Anthropic released Claude Sonnet 5 on Tuesday, its midsize model rebuilt to run autonomous tasks like browsing, planning, and coding. The company's announcement pitches it as the most agentic Sonnet yet, able to handle work that recently needed bigger, pricier models.
The headline claim: performance close to Opus 4.8, the flagship, but cheaper. Those benchmarks are self-reported. On one agentic coding test Anthropic cites, Sonnet 5 hits 63.2% against Opus 4.8's 69.2%, so the gap narrows without closing. It replaces Sonnet 4.6 from February.
Pricing starts at $2 per million input tokens and $10 per million output through August 31, then climbs to $3 and $15. Worth noting: Sonnet 5 uses an updated tokenizer, and the same input can map to up to 1.35x more tokens depending on content, so the intro discount may not mean lower bills.
"That used to stall halfway," said Zapier engineer Daniel Shepard of a two-part Salesforce automation the model finished end to end. It's a vendor testimonial, but it points at the reliability gap that's kept agents stuck in pilots.
Anthropic also says Sonnet 5 has a much lower ability to perform dangerous cybersecurity tasks than its Opus models, a deliberate framing as the company negotiates model access with the Trump administration. It's the default for Free and Pro now, with API access via claude-sonnet-5.
Bottom Line
Claude Sonnet 5 scores 63.2% on Anthropic's agentic coding benchmark versus Opus 4.8's 69.2%, at a fraction of the price.
Quick Facts
- Replaces Sonnet 4.6, released February 2026
- Intro pricing: $2/M input, $10/M output through Aug 31, 2026
- Standard pricing after: $3/M input, $15/M output
- Agentic coding benchmark: 63.2% vs Opus 4.8's 69.2% (company-reported)
- Tokenizer change can map input to up to 1.35x more tokens
