Sonnet Regression July 2026

Discussions
,
1

I’ve been comparing token usage between Claude Sonnet 5 and 5.5 on Max mode, and Sonnet 5 seems to consume significantly more tokens per task — around double the cost in my testing. Not sure if this is intentional (more thorough reasoning?) or something worth flagging. Would be interested to hear if others have benchmarked similar differences.

4

Hey, thanks for flagging that. A difference in token usage between model versions is generally expected. Different models handle reasoning differently and pull in different amounts of context per step, so an average 2x increase can just be a behavior difference, not a regression.

A couple things to keep in mind when comparing:

  • In Max mode, the context window can grow a lot, and models that pull more context on each call will show noticeably higher usage.
  • For a fair comparison, you need to run the same prompt on the same task and with the project in the same state. Otherwise, variance can easily look like 2x.

If you want us to look at specific runs, send the Request ID for a pair of tasks right upper corner in chat > Copy Request ID. With that, we can see the actual token breakdown. Note that if Privacy Mode is enabled, the Request ID might not be available.

Also, it’d be great if others share their measurements in the thread.

5

Hey Dean!

I ran private internal benchmarks on the same prompts and states in Max mode. Sonnet 5 (new tokenizer) uses noticeably more tokens than GPT 5.5. This drives higher effective costs per completed solution.

My private metrics show GPT-5.5 executive functioning plus Composer-2.5 workers give the best cost per solution right now.

Appreciate the work and wish he best for others in their aigentic journey.

Aiggy

6

Thanks for sharing the measurements, that’s helpful.

That kind of difference in token usage between models is expected. Sonnet 5 uses a different tokenizer and a different pattern for pulling in context, so in Max mode the usage can differ a lot compared to GPT-5.5. That affects the effective cost per solution you mentioned.

If you see usage in any run that looks clearly abnormal, not just higher, send the Request ID for the pair of tasks and we’ll check the token breakdown.