Claude Sonnet 5.5: same price list, but your bill depends on tokens

On 28 September Anthropic released Claude Sonnet 5.5, the mid-range model that many companies use behind chatbots and automations ². The price per token is identical to Sonnet 5: $2 per million input tokens and $10 per million output tokens ¹.

That sounds like good news with no strings attached. It isn't quite, because with models that "reason", the bill is set not by the price list but by the amount of text the model produces.

What Anthropic promises

Anthropic states that Sonnet 5.5 generates answers more than 30% faster than Sonnet 5 ². It also states a cost per task up to 30% lower, because the model would use fewer tokens to reach the same result ².

The "up to" matters: it is the best case, not the average. And it is a measurement by the company that sells the model.

What the people who tested it measured

Artificial Analysis, which compares models using the same tests, puts Sonnet 5.5 at 56 points on its index, 2 points behind Opus 5.5 at its maximum level ³. In the same test, however, it counted about 193,000 output tokens per task, the highest consumption they have ever measured ³. The result is a cost of about $7.60 per task, roughly 50% above Sonnet 5 ³.

The two findings don't contradict each other. Artificial Analysis pushed the model to the highest effort level, where it reasons at length; Anthropic is talking about typical work. The point for you is a different one: the same model, at the same price per token, can cost less or more depending on how you set it up.

The effort level is your cost dial

In the Claude apps the default level is Medium; on the developer platform it is High ². Above that there are further levels, up to "max" ⁴, and that is the level at which Artificial Analysis measured the highest cost ³.

For a chatbot that answers questions about opening hours, shipping and returns, high effort is almost always wasted. For an agent that has to read a contract and prepare a draft, it may be worth the spend. The choice should be made task by task, not once for the whole company.

What to do now

• If you use Sonnet 5 through a supplier, ask them which effort level they use and whether they plan to move to Sonnet 5.5.
• If you use it directly, try Sonnet 5.5 on a sample of real requests and compare two figures against Sonnet 5: the quality of the answers and the output tokens.
• Set a monthly spending cap on the API account before you change model, not after.
• If you use a Claude subscription for your employees, the per-token price list doesn't affect you: for you, only the speed changes.

A new model at the same price is an opportunity, not an automatic saving. The saving comes from whoever measures.