A cheaper model at the top of the range
Anthropic released Claude Opus 5.5 on Tuesday and said it performs at a level similar to Claude Fable 5.1, the company’s largest model, on most work. The pricing moved in the other direction: input falls to $4 per million tokens from $5, and output to $20 from $25.
The steeper cut is on cached input, which drops from $0.50 to $0.20 per million tokens — a 60 per cent reduction that matters most to agents replaying long context on every step. Anthropic puts the combined effect at about 40 per cent less on typical workloads, and says the model generates output roughly 30 per cent faster than Opus 5.
The numbers, and whose they are
Every figure below is Anthropic’s own, measured on its own harness, and none of it has been reproduced independently yet.
The company reports Terminal-Bench 4.0 at 66.4 per cent, against 52.3 per cent for Opus 5. On FrontierCode v1.1 it gives 54.4 per cent, up from 48.0 per cent. OSWorld 2.0, which tests a model driving a desktop, rises to 81.8 per cent from 74.0 per cent. On GDPval-AA v2.1, an Elo-scored evaluation of knowledge work, Anthropic puts Opus 5.5 at 1846 against Opus 5’s 1708.
Anthropic told TechCrunch it is “the strongest-performing model we’ve tested to date”. The interesting claim is not the individual scores but the shape of them: a mid-cycle release closing on the flagship rather than inching past its own predecessor.
Safety work shipped with it
Anthropic says Opus 5.5 scored best of its models on the company’s automated behavioural audit, and that attempts to circumvent containment boundaries fell 85 per cent against Opus 5. It also claims improved resistance to prompt injection, which is the failure mode that has produced most of this year’s agent incidents.
The safeguards around the model are unusually specific. Cybersecurity tasks are routed to the older Opus 4.8 unless the user has cleared an expanded Cyber Verification Program. Biology capability is described as comparable to Claude Mythos 5.1, gated behind a new Life Sciences Verification Program. Extended thinking is now mandatory rather than optional, and the anti-distillation safeguard that hides reasoning traces has been kept.
Where it runs
The model is available from launch on Anthropic’s own platform as claude-opus-5-5, and on Amazon Web Services, Google Cloud and Microsoft Azure. GitHub added it to Copilot the same day. Anthropic says zero data retention is available and that the output carries the watermarking required under the EU AI Act.
The release lands three days after Anthropic’s chief executive Dario Amodei argued publicly for slowing the pace of capability work so that alignment research can catch up, and on the same day OpenAI cut its own prices with GPT-6 Sol and Luna. Both companies are now competing on cost at the frontier rather than on capability alone.
What to watch
Claude Sonnet 5.5 and Claude Haiku 5.5 are promised “in the coming weeks”, which would carry the same price and latency changes down the range. The other thing to watch is independent measurement: Anthropic’s benchmark table is a vendor’s account of its own product, and the model’s real position against Fable 5.1 and against OpenAI’s new tier will not be settled until outside evaluators publish.
