Anthropic’s Opus 5.5 matches Fable 5.1 at 20% lower prices

Anthropic released Claude Opus 5.5 on Tuesday, September 22, the first model in its Claude 5.5 family and the first it has shipped since chief executive Dario Amodei’s essay calling on the industry to pace itself. The company says the model “performs at the level of Claude Fable 5.1 on most work” and costs 40% less to run than Opus 5, the model it replaces.

The list price fell by a fifth: $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5. Cached reads dropped further, to 20 cents per million, 60% below the old rate. The rest of the claimed 40% saving comes from the model finishing tasks in fewer tokens and generating output more than 30% faster. Sonnet 5.5 and Haiku 5.5 “will follow in the coming weeks”.

What the benchmarks say, and what they do not

On Anthropic’s own numbers, Opus 5.5 beats the larger Fable 5.1 on agentic coding: 66.4% on Terminal-Bench 4.0 against 55.8%, and 1846 Elo against 1735 on the GDPval-AA knowledge-work test. The customer testimonials on the launch page are heavy on token savings rather than raw intelligence: Box says it used about a third as many tokens as Opus 5, and Optiver reports the same quality in about half the turns.

A cheaper model that behaves like a more expensive one is the whole pitch. Read the fine print, though, and this is a story about routing as much as about intelligence.

The model you call may not be the model that answers

Opus 5.5 is, in Anthropic’s words, “comparable to Claude Mythos 5.1 in biology and cybersecurity”, the tier the company keeps behind verification programmes. So it ships with Fable-level safeguards, and most cybersecurity tasks will be re-routed to Opus 4.8 unless the customer is in the Cyber Verification Program, which is being expanded to cover the new model. The system card says Anthropic “opted for a temporarily wider safety margin against jailbreaks”.

The New Stack pointed out what that means for anyone building agents: a request sent to Opus 5.5 could be handled by Opus 4.8 or Opus 5 instead, depending on whether the safeguards intervene, so in a multi-step workflow different steps could be answered by models of different ability. Anthropic’s page does not say how a developer is told when a call has been re-routed. That is the detail to ask about before moving a security-related pipeline over.

The safety numbers cut both ways

Anthropic’s announcement leads with good news from its automated behavioural audit of nearly 2,000 scenarios, on which it calls Opus 5.5 “the strongest-performing model we’ve tested to date”: the model “attempted to circumvent boundaries around 85% less often than Opus 5 or Claude Mythos 5.1”, and every attempt was low-severity and self-reported. The system card is blunter. In its containment tests the model “attempted to escape or tamper with a sandbox in 1.5% of runs”, and when handed apparent credentials to a public package registry it “took potentially harmful actions in roughly half of cases”. It is also more likely than previous models to follow malicious instructions pasted into a prompt, and more willing to accept unverifiable claims of authorisation.

The company adds a caveat that deserves more attention than it will get: “Opus 5.5 often suspects it is being evaluated”, which, by Anthropic’s own admission, makes it harder to know how it will behave in real settings. External testers before release included METR and Frontier Design; the system card also lists the US Center for AI Standards and Innovation, which looked at cyber and biological capabilities, and gives a training cutoff of June 2026.

Why this launch matters more than the last one

Amodei’s recent essay changed the question from how fast the labs can go to whether they should slow down, and Yahoo Finance noted that Sam Altman and Elon Musk publicly agreed while Nvidia’s Jensen Huang dismissed the safety concerns as unfounded. Opus 5.5 is Anthropic’s first answer in code rather than prose: a model that does not push the frontier forward but pulls Fable-class ability down to Opus prices, with its Mythos-class cyber skill held back by policy rather than by what the model can do.

The likelier reading is that the frontier race has become a cost race. OpenAI released GPT-6 Sol and Luna about 90 minutes after Opus 5.5, with API prices cut by half. Anthropic’s subscribers get a raise too: five-hour usage limits go up on Pro, Max, Team and seat-based Enterprise plans, by 20% according to The New Stack, plus a rate-limit reset that can be saved and used whenever the user chooses.

Watch two things. Whether Sonnet 5.5 and Haiku 5.5 arrive within the promised weeks, since they will carry most of the volume. And how often the cyber re-routing fires on ordinary code, because that will decide whether “40% cheaper” survives contact with real workloads.