Gemini 3.7 Flash halves its price, then makes the case for it

Google unveiled Gemini 3.7 Flash on August 13, 2026, and this update leads with something Flash releases rarely lead with: a price cut. The model costs $0.75 per million input tokens and $3.75 per million output tokens at an introductory price that runs through the end of the year, half the original price of Gemini 3.6 Flash, which shipped three weeks earlier. Google positions 3.7 Flash as its workhorse for coding and agents.

The timing is deliberate. Google has kept a fast cadence across the Gemini lineup this year, with the 3.6 Flash generation arriving in July alongside a lighter, faster Flash-Lite. The quick 3.7 turnaround looks less like a discovery and more like a habit. Google attributes the speed to developer feedback and algorithmic improvements it says it will carry into future models. The cadence had already drawn attention: a leak suggested Google was testing a 3.6 Flash build while the expected pro model stayed absent, per the earlier leak report.

Benchmarks, and a 50% price cut

The sharper story is in the numbers Google published against the previous Flash. The company reports 43.6% on FrontierCode 1.1 Main, up from 34.4%, and 65.3% on DeepSWE v1.1 against 49.0%, a gap it frames as better first-pass code accuracy and more production-ready output. The pattern holds beyond coding: GPA.pdf goes from 22.0% to 34.0% on document-heavy knowledge work, AutomationBench from 17.0% to 30.4% on real-world business workflows, and WebDev Arena moves from 1538 to 1588 Elo.

These are Google's figures, and the usual caveats about first-party benchmarks apply. Read as a set, though, they tell a specific story: the company is claiming consistent agent gains across coding, web development, and documents, then charging much less per token for it. Slightly better in one benchmark, roughly double on another, half the price. That combination is rare in a market where cheap usually means slow, light, or inaccurate. It is also a market where rival benchmark claims shift quickly: xAI's Grok 4.6 tied GPT-5.6 Sol at 61 on a composite index before its component scores split, per the score breakdown.

What the price signals on the agent economy

Google's word for 3.7 Flash is deliberate. "Workhorse" is a bid for high-volume agent workloads, the kind where cost per million tokens compounds inside loops that run repeatedly. The company leans into the framing: the model is built for multi-step planning and tool calls, and Google says it follows instructions with greater fidelity, which it translates into fewer retries and less manual oversight in engineering workflows. The pitch is not exclusive to Google. Meta aimed its 30B Muse Glimmer squarely at agent workloads and shipped it to run in under 20GB, per the Muse Glimmer write-up.

The introductory framing is what makes the release feel strategic. A permanent price cut would read as a cost-of-goods story; a temporary one plants a flag. This, Google is saying, is what it takes to win developers over to its agent stack, and the benchmark page is part of that pitch. Half off is aggressive even in a year of aggressive token pricing: Zhipu's GLM-5.2 went viral at $0.07 per million tokens, a 95% cut that posters called "almost free," with one reply claiming the price had already 10x'd, per the GLM-5.2 pricing drama. At that point, the scores matter less than the price and the cadence. If this becomes the routine, Gemini 3.7 Flash is less a product announcement than a statement about how routinely its replacement will arrive.