Claude Sonnet 5.5
Claude Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 versus Sonnet 5's 10.3%, runs 30% faster, and keeps the $2/$10 price.
The second Claude 5.5 model, faster and up to 30% cheaper per task than Sonnet 5 at unchanged $2/$10. Two points below Opus 5.5 on GDPval-AA, and the first Sonnet to beat Pokemon Red from screenshots. The first Sonnet shipped with Opus-style cyber safeguards and fallbacks, and with preserved-thinking limits against distillation.
- Date
- Monday, 28 September 2026
- Lab
- Anthropic
- Kind
- model
- Access
- closed API
- Price
- $2/$10 per M tokens in/out; cache reads $0.20 (2026-09-28)
Figures
| Measure | Value | Measured by |
|---|---|---|
| Terminal-Bench 4.0 | 70.6% vs 10.3% Sonnet 5 and 66.4% Opus 5.5 (Anthropic run) | company |
| GDPval-AA v2.1 | 1844 Elo vs 1449 Sonnet 5, 1846 Opus 5.5, 1487 GPT-5.6 Sol | third-party |
| CursorBench 4.0 | 55.5% vs 34.1% Sonnet 5, 57.8% Opus 5.5 | third-party |
| OSWorld 2.1 (partial) | 80.1% vs 57.0% Sonnet 5, 81.8% Opus 5.5 | company |
| FrontierCode 1.1 (Main, Max effort) | 46.2% vs 42.4% Sonnet 5, 54.4% Opus 5.5, 49.3% GPT-6 Sol | company |
Anthropic says Opus 5.5 remains clearly stronger on open-ended work needing sustained judgment. Terminal-Bench 4.0 gap vs Sonnet 5 looks very large; Anthropic lists it as measured by itself and notes it did not have a public GPT-6 Sol figure so it shows GPT-5.6 Sol there. Haiku 5.5 was still unreleased on 2026-10-04: Haiku 4.5 remains the current Haiku. Five breaking API changes from Sonnet 5 (forced tool use errors, account-bound thinking blocks, others).
Sources
- www.anthropic.com/claude-sonnet-5-5
- platform.claude.com/docs/en/models/sonnet-5-5/overview
- platform.claude.com/docs/en/release-notes/overview
This record was checked against its sources on 6 October 2026. How we check