Sonnet 5.5: half the price, almost everything else

Pricing is unchanged from Sonnet 5: $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 and cache writes at $2.50. Input and output cost half as much as Opus 5.5, which sits at $4 and $20, and that is the figure the whole announcement revolves around; the fact sheet says nothing about the bigger sibling's cache rates, so the comparison stops there. The Next Web notes, however, that this price doesn't fully square with rate increases previously reported for Sonnet 5. Anthropic claims output is generated more than 30% faster than Sonnet 5 and, in its own tests, a cost per task up to 30% lower, because the model supposedly uses fewer tokens and makes fewer tool calls. That "up to 30%" is an internally measured maximum, not a saving readers will see on their invoice. It is available now on Claude Platform under the ID claude-sonnet-5-5 and on the three major clouds: Amazon Web Services, Google Cloud and Microsoft Azure.

The company's own numbers put Sonnet 5.5 right on the heels of Opus 5.5: 1,844 Elo on GDPval-AA v2.1 versus 1,846, with Sonnet 5 stuck at 1,449. On Terminal-Bench 4.0 the jump is harder to read: 70.6% versus 10.3% for the previous generation, and above the 66.4% The Next Web attributes to Opus 5.5 at its highest effort level. A gain of roughly sixty percentage points in a single generation is unusual, and the announcement doesn't explain under what conditions the two scores were compared. The other results: 46.2% on FrontierCode 1.1 Main at Max effort, 55.5% on CursorBench 4.0, 80.1% on OSWorld 2.1, and 64.5% on Humanity's Last Exam with tools. Keep in mind that all of these figures, including those on speed and cost, come from Anthropic's announcement or from partner customers quoted in it — as of September 29 there are no independent evaluations. The company's notes also flag a bug in pre-release testing of structured outputs, with a "small effect" expected on scores, and warn that a fix applied to GPT-6 Sol may not be reflected in every competitor score.

The customer testimonials come from the same frame and should be weighed for what they are: metrics passed to the company and published in its announcement, not measurements anyone outside can verify. Zendesk, through its head of AI, reports tickets handled 20% faster; Box, through its AI product leadership, describes a model that is "more accurate, 2.4 times faster and uses 12% fewer total tokens" than its predecessor; Epic Games, through its operations leadership, says Sonnet 5.5 met "the same quality bar you'd expect from a higher-tier model". On safety, the news is structural rather than numerical: Anthropic says the model's cybersecurity capabilities are comparable to Opus 5's, and for the first time a Sonnet ships with the cyber safeguards previously reserved for the most powerful models, with the riskiest cyber requests rerouted — visibly to the user — to the previous Sonnet 5. The company also plans to widen access to its Cyber Verification Program, while biological safeguards stay at Sonnet 5's level. Against distillation, there are classifiers that block extraction of the model's reasoning and a "preserved thinking" feature tied to the account that produced it. SiliconANGLE adds two details: it is the first Sonnet to finish Pokémon Red, and the model embeds an invisible text watermark, though how it works has not been disclosed.

One chapter remains open: Haiku 5.5 will arrive "in the coming weeks", with no date. And there is one statement that deserves more attention than most benchmarks: according to The Next Web, Anthropic says the model does not push forward the capability frontier of its systems, and that this is precisely why the alignment evaluation covered a narrower set of risks. It is a declared methodological choice, not a hidden one, but it says something about where things are heading. In a quarter in which OpenAI halved its API prices with GPT-6 Sol and Luna, the middle tier is becoming the real battleground, and the way to win it isn't raising the capability ceiling: it's bringing down what used to sit at the top — price, speed, and now protections too. That last part strikes me as the most interesting: it's the line I'd check first in the next mid-tier announcements.

— Olya