Claude Sonnet 5.5 Launches: 30% Faster at Same Sonnet Price

Anthropic released Claude Sonnet 5.5 today, completing the second tier of its Claude 5.5 model family. The release follows Claude Opus 5.5 from last week, and the headline numbers are striking: the same API pricing as Sonnet 5, output that runs 30% faster, and an agentic coding benchmark score that jumped from 10.3% to 70.6%.

That last number is the one that will get developers’ attention.

What Changed

The pricing stays the same: $2 per million input tokens, $10 per million output tokens, and $0.20 per million cache reads. What changes is how efficiently the model uses those tokens.

Anthropic describes Sonnet 5.5 as saving up to 30% per task not because of a price reduction, but because it uses fewer tokens and makes fewer tool calls to complete the same work. Early tester CodeRabbit put it plainly: the model shows better judgment about when to reach for web search and avoids the “high token use” pattern that made Sonnet 5 more expensive to run in practice.

On raw speed, Sonnet 5.5 generates output over 30% faster than its predecessor. For agentic pipelines where latency compounds across steps, that matters.

The Coding Numbers

Terminal-Bench 4.0 is designed to measure how well a model handles multi-step agentic coding tasks: navigating codebases, running tests, debugging in a live environment. Sonnet 5 scored 10.3%. Sonnet 5.5 scores 70.6%.

That is not a marginal improvement. It is a model that now belongs in a different category for coding work.

On FrontierCode v1.1, Sonnet 5.5 reaches 46.2% at maximum effort. On CursorBench 4.0, it scores 55.5%, second only to Opus 5.5. These are the kinds of benchmarks that translate into real developer productivity, not just leaderboard rankings.

Knowledge Work: Near-Opus at Sonnet Pricing

On GDPval-AA v2.1, which tests performance against real-world professional tasks across 44 occupations and nine industries, Sonnet 5.5 scores 1,844 Elo. Opus 5.5 scores 1,846. That two-point gap is effectively parity, and Opus 5.5 costs $4 per million input tokens to Sonnet 5.5’s $2.

For teams doing document-heavy work, analysis, and professional knowledge tasks, Sonnet 5.5 may have become the obvious default: near-flagship performance at the mid-tier price point.

What This Means for Business

The economics of running AI agents just shifted again.

When Anthropic released Opus 5.5 last week at a 40% cost reduction, the message was that flagship-level work was getting cheaper. Sonnet 5.5 makes a different argument: the middle tier has caught up to where the flagship used to be, and it runs faster and uses fewer tokens to get there.

For business owners and operations teams building AI agent workflows, Sonnet 5.5 removes a common tradeoff. You no longer have to choose between capable models that are expensive and cheaper models that require more babysitting. Sonnet 5.5 handles well-scoped everyday tasks, bug fixing, document drafting, and long-horizon agentic coding with near-Opus results at half the output token cost.

There is also a safety consideration worth noting for enterprise teams. Sonnet 5.5 is the first Sonnet model to carry Opus-level cybersecurity safeguards. That matters when you are giving agents real system access. Anthropic added higher-risk task fallback to Sonnet 5 and introduced new classifiers to prevent reasoning extraction attacks, which have been a growing concern in enterprise deployments.

Availability

Claude Sonnet 5.5 is available now via the Claude Platform using the model ID claude-sonnet-5-5, and through AWS, Google Cloud, and Microsoft Azure. Zero data retention is available for enterprise customers.

Claude Haiku 5.5, the third tier of the family, is still to come.

The combination of Opus 5.5 and Sonnet 5.5 now gives enterprise teams a full range of options: Opus for sustained, complex agentic work where cost is secondary to capability; Sonnet for the broad middle of professional tasks where speed, cost, and performance all need to balance; and Haiku (coming soon) for high-volume, cost-sensitive applications.

If your team has been waiting to commit to an AI agent stack, the model lineup is now more coherent and more affordable than it has been at any point this year.