Sonnet 5.5: what the cheaper, faster claims mean for your bill

The problem

Anthropic released Claude Sonnet 5.5 on September 28, 2026 and pitched it as a faster, cheaper everyday model. If you build with AI coding tools or pay for API usage, the useful question is not whether the launch is exciting but which of the claims will change your own invoice. The two accounts we read agree on the direction (faster, less token use) but emphasize different things, and one of them documents a failure mode at the highest thinking setting. Because the mid-tier model is what most agent workflows default to, a small change in tokens per task compounds quickly. This brief separates what was reported from what still has to be measured on your workload.

Our take

Both sources say the model is faster and uses fewer tokens, and one says the price is unchanged from Sonnet 5. Read together, that means any saving arrives through token count, not through a lower rate card, so it will differ from job to job. Neither piece prints a per-million-token price, and we have not verified one, so do not build a budget from the headline. The two also frame the coding story differently: one says Sonnet 5.5 beats Opus 5.5 on agentic coding because you can run several agents for the same money, while the other says it is almost as good as Opus 5.5 on some coding tasks. Those are different claims, and only the second is a like-for-like comparison. The most concrete data point is a warning: the author reports the top thinking setting used 128,000 tokens, cost $1.28 and returned nothing, while a lower setting cost about six cents. Our read is to leave defaults alone until you replay a sample of your own tasks and compare token counts side by side.

What was announced

Anthropic released Claude Sonnet 5.5 on September 28, 2026, roughly three months after Sonnet 5, according to TechCrunch. The outlet reports that it runs about 30% faster than its predecessor, consumes noticeably fewer tokens, and is aimed at everyday work such as coding and document creation. TechCrunch also says a new Haiku model is due in the coming weeks, with no date given.

What the independent write-up adds

Simon Willison states that the price matches Sonnet 5, that it is more than 30% faster, and that most work costs up to 30% less. He also notes that the free tier on claude.ai now runs on Sonnet 5.5.

His own runs are the only hard numbers in either source. At the maximum thinking effort, one image-generation prompt consumed 128,000 tokens, cost $1.28 and produced no result. At the next level down (xhigh), the same kind of prompt took 41 seconds, cost about $0.057 and worked. He reports that Opus 5.5 shows the same maximum-effort problem.

Where the two accounts differ

What is still unknown

No per-token price appears in either source, and neither reports independent benchmark results beyond the vendor's framing. Haiku 5.5 pricing is not yet public.

Best for

• You run API-billed agent or coding jobs on Sonnet 5 todayReplay a sample of real tasks on Sonnet 5.5 and compare token counts before switching defaultsBoth sources describe savings through fewer tokens at an unchanged price, so the gain only shows up in your own usage.
• You set thinking effort to the maximum levelStay one level below (xhigh) until Anthropic addresses the reported problemThe one documented run at maximum effort spent 128,000 tokens for no output, and the same author reports it on Opus 5.5 too.
• You use the free claude.ai tierDo nothing; you are already on the new modelWillison reports the free tier now uses Sonnet 5.5.