Sonnet 5.5 and GPT-6.1 Sol: which of your jobs still need a top AI model?

Anthropic released Claude Sonnet 5.5 on September 28, and OpenAI released GPT-6.1 Sol the very next day. Both are the middle model of their family, both cost $2 per million input tokens and $10 per million output tokens, and both labs pitch them as close to their top model for a lot less money.

Anthropic says Sonnet 5.5 is strongest at well-scoped everyday tasks, fixing bugs and creating polished documents, slides and spreadsheets, while Opus 5.5 stays clearly stronger on open-ended work that needs sustained judgment. OpenAI says GPT-6.1 Sol nearly matches GPT-6 Astra on coding, computer use and professional work at one fifth of the price. These are each lab's own results, and every lab publishes the tests it wins.

Here is the full line-up, by price per million tokens for input and output.

At the top, Claude Fable 5.1 and the invitation-only Mythos 5.1 cost $10 per million input tokens and $50 per million output tokens, exactly like GPT-6 Astra. Just below, Claude Opus 5.5 costs $4 and $20. In the middle, Claude Sonnet 5.5 and GPT-6.1 Sol both cost $2 and $10. At the small end, Claude Haiku 4.5 costs $1 and $5, and GPT-6 Luna costs $0.10 and $0.50.

So a job that moves from Opus 5.5 to Sonnet 5.5 costs half as much in tokens, and a job that moves from Astra to 6.1 Sol costs one fifth.

My take is that the question has moved from which lab to pick to which jobs still need a top model at all. I would keep the top tier for open-ended work where a wrong call is expensive, like system design, a large refactor or a contract review. I would move the jobs with a clear finish line to the middle tier, like fixing a known bug, drafting a report from a template or turning meeting notes into slides. Anthropic also documents a setup where the cheaper model does the bulk of the work and hands the hard decisions to the top model, so most of the tokens are billed at the lower rate.

Both labs give the same advice on how to decide, and I agree with it: test on your own work before you trust any benchmark. Pick one job your team repeats every week, run twenty real examples on both models, and move it only where the answers hold up. We have not measured the saving on our own work yet, so take this as a method to test rather than a result.

The middle tier just got good enough for most daily work, so the next saving will come from sending each task to the cheapest model that does it well.