Anthropic ships the Claude 5.5 family: flagship Opus 5.5 costs 40% less to run
On September 22, Anthropic released Claude Opus 5.5, the first model in its new Claude 5.5 family. The pitch: it performs at the level of Claude Fable 5.1 on most work while costing about 40% less to run than Opus 5. It is also Anthropic's first release since it called for pacing the frontier, and it was tested before release by external evaluators including Frontier Design and METR. The focus is clear — long-horizon coding and agentic workloads.
Price and speed: where the cut lands
Start with the price list. Input fell from $5 to $4 per million tokens and output from $25 to $20. The biggest move is on caching: cache reads dropped from $0.50 to $0.20, and cache writes from $6.25 to $5. Anthropic singles out cache reads as the bulk of the cost in agentic and coding work — that line is down 60%.
The lower price rests on efficiency rather than a bare sticker change: Opus 5.5 needs less compute to serve and also uses fewer tokens per task, and the two together net out to the 40% drop. On speed, Anthropic says it generates output more than 30% faster than Opus 5. Separately, a fast mode is available in Claude Code and the Claude Platform with up to 2.5x speed, priced at $8 per million input tokens and $40 per million output tokens.
Cache deserves a word of explanation. Every step an agent takes re-sends the system prompt, tool definitions and conversation history to the model, and much of that prefix never changes. A provider can store that prefix and reuse it, billing the hits at a far lower rate — which is why the price of a cache read often drives the final bill for an hour of agent work. Cutting that price from $0.50 to $0.20, while also using fewer tokens per task, is what adds up to the 40%.
Alongside the pricing come usage changes: Anthropic raised the five-hour limits on Pro, Max, Team and seat-based Enterprise plans, and is giving subscription users a rate-limit reset that they can save and use whenever they choose. For teams running agents for long stretches, those matter as much as the price cut.
Benchmarks and real workloads: cheap does not mean weak
On the published evaluations, Opus 5.5 leads in agentic coding, computer use and knowledge work: Terminal-Bench 4.0 at 66.4% against Fable 5.1's 55.8%, Opus 5's 52.3%, GPT-6 Astra's 57.9% and GPT-5.6 Sol's 37.3%; FrontierCode v1.1 (Main) at 54.4%; CursorBench 4.0 at 57.8%; and, on the knowledge-work side, 1846 Elo on GDPval-AA v2.1 versus 1735 for Fable 5.1 and 1708 for Opus 5. It also posts 40.0% on AutomationBench, 67.7% with tools on Humanity's Last Exam, 81.8% partial on OSWorld 2.1 and 89.0% with tools on Chartography.
There are caveats to how those numbers were produced. Anthropic notes that Opus 5.5 was evaluated with its production safeguards enabled: when those intervened, cybersecurity tasks were completed by Claude Opus 4.8 and biology and frontier LLM development tasks by Claude Opus 5, which likely reduces Opus 5.5's performance on those benchmarks. AutomationBench, in turn, was run and reported by Zapier without fallback models, so safeguard interventions counted as failures — producing a lower score than the model would achieve in practice.
Anthropic cautions that at this level of capability, benchmark margins have become a less reliable guide to real-world differences, and that in its own use the gap between Opus 5.5 and Fable 5.1 is narrower than the scores suggest. Where its advantage is clear is efficiency: at its default effort level on FrontierCode, Opus 5.5 beats GPT-6 Astra at roughly a fifth of the cost per task; on Terminal-Bench 4.0 it matches Astra for about 40% of the cost; and on CursorBench it beats GPT-5.6 Sol by 11 points for about a third of the cost.
A few concrete workloads are more persuasive. One early tester completed a 680,000-line code migration in less than a day — work that would have taken an engineering team weeks. Opus 5.5 audited and fixed a 200,000-line codebase in under three hours, where Opus 5 took over 20 hours and used 2.5x as many tokens. In an internal test translating HAProxy from C to Rust, both rewrites passed nearly all of HAProxy's own regression tests, but Opus 5.5 finished in 9.5 hours against Fable 5.1's 12, at 51% less cost.
The knowledge-work results come with checkable detail too. Anthropic asked Opus 5.5, Fable 5.1 and Opus 5 to each write a report on a company's quarterly performance using only a copy of the web where the earnings release was hard to locate; an automated grader then checked every figure and quote against sources. Sixteen of 18 Opus 5.5 reports cleared the quality bar, where any invented figure or quote would have failed, while neither Fable 5.1 nor Opus 5 cleared it in any attempt. In another test, both models analysed a fictional merger; Opus 5.5 finished in 63 minutes at 50% lower cost with a more thorough model and an easier-to-read presentation, while Opus 5's version contained minor errors.
Writing and communication are something Anthropic targeted specifically this round, having heard the most feedback there after Opus 5. The company says Opus 5.5's messages are easier to take in at a glance, put the most important information up front, use less jargon and fewer idiosyncratic phrases, and follow whatever writing rules you give it; one early tester said it "writes the way I do." Anthropic adds that in its own use the output is easier to follow and check — which it treats as a safety benefit as well as a practical one. Customer reports circle the same theme: fewer tokens, fewer steps. GitHub said that across GitHub Copilot CLI and VS Code, Opus 5.5 used among the fewest tokens and steps it has measured, solving more terminal tasks than Opus 5 in less than half the steps in VS Code. Optiver said that on real engineering and trading-desk work, Opus 5.5 matched Opus 5's quality in about half the turns, time and output tokens, cutting that workload's cost by 40% to 50%. A developer at Clio ran it unattended for over 18 hours on a task spanning six repositories.
Safety and how it ships
On safety, Anthropic says Opus 5.5 posted the best scores of any model to date on its automated behavioural audit — its largest alignment suite, covering thousands of simulated scenarios — is much less likely than recent models to take hard-to-reverse actions or act outside the boundaries it has been given, and resists prompt injection better than Opus 5. The company also broadened its alignment testing to cover longer tasks, impossible tasks and scenarios modelled on real incidents. On a benchmark run by the AI security firm Gray Swan, it ties Fable 5.1 for the lowest prompt injection success rate of any model tested. Per the brief, it is deployed with ASL-3 safeguards, and shows roughly 85% fewer instances of boundary evasion than Opus 5 and Mythos 5.1.
For enterprises running agents inside their own systems, Anthropic pairs the model with a stack: a classifier that screens every action before it runs, an open-source sandbox security teams can audit, and code review that catches vulnerabilities before they merge. That matters because agents running autonomously for many hours need to be doing what they were meant to do. And because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, it ships with safeguards similar to those on Fable 5.1, with vetted organisations able to apply to the relevant life-sciences and cyber verification programmes for research use.
Read down the family and the strategy is plain: Claude Fable 5.1 and Mythos 5.1 (the latter limited to a trusted access programme) staked out the frontier on September 1; Opus 5.5 made the flagship faster and cheaper on September 22; Sonnet 5.5 arrived on September 28 more than 30% faster than Sonnet 5 and up to 30% cheaper; and Haiku 5.5 on October 7 cut running costs by roughly 75% against Haiku 4.5. A flagship that is both faster and cheaper, plus a filled-out tier line, is what lets token-hungry work like long-horizon coding and agents run for longer — that is the real play this round.