On September 28, Anthropic shipped Claude Sonnet 5.5 — its second 5.5-generation release in barely a week, right after the flagship Opus 5.5. Nothing about the headline screams breakthrough. But the numbers behind it are worth a look: the same task now runs 30% faster and costs up to 30% less, without Anthropic touching the price per token at all.
The catch is that pricing stayed put — $2 per million input tokens, $10 per million output tokens, same as Sonnet 5. The savings come from the model needing fewer steps, tokens and tool calls to land the same result. Early adopters back this up: Slack reports 14% fewer output tokens per task, Zendesk saw a 20% speed bump, and Box measured 2.4x faster runs while burning 12% fewer tokens overall.
The benchmark jump is even sharper than those percentages suggest. Terminal-Bench 4.0 went from 10.3% under Sonnet 5 to 70.6% — nearly a sevenfold gain. On GDPval-AA v2.1, Sonnet 5.5 scored 1844 against Opus 5.5's 1846, closing a gap that used to separate a flagship from a workhorse model. In a smaller but telling detail, it's also the first Sonnet to beat Pokémon Red using screenshots alone.
Anthropic keeps the same division of labor: Opus 5.5 stays the pick for open-ended work that needs sustained judgment, while Sonnet 5.5 handles well-scoped everyday jobs — bug fixes, documents, spreadsheets, slides. It's live now on Anthropic's API, AWS, Google Cloud and Microsoft Azure under the model ID claude-sonnet-5-5, with an updated Haiku promised for "the coming weeks" — no firm date yet.
There's a real tradeoff buried in the release notes, and Anthropic doesn't hide it. Sonnet now carries the same cyber safeguards previously reserved for Opus and Fable, which the company admits could mean more refusals even on harmless security tasks. Coming right after OpenAI slashed GPT-5.6 pricing last week, the fight between AI labs is clearly less about raw intelligence now and more about who can do the same job for less.
