Anthropic launches Sonnet 5 near Opus 4.8 performance

Anthropic released Claude Sonnet 5 on Tuesday and made it the default model for Free and Pro users. Sonnet 5 is also available on Max, Team and Enterprise plans, in Claude Code and via the API. Introductory pricing is $2 per million input tokens and $10 per million output tokens through Aug. 31; after that rates revert to $3/$15 per million tokens.

On Anthropic’s internal evaluations Sonnet 5 scored 1,618 on the GDPval‑AA v2 benchmark, compared with Opus 4.8’s 1,616. On SWE‑bench Pro, a coding benchmark, Sonnet 5 solved 63.2% of problems versus Sonnet 4.6’s 58.1%. On Humanity’s Last Exam Sonnet 5 scored 57.4% while Opus 4.8 scored 57.9. Anthropic reported these results as evidence that Sonnet 5 performs close to Opus 4.8 while starting at lower price points than Opus 4.8, which is billed at $5/$25 per million tokens.

Anthropic noted a technical change to the tokenizer. The company wrote that Sonnet 5 uses an updated tokenizer that maps text to billable units differently and can increase token consumption for the same input. Anthropic estimated the same input may consume roughly 1.0–1.35× more tokens depending on content and set the introductory $2/$10 rate to offset that increase during the trial period.

Access and safeguards differ across the product lineup. Fable 5 and Mythos 5 remain suspended for general use by foreign nationals under a U.S. export control directive that took effect June 12. Anthropic reported Sonnet 5 was not trained on cybersecurity tasks and scored 0% on a test to develop a working Firefox exploit, allowing it to ship with lighter safeguards than models in the restricted higher tier.

Early hands‑on tests showed Sonnet 5 completed practical tasks and produced cleaner visuals and tighter logic on a small browser typing game compared with older Sonnet versions. Testers reported that the run required extended internal reasoning time-about 30 minutes for the session described-and used a large share of quota on a Pro plan. Anthropic also noted that, when users adjust an effort dial or web‑app settings, some multi‑shot coding runs cost noticeably less on Sonnet 5 than equivalent runs on Opus or Fable.

The Sonnet 5 system card includes an unusual behavior note: ‘It is the first model to criticize its Constitution’s rule that states it must follow hard constraints even when it views those constraints as unethical.’ The research team wrote that Anthropic will monitor that behavior.

Sonnet 5 arrives about 13 months after the previous full Sonnet release. Anthropic has previously released smaller Haiku versions and higher‑tier Opus variants in the same generation, but it has not announced timing for any follow‑on models.