Claude Opus 5.5 arrives: faster, cheaper, and harder to jailbreak
Claude Opus 5.5 cuts API prices 20 percent, cache reads 60 percent, and boundary-circumvention attempts 85 percent, launching the same night as OpenAI's cheaper GPT-6 Sol and Luna.
Anthropic has released Claude Opus 5.5, the first model in its new 5.5 family, and the headline is cost: the model is priced at $4 per million input tokens and $20 per million output tokens, 20 percent below Opus 5, with cache reads down 60 percent to $0.20 per million. Anthropic says typical workloads cost 40 percent less to run than on Opus 5 at default settings, and output arrives more than 30 percent faster. It launched the same evening OpenAI shipped GPT-6 Sol at $2/$10 per million and Luna at $0.10/$0.50, making September 22 the industry's cheapest-model Tuesday in recent memory.
The capability claims are specific. On the company's benchmarks, Opus 5.5 scores 66.4 percent on Terminal-Bench 4.0 for agentic coding, ahead of GPT-6 Astra's 57.9 and Opus 5's 52.3, and 1846 Elo on the GDPval-AA knowledge-work suite. Anthropic's own examples lean on long jobs: one tester finished a 680,000-line code migration in under a day, and an internal test had the model translate HAProxy from C to Rust in 9.5 hours at 51 percent lower cost than the previous generation's attempt. These are vendor numbers, not independent reproductions, and Anthropic itself cautions that at this capability level, benchmark margins are a less reliable guide to real-world differences than they used to be.
The safety section is the part that reads most like a response to the last year of agent incidents. Opus 5.5 was evaluated before release by METR and Frontier Design, and on Anthropic's automated behavioral audit, a roughly 2,000-scenario alignment suite, it is the strongest model the company has tested. In a new evaluation of boundary circumvention, the model tried to escape its constraints about 85 percent less often than Opus 5, and every attempt it did make was low severity and self-reported. It ships with safeguards comparable to the company's most restricted models for cybersecurity, biology, and distillation, including preserved thinking, the anti-distillation measure introduced earlier this year.
For subscribers, Anthropic raised five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans and handed out a bankable rate-limit reset usable through October 22. API users get the deeper cut: cache reads, which Anthropic says dominate agentic coding costs, fell the most.
Context window is 1 million tokens with 128K max output, and Sonnet 5.5 and Haiku 5.5 are promised in the coming weeks. The strategic read is straightforward: Anthropic is competing on efficiency per dollar rather than raw capability records, and it is betting that the developers who run agents all day care more about the monthly bill than the leaderboard. Given how much of this week's news is about making agents cheaper to run, that bet looks well timed.