Claude Sonnet 5: Near-Opus Performance

Released June 30: Anthropic's "most agentic Sonnet yet", near-Opus-4.8 performance.

• 63.2% SWE-bench Pro
• 80.4% Terminal-Bench 2.1
• A GDPval-AA score of 1,618 that edges out Opus 4.8's 1,615
• Native 1M-token context by default, 128K max output

It's the new default in Claude Code and on Free/Pro plans, with intro pricing of $2/$10 per million tokens through August 31 (then $3/$15) — read by everyone as pre-IPO land-grab pricing.

#The fine print

Simon Willison flags two things worth checking before you switch: a new tokenizer that produces 1.0–1.35x more tokens (check your budgets), and new cyber-safety refusal behavior.

And a day after launch, Anthropic quietly swapped the BrowseComp cost-performance chart, saying the original methodology underestimated Sonnet 5 — cue a small but healthy benchmark-transparency debate.