Frontier Now
Search中文

Model releases#Agents#Long context

Anthropic releases Claude Opus 5.5: near Fable 5.1 on most tasks, with API prices cut 20%

Anthropic released Opus 5.5, the first Claude 5.5 model, on September 23. The company says it matches Fable 5.1 on most tasks and costs 40% less per typical task than Opus 5, at $4 input and $20 output per million tokens.

Author
Frontier Now Editorial
Published
Event date
Anthropic’s launch artwork for Claude Opus 5.5: a sunrise over the horizon with the model name

Anthropic released Claude Opus 5.5 on September 23, the first model in the Claude 5.5 family, with Sonnet 5.5 and Haiku 5.5 due in the coming weeks. It’s positioned as the workhorse for long-running coding agents and knowledge work.

The facts

  • Pricing: $4 per million input tokens and $20 per million output, 20% below Opus 5. Cache reads drop from $0.50 to $0.20. A Fast mode of up to 2.5x speed costs $8 / $40.
  • Specs: 1M-token context, 128K max output, thinking always on with medium effort by default. The API name is claude-opus-5-5. It’s live in Claude, Claude Code and the Claude Platform, on AWS, Google Cloud and Azure, and is now the default model on paid plans.
  • Vendor benchmarks: 66.4% on Terminal-Bench 4.0 (Opus 5: 52.3%) and 81.8% on OSWorld 2.0 (up from 74.0%). It trails in places too: 40.0% on AutomationBench versus GPT-6 Astra’s 41.4%.
  • Plan limits: five-hour usage limits on Pro, Max and Team go up 20%.
  • Safety: Anthropic says boundary-pushing attempts are down about 85% from Opus 5. Most cybersecurity tasks are routed to Opus 4.8, with full capability reserved for verified researchers.

Our take

The story here is price and long tasks more than the leaderboard. In early testing, a 200,000-line codebase audit took under three hours, where Opus 5 needed more than 20 for a similar job. If you’d written Opus off as too expensive for daily use, it’s worth redoing the math.

One caution: the lower per-task cost is measured at default settings on typical workloads. Developers citing Artificial Analysis runs note its output-token use at max effort is the highest of any model tested, so trial a small batch before running high-effort jobs at scale.