Claude Opus 5.5 cuts prices and beats Fable 5.1
Anthropic cut input to $4 and output to $20 per million tokens, about 40 percent below Opus 5, and says the model also runs 30 percent faster.
In short
Anthropic released Claude Opus 5.5 on September 22, 2026, matching Fable 5.1-level performance at roughly 40 percent lower running cost.
At a glance
- Pricing per million tokens: $4 input (was $5), $20 output (was $25), $0.20 cache reads, $5 cache writes.
- Terminal-Bench 4.0: 66.4 percent for Opus 5.5 versus 55.8 percent for Fable 5.1.
- Context window 1M tokens, max output 128K tokens, output speed 30 percent faster than Opus 5 per Anthropic.
- Available on Claude Platform, AWS Bedrock, Google Cloud Vertex AI, Microsoft Azure Foundry and OpenRouter; model ID claude-opus-5-5.
- Limits: thinking can no longer be disabled; max effort can use up to 119,000 output tokens per task.
Anthropic shipped Claude Opus 5.5 on September 22, 2026, and calls it "the strongest-performing model" it has tested to date. Output falls to $20 per million tokens from $25 and input to $4 from $5, which the company frames as roughly 40 percent lower running cost than Opus 5. The model is live on the Claude Platform, AWS Bedrock, Google Cloud Vertex AI, Microsoft Azure Foundry and OpenRouter under the ID claude-opus-5-5.
Where the 40 percent actually comes from
The headline number does not fall out of the price list. Input and output each drop 20 percent. Caching is what moves: cache reads go to $0.20 per million tokens from $0.50, and cache writes to $5 from $6.25. Treat the 40 percent as a blended claim about typical workloads rather than a figure any single line item supports.
Benchmarks against Fable 5.1
Anthropic reports results from several suites. Opus 5.5 leads the larger Fable 5.1 in every run the company published, with the widest gap on Terminal-Bench 4.0.
| Benchmark | Opus 5.5 | Fable 5.1 |
|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% |
| Humanity's Last Exam | 67.7% | 65.6% |
| OSWorld 2.0 | 81.8% | 80.7% |
| FrontierCode v1.1 | 54.4% | 50.3% |
| CursorBench 4.0 | 57.8% | 51.8% |
Artificial Analysis scored the model 58 on its Intelligence Index, the highest figure it has recorded, according to the-decoder. That number appears in only one of the three reports reviewed here and has not been independently checked.
The "Claudish" complaint
One part of the launch is about prose, not compute. Claude output has drawn steady criticism for a formulaic, roundabout register that readers label "Claudish." Anthropic says Opus 5.5 leads with the important information, leans less on jargon and follows writing instructions more closely. No source published a measurement behind that claim.
What to check before migrating
The context window is 1M tokens with a 128K token output ceiling, and Anthropic puts output speed 30 percent above Opus 5. Thinking can no longer be disabled, which matters for existing API integrations built around toggling it. At the maximum effort setting the model can burn up to 119,000 output tokens on a single task. Cybersecurity, biology and frontier-model requests are routed to older models, and the weights are not being released.
Next in the line
Sonnet 5.5 and Haiku 5.5 are due in the coming weeks. None of the reports give a date.
FAQ
How much does Claude Opus 5.5 cost?
Per million tokens it is $4 for input and $20 for output, with $0.20 cache reads and $5 cache writes. Opus 5 charged $5 and $25, with $0.50 and $6.25 for cache.
Is Claude Opus 5.5 better than Fable 5.1?
On the runs Anthropic published, yes: 66.4 percent versus 55.8 percent on Terminal-Bench 4.0 and 81.8 percent versus 80.7 percent on OSWorld 2.0. These are vendor-reported figures.
Where can I use Claude Opus 5.5?
It is served through the Claude Platform plus AWS Bedrock, Google Cloud Vertex AI, Microsoft Azure Foundry and OpenRouter. The model ID is claude-opus-5-5, and there are no open weights.
