Reviews · SEPTEMBER 25, 2026
Claude Opus 5.5 lands at $4/$20 per million tokens, 40% under Opus 5
Anthropic's September 22 release matches Fable 5.1 on most work while cutting per-token prices 20% and total run cost 40%, with a 30%+ speed gain and cache reads down to $0.20 per million.
Anthropic released Claude Opus 5.5 on September 22 at $4 per million input tokens and $20 per million output tokens, 20% below Opus 5 per token and roughly 40% cheaper on typical workloads once cache reads (down 60%, to $0.20 per million) and a 30%-plus speed improvement are folded in. It's the first frontier launch since CEO Dario Amodei embraced what Anthropic calls "pacing the frontier," and the release reads less like a capability push than a pricing one.
The benchmark story is deliberately understated. Opus 5.5 posts 66.4% (±2.6) on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1, and 57.8% on CursorBench 4.0, against the recently launched Fable 5.1's 55.8%, 50.3%, and 51.8%. On Zapier's AutomationBench, Opus 5.5 hits 40.0% versus Opus 5's 26.9%, the one place the jump looks generational rather than incremental. Anthropic itself waves the reader off the scoreboard: "In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest," the company writes, adding that "at these levels of capability we've found that benchmark margins have become a less reliable guide to real-world differences."
That's a striking admission from a frontier lab weeks before an IPO. Bloomberg framed the release against Anthropic's public-market preparations and the same-day launch of OpenAI's GPT-6 Sol and Luna, with xAI's Grok 4.7 already in the pricing conversation. When the top of the S-curve flattens, the competition migrates to cost per unit of work.
The internal task numbers matter more than the leaderboards. Opus 5.5 completed an HAProxy C-to-Rust translation in 9.5 hours against Fable 5.1's 12, at 51% lower cost. An early tester audited a 200,000-line codebase in under three hours; Opus 5 needed over 20 and burned 2.5 times the tokens. Reuters reports Opus 5.5 outscored GPT-5.6 Sol on a software development benchmark at roughly one-third the cost.
Safety framing has shifted too. Anthropic says Frontier Design and METR ran pre-release evaluations; Reuters reports the model is about 85% less likely than Opus 5 or Mythos 5.1 to attempt to bypass containment boundaries. It's classified with CB-1 but not CB-2 biological capabilities.
Forthcoming Sonnet 5.5 and Haiku 5.5 releases will test whether the same curve holds down-market. For now, the frontier's most important axis is denominated in dollars.