computer use | Improved visual interpretation and multi-app planning | Task success under realistic browser state |
| Writing | Clearer, less verbose communication | Whether outputs follow house style without extra cleanup |
Benchmark results need their footnotes
Anthropic’s launch page places Opus 5.5 at 57.8% on CursorBench 4.0 for agentic coding and 1,846 on GDPval-AA v2.1 for knowledge work. It also reports 40.0% on AutomationBench, 67.7% on Humanity’s Last Exam with tools and 81.8% partial reward on OSWorld 2.0.
Those numbers are useful only with the methodology. Anthropic says Opus 5.5 results generally use adaptive thinking at max effort. Terminal-Bench 4.0 uses xhigh effort. Some cybersecurity, biology and frontier-development tasks fall back to earlier models when safeguards intervene, and the company says this can lower reported performance. Scores also carry sampling error, and competitor values may come from the competitor’s own reports.
The cache-read cut may matter most for agents
Long-running agents repeatedly read instructions, tool definitions, repository context and prior state. Dropping cache reads from $0.50 to $0.20 per million tokens can therefore reduce cost more than the 20% base-token cut suggests. A system with a stable prefix and many turns benefits more than a short one-shot request.
For teams building coding workflows, this reinforces a lesson from our Codex versus Claude Code comparison: model quality, context reuse, permissions and the agent harness should be measured together. The cheapest model call is not necessarily the cheapest completed change.
Availability and model ID
Opus 5.5 is available in Claude for Pro, Max, Team and Enterprise users. Developers can use it on the Claude Platform with the model ID claude-opus-5-5. Anthropic also lists availability through Amazon Web Services, Google Cloud and Microsoft Foundry. US-only inference carries a 1.1 times price multiplier for input and output.
Fast mode is available in Claude Code and on the Claude Platform. It doubles the standard token rates in exchange for up to 2.5 times faster output. That option is useful for interactive work where waiting is more expensive than tokens, but background jobs may be better served by standard mode or batch processing.
A disciplined upgrade test
- Freeze the existing Opus 5 prompt, tools and permissions.
- Replay a balanced set of successful, failed and edge-case tasks on Opus 5.5.
- Record input, cache-read and output tokens separately.
- Measure tool calls, elapsed time and human correction effort.
- Run Fable 5.1 only on tasks where Opus 5.5 misses the acceptance threshold.
- Set a rollback path and monitor the first production week.
Teams migrating from older Claude versions should also track model retirement dates and provider-specific IDs. Our report on GitHub Copilot model retirements shows why a model change can become an operational project when aliases, policy controls and enterprise approvals differ across surfaces.
Primary sources
- Anthropic: Claude Opus product and launch page
- Claude Opus 5.5 System Card
- Claude’s official Opus 5.5 launch post on X
Checked September 22, 2026. Benchmark and cost-per-task comparisons are Anthropic’s reported results unless otherwise stated.
