Anthropic Releases Claude Sonnet 5.5 with Major Agentic Coding and Cost Efficiency Gains

Anthropic has released Claude Sonnet 5.5, the second model in the Claude 5.5 series, positioning it as a faster and more cost-effective complement to Opus 5.5. While Opus 5.5 remains superior for complex, open-ended judgment tasks, Sonnet 5.5 is optimized for well-defined daily work, bug fixes, and document generation. A high-volume, cost-sensitive variant, Claude Haiku 5.5, is expected to launch within weeks.

Performance Benchmarks

Sonnet 5.5 demonstrates significant improvements over Sonnet 5 across key benchmarks, often approaching Opus 5.5 performance at a lower cost:

• Terminal-Bench 4.0 (Agentic Coding): Scored 70.6%, a substantial increase from Sonnet 5's 10.3% (Opus 5.5 scored 66.4%).
• GDPval-AA v2.1 (Knowledge Work): Achieved 1844 points, nearly matching Opus 5.5's 1846 points and surpassing Sonnet 5's 1449 points.
• OSWorld 2.1 (Computer Use): Reached 80.1%, up from Sonnet 5's 57.0% (Opus 5.5: 81.8%).
• Chartography (Visual Chart Recognition): Scored 61.6%, a dramatic rise from Sonnet 5's 15.6%.
• FrontierCode 1.1: Recorded 52.1% at Xhigh effort, compared to Sonnet 5's 42.4%.
• Pokémon Red: Became the first Sonnet model to clear the game using only screenshots.

Cost and Speed Efficiency

Pricing remains identical to Sonnet 5 ($2/1M input tokens, $10/1M output tokens), but efficiency gains reduce task costs by up to 30% due to lower token usage. Output generation is 30%+ faster than Sonnet 5, making it the fastest Sonnet model to date. At Low/Medium effort settings, it achieves Sonnet 5's peak scores at approximately 1/10th the cost per task.

Safety and Guardrails

Sonnet 5.5 introduces enhanced safety measures, particularly in cybersecurity and reasoning extraction:

• Cybersecurity: Features Opus 5-level capabilities with guardrails similar to Opus 5.5. High-risk cyber tasks fall back to Sonnet 5, while advanced access is available via the Cyber Verification Program.
• Distillation Protection: It is the first Sonnet model with safety classifiers to prevent reasoning extraction, ensuring thinking processes remain separated from accounts.
• Biology: Maintains Sonnet 5-level guardrails, blocking only high-risk requests while allowing most research and clinical work.

Enterprise Feedback

Early testers reported significant operational improvements:

• Early testers: Achieved Opus 5-equivalent scores in 118 app builds with fewer iterations (3.6 vs. 7.7 for Opus 5).
• Early testers: Ticket processing speed improved by 20% with fewer incorrect decisions.
• Balyasny Asset Management: Used ~121k tokens per answer vs. 497k for Sonnet 5 in financial tasks.
• Early testers: Reported 2.4x faster speeds and a 12% reduction in total token usage.

The model is available on AWS, Google Cloud, and Microsoft Azure under the model name claude-sonnet-5-5, with Zero Data Retention options supported.