Anthropic’s Claude Sonnet 5.5 Outperforms Opus in Agentic Coding While Cutting Token Costs

Anthropic has replaced its default chatbot engine with Claude Sonnet 5.5, introducing a mid-tier artificial intelligence model that matches or exceeds the performance of the company’s most expensive systems in software development tasks while simultaneously reducing the computational resources required to run it.

The release arrives days after Anthropic debuted Opus 5.5, its flagship model designed for maximum reasoning depth. Historically, the San Francisco-based AI developer has maintained strict capability boundaries between its tiers — Haiku for speed and low cost, Sonnet for balanced daily use, and Opus for complex research and heavy enterprise workloads. Sonnet 5.5 disrupts that hierarchy by posting benchmark scores in autonomous programming that surpass Opus 5.5 running at its highest compute setting.

On Terminal-Bench 4.0, an evaluation framework that tests an AI system’s ability to independently navigate command-line environments, debug errors, and execute multi-step coding workflows without human intervention, Sonnet 5.5 achieved a 70.6% success rate. By comparison, the preceding Sonnet 5 model managed only 10.8% on the same assessment. Opus 5.5, configured at its maximum “xhigh” reasoning allocation, scored 66.4%.

agentic coding benchmarks have become the primary battleground for AI companies targeting enterprise software developers. Unlike standard code generation, which typically involves completing isolated functions or answering syntax questions, agentic tasks require models to plan, use tools, read file systems, and self-correct over extended sequences. The nearly sevenfold improvement from Sonnet 5 to Sonnet 5.5 suggests architectural changes in how the model handles tool use and context retention, rather than simple parameter scaling. Other standard programming evaluations show more modest gains, indicating the leap is highly specific to autonomous, multi-turn execution rather than raw code synthesis.

Beyond software engineering, Anthropic reports measurable improvements in computer use — the model’s ability to visually parse and interact with graphical user interfaces — alongside better recognition of structured data within visual tables and stronger multidisciplinary reasoning capabilities.

These performance gains are paired with structural efficiency improvements that directly impact operating costs for developers integrating the model via API. Sonnet 5.5 completes most standard queries using up to 30% fewer tokens than its predecessor, while also generating responses up to 30% faster. In large-scale enterprise deployments processing millions of requests daily, this reduction in token consumption translates into significant infrastructure savings without requiring changes to existing API integrations.

Anthropic held pricing steady despite the capability jump. API access remains at $2 per million input tokens and $10 per million output tokens. Prompt caching, a feature critical for applications that repeatedly send similar system instructions or context blocks, retains the established rates: 20 cents per million tokens for cache reads and $2.50 per million tokens for five-minute cache writes. This pricing structure keeps Sonnet positioned below Opus but above the cheaper Haiku tier, maintaining Anthropic’s three-tier monetization strategy even as the technical gap between Sonnet and Opus narrows.

Sonnet 5.5 now serves as the default engine across the consumer-facing Claude application. Users accessing the service through the desktop client may need to install the latest software update before the new model activates automatically.