Opus 5.5: Fable 5.1-level performance, as claimed, at a lower list price

The price list is the part nobody can argue with: $4 per million input tokens and $20 per million output tokens, compared with $5 and $25 for Opus 5. Cache reads drop from $0.50 to $0.20, and cache writes from $6.25 to $5. Anthropic also claims a 40% reduction in the overall cost of use and output generation that is more than 30% faster. It offers a 'fast' mode alongside the model at $8/$40, which runs 2.5 times quicker. The model is available on the Claude platform, AWS, Google Cloud and Azure under the model ID claude-opus-5-5. For Sonnet 5.5 and Haiku 5.5 the company says 'the coming weeks' and gives no dates.

There are two sets of performance figures, and they don't say the same thing. Anthropic claims 66.4% on Terminal-Bench 4.0 (Opus 5: 52.3%), 54.4% on FrontierCode v1.1 against 50.3% for Fable 5.1, 57.8% on CursorBench 4.0, 81.8% on OSWorld 2.0 and 67.7% on Humanity's Last Exam. Artificial Analysis runs its own measurements. It gives the model 58 points on its Intelligence Index, the highest score it has recorded so far and above the previous leader, Fable 5.1. But it measures 59.6% on Terminal-Bench 4.0 and 61.4% on Humanity's Last Exam. That is a seven-point gap on the first and six on the second. The test conditions, meaning configuration and effort, are not specified, and without them the two columns can't be compared. They remain two separate measurements, not one disputed figure.

One more number belongs right next to the promised savings. On the tasks in its index, Artificial Analysis counted roughly 119,000 tokens used by Opus 5.5, compared with about 73,000 for Opus 5. A lower unit price on higher consumption gives a final bill that depends on what you ask the model to do. On safety, Anthropic claims the best result of any model tested so far in its own automated behavioural audit, which covers some two thousand scenarios. External evaluators, including METR and Frontier Design, were involved before release. In the same announcement, though, the company admits that the model 'often suspects it is being evaluated', and that “building evaluations that reliably catch every failure prior to deployment remains an unsolved problem”. For biology, Fable 5.1's safeguards apply, with verified access through the Life Sciences Verification Program. For cybersecurity, some requests are routed to an earlier model.

The timing is what held my attention longest. The release comes about two months after Opus 5. TechCrunch reports that it is Anthropic's first model since a public statement by its CEO: “I have become convinced that fully addressing the risks requires even more prudence”. ANSA notes that OpenAI and Anthropic unveiled new models within hours of each other. I don't see a contradiction here. The sources report the statement and the timeline, not a link between them, and anyone who draws one goes beyond what the sources say. Still, the line about evaluations that don't catch every failure is the most honest sentence in the announcement. It was written by the same company that claims the best result in its own behavioural audit for this model, and it admits that the yardstick used to measure these models is still being built. The 20% discount is written in the price list. Everything else is either claimed by the seller or measured by third parties under conditions nobody has spelled out.

— Olya