Anthropic has released Claude Sonnet 5, its most capable Sonnet model yet, narrowing the performance gap with the company’s more expensive Opus models across coding, reasoning and autonomous tool use.

The model can plan complex tasks, operate browsers and terminals, and continue working through multi step assignments with less user intervention than Claude Sonnet 4.6. 

Anthropic said its performance is close to Claude Opus 4.8 in several agentic workloads, though Opus remains the stronger option when users prioritize accuracy over cost.

Sonnet 5 showed improvements over its predecessor on BrowseComp, an evaluation of agentic web research, and OSWorld Verified, which measures a model’s ability to complete tasks across computer interfaces. Anthropic said developers can adjust the model’s effort level to balance accuracy, speed and computing costs.

The model supports a one million token context window and up to 128,000 output tokens through the standard API. It is available under the API identifier claude-sonnet-5 through Anthropic’s platform, Amazon Bedrock, Google Cloud and Microsoft Foundry.

Claude Sonnet 5 is now the default model for Claude’s Free and Pro plans. It is also available to Max, Team and Enterprise customers, as well as through Claude Code and Claude Cowork.

Anthropic introduced temporary API pricing of $2 per million input tokens and $10 per million output tokens through August 31. Pricing will rise to $3 for input and $15 for output beginning September 1, matching the previous standard pricing for Sonnet models. Claude Opus 4.8 costs $5 per million input tokens and $25 per million output tokens.

The company cautioned that Sonnet 5 uses an updated tokenizer that may convert the same input into as many as 35% more tokens depending on the content. Anthropic said the introductory pricing was designed to make the initial transition roughly cost neutral for developers.

Anthropic also reported lower rates of hallucination, sycophancy and other undesirable behavior compared with Sonnet 4.6. 

The company said the model was better at rejecting malicious requests and resisting prompt injection attempts, though it recorded somewhat more undesirable behavior than Opus 4.8 and Mythos Preview in its automated audit.

Sonnet 5 remains less capable than Anthropic’s Opus and Mythos models on advanced cybersecurity tasks. It failed to produce a complete working exploit in an evaluation involving patched Firefox vulnerabilities, though it showed slightly more partial progress than Sonnet 4.6. Anthropic launched the model with automated cyber safeguards enabled by default.