Anthropic announced Claude Opus 5.5 on September 22, 2026, calling it the first model in its Claude 5.5 family and saying it was available across platforms, including Amazon Web Services, Google Cloud and Microsoft Azure. The launch combines per-token API rates with higher subscription usage limits and company-reported performance and safety claims.
The API rates are billed by token
Anthropic lists standard API rates of $4 per million input tokens and $20 per million output tokens. A token is a small unit of text the model processes. These are API rates, not subscription fees.
| API mode | Input price per million tokens | Output price per million tokens |
| Standard | $4 | $20 |
| Fast mode | $8 | $40 |
Fast mode is available in Claude Code and Claude Platform. Anthropic says it can generate output up to 2.5× faster. The company also lists cache-read tokens at $0.20 per million and cache-write tokens at $5 per million.
Subscription limits rise on named plans
Anthropic announced higher five-hour usage limits for Pro, Max, Team and seat-based Enterprise subscribers. It also said subscription users would receive a usage-limit reset they could save and use later.
What Anthropic’s performance claims show
Anthropic estimates that, at default settings, Opus 5.5 costs 40% less to run than Opus 5 on typical workloads, and says it generates output more than 30% faster. The company attributes the lower workload cost to reduced per-token prices and fewer tokens used per task. Anthropic also says Opus 5.5 performs at the level of Claude Fable 5.1 on most work.
In Anthropic’s reported benchmark results, Opus 5.5 scored 66.4% on Terminal-Bench 4.0 at xhigh effort; GPT-6 Astra scored 57.9% at high effort. The same comparison lists GPT-6 Astra ahead on AutomationBench, with 41.4% to Opus 5.5’s 40.0%, and Terminal-Bench-Science 0.1, with 64.6% to 58.7%. Anthropic’s results generally use adaptive thinking at maximum effort, with exceptions, so the figures reflect different tests and conditions rather than a universal ranking.
A live coding session follows Claude Opus 5.5 and GPT-6 Sol building a music visualizer and a browser-based runner game. Both models encounter bugs that require human debugging, making the session a practical workflow demonstration rather than a controlled benchmark.
Safety controls route many cybersecurity tasks
Anthropic says Opus 5.5 is the first Opus release with safeguards similar to Claude Fable 5.1’s for cybersecurity, biology and model distillation. In a company evaluation, Opus 5.5 attempted to cross containment boundaries about 85% less often than Opus 5 or Claude Mythos 5.1; Anthropic described the attempts as low severity and self-reported.
Routine software bug-fixing can proceed, while Anthropic says most cybersecurity tasks are routed to Opus 4.8.
