Anthropic has released Claude Sonnet 5 and made it the default model for Free and Pro users. Under the company’s standard API pricing, Sonnet 5 costs $3 per million input tokens and $15 per million output tokens. A promotional rate available until August 31 lowers that to $2/$10. Against Opus 4.8 at $5/$25, the new model comes in at roughly 60% less.
Company benchmarks place Sonnet 5 close to Opus 4.8
Anthropic says Sonnet 5 narrows the gap with its flagship model across several benchmark categories, though the figures are based on internal testing and have not been independently verified. On SWE-bench Pro, which targets agentic coding ability, Sonnet 5 scored 63.2%, up from Sonnet 4.6 at 58.1% and below Opus 4.8 at 69.2%.
On Terminal-Bench 2.1, Sonnet 5 posted 80.4% versus 82.7% for Opus 4.8. In Humanity’s Last Exam, Sonnet 5 reached 57.4% with tools enabled, nearly matching Opus 4.8 at 57.9%. On GDPval-AA v2, a benchmark focused on knowledge work, Sonnet 5 scored 1,618, slightly ahead of Opus 4.8 at 1,615.
Larger context window and stronger desktop task results
Anthropic also reported gains in computer-use tasks. On OSWorld-Verified, Sonnet 5 scored 81.2%, compared with 78.5% for the previous generation. That benchmark is designed around real desktop actions such as taking screenshots, dragging items, and moving data across applications.
Sonnet 5 supports a context window of up to 1 million tokens and a maximum output of 128k tokens. The configuration is aimed at workloads that require large document sets to remain available in a single session, including cross-file comparison, summarization, and decision support.
Lower list pricing may not mean a lower final bill
The model uses an updated tokenizer, and Anthropic said the same input can produce between 1.0x and 1.35x as many tokens depending on the content. The company described the pricing adjustment as broadly cost-neutral, but advised high-volume users to run their own benchmarks because actual billing may not fall in line with the lower headline rate.
On safety, Anthropic said Sonnet 5 shows lower hallucination and sycophancy rates than Sonnet 4.6 and is better at refusing malicious requests. Even so, the company’s own report indicates Sonnet 5 still has a higher rate of improper behavior than Opus 4.8 and the more tightly restricted Claude Mythos Preview.
In a Firefox 147 vulnerability development evaluation conducted with Mozilla, Sonnet 5 produced no usable exploit, for a success rate of 0%. Its partial success rate was 13.2%, above Sonnet 4.6 at 8.8% but far below Opus 4.8 at 68.8%. Anthropic said cybersecurity protections are enabled by default.