Anthropic ships Claude Opus 5.5, the first Opus since ‘pace the frontier,’ at 40% less to run
Sep 22: Fable 5.1-class work at $4/$20 per million tokens, 20% below Opus 5; cache reads $0.20. Terminal-Bench 4.0 66.4%. The Register: labs still racing. Artificial Analysis, via Claburn, puts it at 58 on the Intelligence Index.
Deep dive3 minIn-depth
James Okonkwo3 min read
- Claude
- Opus 5.5
- Anthropic
- Pricing
- Safety

Photo: AnthropicSource page
Anthropic published Claude Opus 5.5 on Tuesday, the first model in a Claude 5.5 family and the first Opus since Dario Amodei told the industry to pace the frontier on 12 September. The company says it performs at Claude Fable 5.1’s level on most work and costs about 40 percent less to run than Opus 5 on typical workloads. List price is $4 per million input tokens and $20 per million output — 20 percent below Opus 5. Cache reads, which Anthropic says are most of the bill on agent and coding jobs, fall 60 percent to $0.20. Output is more than 30 percent faster than Opus 5. A Fast mode in Claude Code and the Claude Platform is $8/$40, up to 2.5 times the speed. The API id is `claude-opus-5-5`. It is live on the Claude Platform, AWS, Google Cloud and Azure. Sonnet 5.5 and Haiku 5.5 are due in the coming weeks.
Thomas Claburn at The Register, filing at 01:02 UTC Wednesday, put the ship next to the July staff letter and Amodei’s essay. Ten days after the CEO’s post, Anthropic called Opus 5.5 “the strongest-performing model we’ve tested to date.” Claburn counted the cadence: Fable 5.1 and Mythos 5.1 21 days earlier; Opus 5 39 days before those. Anthropic’s releases have gone from roughly quarterly in 2025 to almost monthly this year. Duncan Riley at SiliconANGLE had the same Tuesday pairing: Opus 5.5, then OpenAI’s cheaper GPT-6 Sol and Luna minutes later. Artificial Analysis, Claburn wrote, puts Opus 5.5 on par with GPT-6 Astra for agentic work and at 58 on the Intelligence Index, first among the ten evaluations on that board.
Our tests show that at default settings it will cost 40 percent less than Opus 5 on typical workloads.
— Anthropic, Introducing Claude Opus 5.5, Sep 22, 2026
A vendor table, and a footnote that cuts the cyber rows
Anthropic’s published grid, at adaptive thinking and max effort unless noted, has Terminal-Bench 4.0 at 66.4 percent against Fable 5.1’s 55.8, Opus 5’s 52.3 and GPT-6 Astra’s 57.9 — the last as reported by OpenAI; Opus 5.5 at xhigh. Standard error on that row is ±2.6 points. FrontierCode v1.1 Main is 54.4 percent. CursorBench 4.0 is 57.8. GDPval-AA v2.1 is 1846 Elo against Fable’s 1735 and Opus 5’s 1708. AutomationBench, run and reported by Zapier without fallback models, is 40.0 percent; GPT-6 Astra is 41.4 on Zapier’s public board. Humanity’s Last Exam with tools is 67.7 percent. Terminal-Bench-Science 0.1 is 58.7 against Opus 5’s 29.0. Anthropic’s own footnote is the one that matters for the cyber and bio rows: production safeguards were on. When they fired, cybersecurity tasks went to Opus 4.8 and biology and frontier-LLM-development tasks to Opus 5. That, the company says, likely reduces the scores. Zapier treated safeguard interventions as failures.
The customer quotes are the usual launch packet. GitHub’s Mario Rodriguez: in VS Code it solved more terminal tasks than Opus 5 in less than half the steps. Box’s Yashodha Bhavnani: 40 percent less verbose without losing accuracy. One unnamed tester finished a 680,000-line migration in under a day. Anthropic’s own HAProxy C-to-Rust rewrite: both Opus 5.5 and Fable 5.1 passed nearly all of HAProxy’s tests; 5.5 finished in 9.5 hours against 12 and cost 51 percent less. The company also says the gap to Fable 5.1 in daily use is narrower than the table.
Safeguards that look like Fable, and a race Claburn will not pretend stopped
On Anthropic’s automated behavioral audit — nearly 2,000 scenarios — Opus 5.5 is the strongest model the lab has tested. Attempts to circumvent containment boundaries dropped about 85 percent against Opus 5 or Mythos 5.1, the company says, and every attempt was low severity and self-reported. Prompt-injection resistance matches Fable 5.1 on Gray Swan. Because the model is comparable to Mythos 5.1 on biology and cyber, it ships with Fable-class fences: most cybersecurity work routes to Opus 4.8; high-risk biology is gated, with a Life Sciences Verification Program for vetted labs. Frontier Design and METR tested it before release. Preserved thinking, the anti-distillation lock from Fable 5.1, applies to API accounts created on or after 31 August. Thinking mode cannot be switched off. The essay from 12 September is already on this site. Tuesday’s fact is a priced Opus, an AA 58, and Claburn’s observation that the race did not pause.


