Anthropic released Claude Opus 5.5 on September 22, 2026, just one week after CEO Dario Amodei published an essay stating, "We must slow the pace at which we improve the capabilities of AI models." The timing, landing only three weeks after Fable 5.1, drew immediate skepticism. However, for developers, the model represents a significant correction to Opus 5’s bloat, offering a 20% cost reduction and a 30% speed increase in response times.

Pricing and Performance Metrics

Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 respectively. Cache read costs dropped significantly from $0.50 to $0.20 per million tokens. In coding benchmarks, Opus 5.5 scored 66.4% on Terminal-Bench 4.0, outperforming both Opus 5 (52.3%) and the more expensive Fable 5.1 (55.8%). While it approaches Fable’s performance, it does so at less than half the cost. However, users must avoid setting thinking effort to max, as independent tests showed costs leveling out at roughly $6 per task due to increased token generation.

Breaking Changes and Migration Risks

Switching from Opus 5 is not a simple plug-and-play operation. Several API parameters now return 400 errors, including disabling thinking (thinking: {"type": "disabled"}) and forcing specific tool use (tool_choice set to any). The computer_20251124 tool is rejected on the Claude API and Google Cloud, requiring migration to computer_toolset_20260801. Additionally, progress messages between tool calls may appear as empty thinking blocks, potentially making streaming UIs look frozen without throwing explicit errors.

Safety Filters and Hidden Reroutes

A critical operational detail is the new safety routing. Security tasks, such as exploit finding, are automatically rerouted to the older Opus 4.8, while flagged biology requests go to Opus 5. These filters scan file contents and search results, meaning a repository’s existing code can trigger a reroute unexpectedly. Furthermore, Sonar analysis indicates that while Opus 5.5 writes 27% less code with fewer overall issues, bugs per line increased by 12%, and concurrency-related problems jumped by 44%.

Key Takeaways

  • Cost Efficiency: Input/output prices dropped to $4/$20 per million tokens, with cache reads at $0.20.
  • Performance: Scored 66.4% on Terminal-Bench 4.0, beating Fable 5.1’s 55.8% at 60% of the cost.
  • Migration Friction: Four major API breaking changes require code updates, particularly around thinking blocks and tool choice.
  • Safety Reroutes: Security and biology tasks are silently redirected to older models (Opus 4.8/5) based on content analysis.

The Bottom Line

Opus 5.5 is what Opus 5 should have shipped as, but the silent safety reroutes and API breaking changes mean this is a migration, not an upgrade.