Anthropic's Claude Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, cuts task cost up to 30%

Sonnet 5.5 arrives as the second release in Anthropic's Claude 5.5 family, following the top-tier Opus 5.5. The company positions the pair by task type: Opus 5.5 for complex work needing careful judgment, and Sonnet 5.5 for clearly scoped everyday tasks, bug fixes, and polished documents, slides and spreadsheets. Sonnet 5.5 also brings a sharp eye for design and, per Anthropic, generates clearer prose than the previous generation.

The benchmark spread is wide. On Terminal-Bench 4.0, an agentic coding evaluation, Sonnet 5.5 scored 70.6% against Sonnet 5's 10.3%, while on GDPval-AA it landed 2 points below Opus 5.5. It is said to be close to Opus 5.5 in computer use and chart recognition, and well above Sonnet 5 and GPT-6 Sol on long-horizon knowledge work. A single screenshot alone was enough for it to play through Pokémon Red. In some evaluations at Max effort it matches Opus 5.5, yet both internal and outside testing reportedly favor Opus 5.5 on complex, open-ended work. At Low and Medium effort levels it keeps per-task cost to roughly one-tenth while beating Sonnet 5's top score.

The alignment and safety picture is deliberately nuanced. Sonnet 5.5 holds or exceeds Sonnet 5 on most alignment metrics in Anthropic's automated behavior audits. Its cybersecurity capability improved markedly, reaching a level similar to Opus 5. It is the first Sonnet model to carry the same cyber safeguards and fallback functions as top-capability models, applied like Opus 5.5's safeguards, with biological safeguards unchanged from Sonnet 5. Both safeguards target only rare high-risk requests, so most everyday software development and life-sciences work should be unaffected. The broader question of whether same-price models can absorb the highest-judgment tasks will hinge on how the reported Opus 5.5 edge holds up once more outside testers use Sonnet 5.5 in production.