AitopkitAnthropic has released Claude Opus 5.5, which it describes as its most powerful model to date. The headline number: it matches the flagship Claude Fable 5.1 on most tasks while costing roughly 40% less to run than the previous Opus 5 generation — $4 per million input tokens and $20 per million output tokens.
Alongside the price cut, Anthropic claims the model is 30%+ faster, writes more naturally, and — in a notable safety claim — is 85% less likely than Opus 5 to attempt bypassing its own safeguards. That safety framing matters: the launch lands days after Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman publicly debated slowing down AI development, with both sides now racing to show they can ship faster and safer at once.
The timing is also commercial. OpenAI answered within hours with GPT-6 Sol and GPT-6 Luna — cheaper, faster versions of GPT-6 Astra — and both companies are fighting for the same enterprise and developer workloads ahead of Anthropic's widely expected IPO. For builders, the practical takeaway is simple: frontier-class performance keeps getting cheaper, so re-benchmark your costs before locking into a provider.
One caveat, as always with launch-day claims: the performance and safety figures are Anthropic's own, reported at release. Independent benchmarks and real-world usage over the coming weeks will show how much of the gap holds up in production.