Claude Sonnet 5.5 Launches at Half Opus 5.5 Price, Nears Its Performance

Performance

Anthropic announced Claude Sonnet 5.5, the second model in the Claude 5.5 family. Compared to the previous Sonnet 5, version 5.5 delivers a 30%+ speed increase and up to 30% cost reduction for most workloads. Sonnet 5.5 complements Opus 5.5: Opus targets complex work requiring careful judgment, while Sonnet 5.5 excels at well-scoped daily tasks, code fixes, and creating polished documents, slides, and spreadsheets, also demonstrating strong design sensibility.

Sonnet 5.5 outperforms Sonnet 5 across all domains, with particularly large gains in programming. On the agentic coding benchmark Terminal-Bench 4.0, Sonnet 5.5 scores 70.6% versus 10.3% for Sonnet 5. On FrontierCode with High effort, Sonnet 5.5 exceeds Sonnet 5 by 10 points while per-task cost is roughly one-fifteenth. On CursorBench (real Cursor coding sessions), Sonnet 5.5's best result is within 2 points of Opus 5.5.

On GDPval-AA, a test covering real work tasks across 9 industries and 44 occupations, Sonnet 5.5 trails Opus 5.5 by only 2 points but beats Sonnet 5 by about 400 points. It approaches Opus 5.5 on computer operation and chart reading, and clearly outperforms Sonnet 5 and GPT-6 Sol on long-horizon knowledge work. Sonnet 5.5 is also the first Sonnet model to complete Pokémon Red using only screen captures.

However, benchmarks capture only one facet. In internal and external testing on complex, open-ended work requiring sustained judgment, Opus 5.5 remains noticeably stronger.

Anthropic published charts showing scores and per-task cost at different effort settings. Data points nearer the top-left indicate more capability per dollar. At Low or Medium effort, Sonnet 5.5 surpasses Sonnet 5's best score at roughly one-tenth the per-task cost. At lower effort settings Sonnet 5.5's cost advantage is greatest and it complements Opus 5.5 best; at higher settings it achieves comparable performance at similar cost.

Cost and Speed

Sonnet 5.5's input, output, and cache-write prices are all half of Opus 5.5's, while cache-read remains $0.20 per million tokens, making it well-suited as a default model for high-volume agent workloads.

Sonnet 5.5 uses fewer tokens per task than Sonnet 5, further lowering run cost. Output generation speed is also over 30% faster, a tangible efficiency gain.

Safety and Alignment

Because Sonnet 5.5 does not push the capability frontier further, Anthropic focused alignment evaluation on targeted risks applicable at any capability level: actions harming users, misleading users, and assisting misuse with severe consequences.

Automated behavioral audits tested Claude across ~1,850 scenarios. On most alignment, misuse-resistance, and honesty metrics, Sonnet 5.5 matches or exceeds Sonnet 5.

In newer sandbox-containment evaluations, Sonnet 5.5's sandbox-escape attempt frequency approaches the best-performing model tested (Opus 5.5), and it is the least likely of all Anthropic models to probe container limits. Overall, Opus 5.5 still edges out slightly, but no evidence was found of Sonnet 5.5 pursuing goals conflicting with user intent. Anthropic acknowledges that no test suite reliably catches all failures, so undiscovered tendencies may exist; ongoing alignment work is paired with the mitigations described below.

Cybersecurity

Sonnet 5.5's cybersecurity capabilities improved substantially over Sonnet 5, prompting deployment of safeguards similar to Opus 5.5. It remains usable for routine software development (finding and fixing code issues), but high-risk cybersecurity tasks fall back to Sonnet 5, with the switch visible to users. Soon, cybersecurity defenders can apply to Anthropic's expanded Cybersecurity Verification Program for tiered access to higher capabilities in Sonnet 5.5, Opus 5.5, and Claude Mythos.

Biosafety

Sonnet 5.5 retains the same biosafety guards as Sonnet 5. These target harmful requests while leaving most research, education, and clinical work unaffected, though some microbiology and virology queries may be misflagged. Organizations can join the Life Sciences Verification Program for access with specialized safeguards supporting the full range of biology work.

Distillation Protection

Distillation attacks use thousands of fake accounts to extract model capabilities at industrial scale, enabling malicious actors to build powerful models without Anthropic's safety guards. Because Sonnet 5.5 is far more capable than its predecessor, it is the first Sonnet model to ship with a safety classifier that prevents reasoning extraction at launch. Sonnet 5.5 also expands the preserved-thinking mechanism, making Claude's reasoning inseparable from the account that generated it.

Claude Sonnet 5.5 is available now on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Developers can call it via the model identifier claude-sonnet-5-5 on the Claude Platform. Claude Haiku 5.5, aimed at high-volume, cost-sensitive applications, will join the family in the coming weeks.

Source:

https://www.anthropic.com/claude-sonnet-5-5