Anthropic launches Claude Opus 5.5 and cuts token prices

Anthropic released Claude Opus 5.5 on 22 September, and the news is not that it is smarter — it is that it is cheaper. The company says its new model performs at the level of its larger Claude Fable 5.1 on most work, while costing 40% less to run than Opus 5, the model it replaces [1]. For anyone paying for Claude by the token or running agents on it for hours at a time, that is the number that matters [2].

What was announced

Claude Opus 5.5 is the first model in what Anthropic calls a new Claude 5.5 family [1]. The company says it is the new leading model in the line-up, and that early testers saw large jumps on their most complex work — one finished a 680,000-line code migration in under a day, work the company says would have taken an engineering team weeks [1]. In another example, Anthropic says that when it asked the model to cut load times across every page of a web app, Opus 5.5 succeeded 39 of 40 times, while Opus 5 made smaller changes that also altered the app’s behaviour [1].

The pricing is the headline. Anthropic says input and output tokens cost $4 and $20 per million, which it describes as 20% less than Opus 5 [1]. Cache reads — which the company says make up the majority of agentic and coding costs — drop to $0.20 per million tokens, 60% less than Opus 5 [1]. The model also generates output more than 30% faster, and a Fast mode runs at up to 2.5x speed for $8 per million input tokens and $40 per million output tokens [1]. Anthropic frames the overall effect as a 40% drop in cost on typical workloads, because the model both costs less per token and uses fewer tokens per task [1].

On top of the price change, Anthropic says it is increasing the five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans, and giving subscription users a rate-limit reset they can save and use whenever they choose [1]. The company says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks with many of the same improvements [1]. TechCrunch reads the release as Anthropic pushing frontier-level coding performance down the price curve rather than chasing a new capability ceiling [2].

Anthropic also leans on efficiency in its benchmark framing, with the caveat that at this level of capability it considers benchmark margins a less reliable guide to real-world differences [1]. On its FrontierCode benchmark, the company says Opus 5.5 scores 54.6% at default effort, above GPT-6 Astra’s top score of 53.3%, for about a fifth of the cost per task [1]. On CursorBench, it reports 52.5% at default effort, compared with 51.8% for Fable 5.1 and 46.6% for Opus 5 at maximum effort, and says it beats GPT-5.6 Sol’s top score of 41.7% by 11 points for about a third of the cost [1]. In one internal test translating the HAProxy load balancer from C into Rust, Anthropic says Opus 5.5 finished in 9.5 hours against 12 for Fable 5.1, at 51% less cost [1].

What changed, and for whom

The shift here is about the cost of the same work, not a new tier of work. Anthropic says an early tester used Opus 5.5 to audit and fix a 200,000-line codebase in under three hours, where Opus 5 took over 20 hours and used 2.5 times as many tokens [1]. In a separate test, the company says it asked both models to analyse a proposed merger and build a financial model and presentation; both reached the same conclusion, but Opus 5.5 finished in 63 minutes against 93 for Opus 5 and cost 50% less to produce [1]. The pattern across the company’s own examples is fewer tokens, fewer steps, and less time for a result of similar quality.

On safety, Anthropic says Opus 5.5 posts the best scores of any of its models on its automated behavioural audit, an alignment suite that tests the model across thousands of simulated scenarios [1]. It also says the model is more resistant than Opus 5 to prompt injection, and ties Fable 5.1 for the lowest prompt-injection success rate on a benchmark run by the security firm Gray Swan [1]. But because Opus 5.5 is comparable to the company’s Mythos 5.1 model in biology and cybersecurity, Anthropic is deploying it with safeguards similar to those on Fable 5.1: vetted organisations can apply to its Life Sciences Verification Program to use the model for biology research, and it says it will expand access to a Cyber Verification Program in the coming weeks [1].

Who it is for — and not

Opus 5.5 is aimed squarely at people already spending real money on Claude: agent operators, coding teams, and anyone whose bill is dominated by long context and cache reads. For them, the sensible move is not to assume the savings appear on their own. Default model settings, pinned model versions, and effort levels all decide whether you actually get the cheaper, faster path or keep paying Opus 5 rates. Anthropic’s own framing — that the clearest advantage is efficiency rather than a jump in raw capability — is a reason to treat this as a cost review, not a reason to hand an agent more autonomy [1].

It is less relevant if you use Claude lightly through a chat interface, where the per-token price is not what you feel. And the safety terms are a real constraint for a narrow group: because the model routes biology and cyber work through stricter safeguards, teams doing legitimate work in those areas may hit classifier friction and need the verification programs Anthropic describes [1]. As always with a vendor launch, the benchmark numbers and tester quotes here are the company’s own; the useful parts are the prices and limits, which are concrete and checkable [1][2].

We will update this page if Anthropic publishes the Opus 5.5 system card details, or when Sonnet 5.5 and Haiku 5.5 ship with their own prices.

∴ Anthropic is competing on price and efficiency now: same-tier coding work for roughly 40% less than Opus 5.