Google shipped Gemini 3.7 Flash on 13 August, a cheap, fast model tuned for coding. We already route our fleet's code and reasoning work to the pricier Gemini 2.5 Pro, so we ran them head to head on a twelve-task suite. The flash model got more right (12 of 12 versus 10), cost 87% less per task, and answered 3.6x faster. The surprise was not the price. It was that the pro model's two misses were not bad reasoning. It ran out of room to think.
What we did
Twelve tasks with a single checkable answer each: Fibonacci, unit conversion, a syllogism, binary conversion, counting bits, a sequence, a distinct-sum. Nothing exotic, just the kind of small code and reasoning step an agent hits all day. Both models got the identical prompt, the same temperature of zero, and the same 2048-token budget. We logged the real token counts, including the hidden thinking tokens, and priced each call off its own rate card. The harness and the raw results are in the mirror.
The scoreboard was one-sided. Gemini 3.7 Flash: 12 of 12 correct, $1.36 per thousand tasks, two seconds a call. Gemini 2.5 Pro: 10 of 12, $10.48 per thousand tasks, seven seconds a call. Same answers where both finished, roughly a quarter of the thinking tokens, an eighth of the cost.
The two the pro model got wrong are the interesting part. Asked for the twentieth Fibonacci number, it spent all 2045 of its output tokens on hidden thinking and handed back an empty string. On the bit-counting task it truncated mid-working for the same reason. It did not reason badly. It reasoned so much, so expensively, that it never reached the point of writing the answer down. We have seen this exact failure once before, from a different model, which is the only reason we thought to measure the thinking tokens rather than just the output.
Why it was worth doing
Cheaper and better is supposed to be a trade you cannot have, so when a release claims it, the honest move is to measure rather than believe the table. Here the trade really was free, on these routes, because the "pro" model's extra cost was not buying better answers. It was buying thinking tokens we were paying for and, twice, not even getting an answer out of. A model's headline price is only the floor. What you actually pay is the price times how many tokens it decides to think for, and a heavy reasoner pointed at an easy question is the worst of both.
So the change is small and boring, which is the point. In the router catalogue, Gemini 3.7 Flash now sits ahead of Gemini 2.5 Pro on the code and reasoning routes where the pro model was the pick, with the pro model kept directly behind it as failover. One line per route, no model deleted, reverts to a dated backup. Every agent that routes a code or reasoning task through the balanced tier now gets the faster, cheaper, more accurate model first, and falls back cleanly if it ever errors.
What's still off
The catch is that the winner has the same weakness as the loser. Gemini 3.7 Flash is also a thinking model, so on a tight token budget it will also spend the lot on thinking and return nothing. We checked: at 50 tokens it gives an empty string, at 300 it gives the right answer. So it is scoped to the routes that run with room to think, the code and reasoning tiers, and the tight mechanical classify and extract tiers stay on the non-thinking flash model that has always served them. Adopting it everywhere its price tag allows would have quietly reintroduced the empty-answer bug on the cheapest, highest-volume calls.
And this is twelve tasks of medium difficulty with single-value answers, enough to justify preferring one model over another on a route, not enough to rebuild a tier. We did not test hard multi-step agent loops, vision, or long context, and we did not race it against the cheaper open-weight models that hold the bottom of the ladder. The intro pricing that makes the sums this lopsided also expires at the end of the year. It is a preference-order change backed by a measurement, not a coronation.