Google has released Gemini 3.7 Flash — just three weeks after the launch of its predecessor, 3.6 Flash. The new model is presented as the company's most capable workhorse yet for coding and AI agents.

Two claims

By Google's own benchmark results, 3.7 Flash wins on two fronts at once:

  • Capability — it beats Claude Sonnet 5 and GPT-5.6 Terra.
  • Price — it does so at half the price of those models. It is also 50% below the price of its own predecessor from three weeks earlier.

It is worth keeping in mind that the benchmarks come from the company itself; until independent verification arrives, the numbers are a claim rather than a finding. The pricing side needs no verification, though — that is a published figure.

What three weeks means

The real story may not be the model but the interval. Refreshing a mid-tier member of a model family in three weeks is not an established pace in this industry; a few months ago that gap was measured in months.

That has two consequences. The first is on the developer side: a workflow optimised around one model goes stale before that model has settled. The cost of adapting can exceed the performance gained.

The second is on price. Halving the price of your own three-week-old model is not a routine update; it is a public declaration of how fast a model's price loses its value. A product costed against today's rates lives on a different cost base within weeks.

Why the workhorse tier matters

The Flash series is not Google's most powerful model; it is its most used. In agentic workflows a single task generates dozens or even hundreds of model calls, and the vast majority of those do not require a frontier model: reading a tool's output, choosing the next step, formatting a piece of text.

In that high-volume, unglamorous work the deciding factors are not superior reasoning but unit cost and latency. Halving the price means twice the work on the same budget — in agentic systems that moves the boundary of what is feasible.

The competitive effect

The move also shows where differentiation in the model market is heading. As the quality of frontier models converges, competition drops from the top tier into the middle; the winner is not whoever ships the smartest model but whoever does the same job most cheaply.

Google's ability to price aggressively in that tier is a direct consequence of building its infrastructure on chips it designed itself. Rivals can match such a cut only out of their margins; Google takes it out of cost.