One model, two speeds: Claude 3.7 and the end of picking a model per task

Anthropic released Claude 3.7 Sonnet today, 24 February 2025, alongside a command-line coding tool. Its notable feature is hybrid reasoning: the same model answers quickly for simple requests, and can be told to think for longer on hard ones.

Until now these were separate products. You picked a fast cheap model or a slow expensive reasoning model, and you had to know in advance which your task needed.

Why merging them matters practically

Because in an SME, nobody knows in advance. The person comparing a contract and the person rewriting an email are the same person, twenty minutes apart, in the same window. Asking them to switch models by task is asking them to make a technical judgement they have no basis for.

A model that adjusts its own effort removes that decision. It also removes a category of waste: paying reasoning prices for « make this friendlier ».

What to actually watch

Cost predictability. If effort varies per request, so does cost. For a company budgeting this, ask your vendor how the extended mode is billed and whether you can cap it. A surprise invoice is the most common way a promising pilot gets cancelled.

Latency. Thinking longer means waiting longer. Fine for a contract comparison you started before coffee. Not fine for a lookup during a customer call. Know which of your tasks are in which category.

The pattern across this month

Deep research on the 2nd, hybrid reasoning today. Both are the same underlying move: spend variable amounts of computation depending on how hard the question is, rather than a fixed amount every time.

That is a quiet but real shift in what you are buying. Twelve months ago you bought access to a model. Increasingly you are buying an amount of thinking, and the interesting commercial question is becoming how that gets metered.

Worth one question at your next renewal: what exactly am I billed for, and can I see it per task? Vendors who can answer have built something considered. Those who cannot will discover the question from their own customers soon enough.