Anthropic has unveiled Claude Sonnet 5.5, a more efficient and cost-effective iteration of its AI model that delivers significant gains in coding test performance and introduces frontier-level cyber defenses previously limited to its top-tier models.
- Scores 70.6% on Terminal-Bench 4.0, up from 10.3% for Sonnet 5
- Runs 30% faster and costs up to 30% less per task
- First Sonnet with frontier-style cyber safeguards and reasoning extraction blockers
What happened
Anthropic has released Claude Sonnet 5.5, an upgraded version of its Sonnet AI model line that delivers a dramatic improvement in coding performance. Specifically, Sonnet 5.5 scores 70.6% on the Terminal-Bench 4.0, a challenging agentic coding test, compared with only 10.3% for the earlier Sonnet 5 model. This new release also runs more than 30% faster and is priced up to 30% cheaper per task, at $2 per million input tokens and $10 per million output tokens.
In addition to efficiency gains, Sonnet 5.5 introduces enhanced security features. It is the first Sonnet model to include frontier-style cybersecurity safeguards such as classifiers that block reasoning extraction, which prevents the model’s internal logic from being decoded and misused. Higher-risk cybersecurity requests will automatically fallback to the older Sonnet 5 model to maintain safety. The model’s biology-related safeguards remain unchanged from previous versions.
Why it matters
Claude Sonnet 5.5’s improvements represent a significant leap in both performance and security for Anthropic’s AI offerings. Its high score on the Terminal-Bench 4.0 test demonstrates its superior coding capabilities compared to prior models and competitors while maintaining competitive pricing. This balance of power, speed, and cost could make it an attractive solution for developers and enterprises seeking advanced AI coding assistants.
The inclusion of frontier-level cyber safeguards is especially compelling given increasing regulatory pressures, such as the EU AI Act, which mandates robust risk management and cybersecurity protections for general-purpose AI providers. By binding model reasoning outputs to the user account and blocking extraction attempts, Anthropic is taking proactive steps to reduce misuse risks and comply with evolving legal frameworks, helping to set industry standards for responsible AI deployment.
What to watch next
It will be important to monitor how customers adopt Sonnet 5.5 in real-world applications and whether the enhanced cybersecurity features effectively prevent unauthorized reasoning extraction without degrading performance. Anthropic’s approach of selectively falling back to older models for higher-risk requests may also be scrutinized for scalability and user experience impacts.
Additionally, developments around AI regulation enforcement—particularly the application of the EU AI Act’s incident reporting and adversarial testing requirements—will influence how companies like Anthropic evolve their models. Further advancements in open-ended AI tasks, where competitors’ Opus models currently excel, may also drive future iterations of Sonnet 5.5 or new Claude models focused on sustained, nuanced judgment.