It is the best model Anthropic has tested so far on its internal alignment tests and comes with strong safeguards inherited from Fable.
40% cheaper on typical tasks
Opus 5.5 costs $4 per million input tokens and $20 per million in output, 20 percent below Opus 5. Cache reads fall 60 percent to $0.20 per million tokens.
Anthropic says typical workloads cost around 40 percent less overall, while using fewer tokens per task. Output is also more than 30 percent faster.
The efficiency gains become more obvious when measuring completed tasks rather than token prices, Anthropic says.
Leading the benches
At medium effort, Opus 5.5 scored 54.6 percent on FrontierCode, slightly above GPT-6 Astra’s best reported score of 53.3 percent, while Anthropic says it did so at around one-fifth of Astra’s cost per task.
On other benchmarks, the new Opus completes a tour de force, topping all other models in seven of nine popular benchmarks, specifically beating Fable 5.1 with more than ten percentage points in Terminal-Bench 4.0. Only two of the nine benchmarks posted by Anthropic have Astra winning.
Opus also jumps ahead of Fable 5.1 with a full five points on the Artificial Analysis benchmark of benchmarks, clocking in a healthy lead and the the top spot so far.
Anthropic says the model performs at roughly the level of its higher-end Fable 5.1 on most real-world work tasks, despite Opus 5.5 costing $4/$20 per million tokens compared with Fable’s $10/$50.
Opus 5.5 is available now through the Claude Platform and major cloud providers.
Read more: Anthropic’s presentation and launch thread. Reuters, TechCrunch and VentureBeat. Discussion on Hacker News and r/Singularity.