30 SEP 2026 — Anthropic released Claude Sonnet 5.5 on 28 September at Sonnet 5's prices: US$2 per million input tokens and US$10 per million output. The company says the new model "costs up to 30% less for most work" because it needs fewer tokens. The independent benchmarker Artificial Analysis measured the opposite at the model's highest setting: about 50% more per task than Sonnet 5.
The two claims can both be true. They measure different effort settings, and the gap between them is the practical question for anyone paying the bill.
Anthropic's pitch
Anthropic pitches Sonnet 5.5 as the faster, cheaper complement to Opus 5.5, strongest at well-scoped everyday tasks such as fixing bugs and making documents. It says output is more than 30% faster than Sonnet 5.
On its own table, Sonnet 5.5 scores 70.6% on the Terminal-Bench 4.0 agentic coding test against 10.3% for Sonnet 5, and sits within two points of Opus 5.5 on the GDPval-AA work test. Anthropic adds that Opus 5.5 "remains clearly stronger" at complex, open-ended work.
Because its cybersecurity abilities are comparable to Opus 5's, Anthropic says this is also the first Sonnet to ship with the cyber safeguards it uses on its most capable models.
The independent measurement
Artificial Analysis scores Sonnet 5.5 at 56 on its Intelligence Index at maximum effort, second only to Opus 5.5 and two points behind it. To get there, the model used about 193,000 output tokens per task, the highest it has measured, around seven times GPT-6 Astra. That put its cost at US$7.60 per task, about 50% more than Sonnet 5's.
At lower settings the picture changes, and the tester says those configurations are beaten on value by OpenAI's models. Its conclusion is that Sonnet 5.5 sits off the best price-for-performance line at every setting, with the high setting coming closest to GPT-6 Sol.
Why the claims diverge
Per-token prices say little on their own. A reasoning model can think for more or less time before answering, and each setting changes how many tokens it spends. Anthropic says in the same announcement that Sonnet 5.5 at low or medium effort beats Sonnet 5's best score for about a tenth of the cost on several tests. The tester's 50% figure applies at the maximum setting.
One developer's test
Simon Willison found the same edge in his own test. At max effort, a drawing prompt ran for 128,000 tokens and failed to produce an answer, the same behaviour he had seen from Opus 5.5. At the next setting down it succeeded for under six cents.
Same rate card as OpenAI
Sonnet 5.5's list price matches GPT-6 Sol and GPT-6.1 Sol, which OpenAI released a day later at the same US$2 and US$10. With identical rate cards, cost per task is now the only price difference that matters, and it depends on the effort setting chosen.
Willison also notes that Sonnet 5.5 is now the model behind the free tier of Claude's consumer app. Anthropic says a cheaper Haiku 5.5 will follow in the coming weeks.
RECATOOLS is written with the assistance of Claude, made by Anthropic, the company whose model this article assesses. We have relied on independent measurements where they exist, and readers should weigh that interest.
View author profile → · Editorial policy
