Google Gemini 3.7 Flash: Same-Base, Three Weeks, Half the Price

Three weeks after Gemini 3.6 Flash, DeepMind ships 3.7 Flash on the same base model with a +34.4% relative jump on DeepSWE, a +78.8% jump on AutomationBench, and an API price halved until December 31, 2026 -- then doubling on January 1, 2027. The workhorse is the lead indicator for what post-training alone can buy when the harness is locked.

August 17, 2026 — On August 13, DeepMind shipped Gemini 3.7 Flash to general availability. Three weeks earlier, Gemini 3.6 Flash shipped. The model page describes the new release as “based on Gemini 3.6 Flash” with “algorithmic improvements to its core reasoning foundation” — not a fresh pretraining generation. Same base, post-training iteration, three weeks. The jump on coding and agent benchmarks is large enough that DeepMind’s own numbers cross several of the lines we drew at the Aug 10 harness post, and the API price is half of what 3.6 Flash charged on day one.

This is the workhorse tier, not the frontier. DeepMind’s own model page puts it there, and so do the benchmark numbers — 3.7 Flash lands within a few points of GPT-5.6-Terra on long-horizon benchmarks while costing roughly half. The headline economic fact is the dated discount: $0.75 input / $3.75 output per million tokens through December 31, 2026, then $1.50 / $7.50 starting January 1, 2027. Anyone building an agent product on Flash needs to know that the clock is running now, not later.

What moved in three weeks

DeepMind’s release table publishes results for both 3.6 Flash and 3.7 Flash on the same evaluation harness, so the deltas are real:

Benchmark 3.6 Flash 3.7 Flash Delta
FrontierCode 1.1 Main 34.4% 43.6% +9.2pp (+26.7% relative)
DeepSWE v1.1 48.6% 65.3% +16.7pp (+34.4% relative)
AutomationBench 17.0% 30.4% +13.4pp (+78.8% relative)
Terminal-bench 2.1 78.0% 85.8% +7.8pp
Terminal-bench 3.0 5.4% 14.9% +9.5pp
WebDev Arena Elo 1538 1588 +50

Two rows are worth pausing on. DeepSWE — the coding-agent benchmark that matters most for product work — moves from sub-half to two-thirds in three weeks on the same base. AutomationBench more than halves the gap to its ceiling in a single revision. The numbers DeepMind quotes from its own table are 43.6, 65.3, 30.4, 1588, and 85.8 — they are not rounded up to nicer fractions because DeepMind does not round them up, and the post does not either.

What didn’t move

WebDev Arena gained 50 Elo. That is real movement on a public leaderboard, but 50 Elo is roughly the noise floor for a model that gets evaluated in batches over a week. Terminal-bench 3.0 — the harder, less-publicized successor — moved from 5.4% to 14.9%. Both improvements matter, but neither tells the same story as the coding-agent rows. The honest read is that post-training on a locked base can buy a lot on the benchmarks where the reasoning trace and tool-use scaffolds already exist, and substantially less on the benchmarks where the harness has to invent its own workflow on the fly.

The price schedule is the story

The 3.7 Flash launch rate is $0.75 input / $3.75 output per million tokens — half of 3.6 Flash’s prior list rate, and the rate any new product shipping this week will budget against. The catch is the date on the discount page. CometAPI’s release notes and Google Cloud’s Gemini Enterprise Agent Platform pricing page both document the same step: the intro rate holds through December 31, 2026, and resets to $1.50 / $7.50 on January 1, 2027. That is a 138-day window before the price doubles, with no announced extension.

For anyone running an agent product on Flash today, the implication is operational, not architectural. Budget the lower rate for the rest of 2026. Plan the architecture assuming the higher rate is what Flash costs in steady state. If the product can’t tolerate $1.50 / $7.50, the migration path is not “wait and see whether Google extends the intro rate.” It is “design for the reset now.”

Spec sheet, distribution, and what’s missing

The spec is the workhorse anchor: 1M-token input context, 64K output, knowledge cutoff March 2026 (some domains January 2025), function calling, code execution, search grounding, structured output, context caching, batch / Flex / Priority inference, computer use in preview, multimodal input across text, image, video, audio, and PDF. Image generation and audio generation are still off the table — the model reads them, it does not produce them.

The distribution is unusually wide for a three-week iteration. Flash 3.7 is live in the Gemini API, Google AI Studio, Vertex AI, Antigravity, Android Studio, Gemini Enterprise Agent Platform, GitHub Copilot, and Gemini Spark (now in 160+ countries). DeepMind’s own showcase this week is an Antigravity-built 3D game that uses 3.7 Flash plus Nano Banana for real-time character generation. The Antigravity angle is worth a sentence: a workhorse model that can hold a multi-step build session against a generative image model in the same loop is the unit of agent product that the post-Aug-12 browser-bifurcation stack was built to host.

The inverse of the harness thesis

On August 10, the post was that wrapping the same model in different agent harnesses can swing SWE-bench Pro pass@1 by 29 percentage points (TopClanker, 2026-08-10). On August 12, the post was that the browser is splitting into a human-UX stack and an agent-cost stack around that exact problem (TopClanker, 2026-08-12). Today’s post is the third leg.

When the harness is locked – when the wrapper, the scaffold, and the browser are stable enough to compare apples to apples – the model inside the wrapper can still move 34.4% relative on DeepSWE in three weeks, with no change to the base weights that anyone outside DeepMind can see. The harness swing does not disappear. The post-training gain on top of the locked harness is the inverse pressure: every benchmark number published in 2026 is partly a harness number and partly a post-training number, and the frontier labs are optimizing both at once. The honest reading is that the leaderboards are noisier than they look, in two dimensions at once, for the same reasons — and the workhorse tier is where you can see the post-training axis most clearly, because DeepMind publishes the comparable rows side by side.

The practical takeaway is the dated discount. Build for the $1.50 / $7.50 reset. The model is good enough either way.

Sources