Three weeks. Half price. A new floor for agent work.
Google shipped Gemini 3.7 Flash on August 13 — just three weeks after Gemini 3.6 Flash, at roughly half the intro price of its predecessor. Google calls it “the most intelligent workhorse model we’ve ever released,” and for once the positioning is the story. This is not a routine point release. It is the first product move of the post-Hassabis DeepMind, and it tells you exactly where the company thinks the model war is now being fought.
The numbers: where 3.7 Flash actually improved
Every meaningful gain is in code, tool use, and long-horizon agent execution:
- FrontierCode 1.1 Main: 34.4% → 43.6% production code quality — above Claude Sonnet 5 (42.7%) and GPT-5.6 Terra (41.3%)
- DeepSWE v1.1: ~49% → 65.3% on long-horizon software engineering
- Terminal-bench 3.0: 5.4% → 14.9% on hard agentic terminal tasks
- OSWorld 2.0: 33.8% → 47.9% on computer use
- AutomationBench: 17.0% → 30.4% on real business workflows
- WebDev Arena: Elo 1538 → 1588
Notice what is missing: no headline math or knowledge benchmark. The model card reads like a checklist of the exact areas Google needed to fix — and the ones where agent workloads actually burn tokens.
The real weapon is price, not just performance
Until December 31, 2026, Gemini 3.7 Flash costs $0.75 per million input tokens and $3.75 per million output tokens — half of 3.6 Flash’s original pricing. From January 1, 2027 it reverts to $1.50 / $7.50. Even the list price is aggressive.
Think about why this matters. A chatbot calls a model once or twice. A working agent plans, searches, reads a dozen files, calls tools, retries, and feeds results back — dozens or hundreds of model calls per completed task. Once agents scale, the cost structure of AI flips from “per answer” to “per task.” The model that can run a smart-enough loop for the least money becomes the default substrate. That is the economic argument for a workhorse model, and Google is now selling it at half price to seed the habit.
A deliberate cadence, not an accident
The timeline is the tell. DeepMind’s leadership changed on August 5; 3.7 Flash shipped eight days later. Koray Kavukcuoglu, the new SVP running Gemini, has been the main interface between DeepMind and Google Cloud, and his mandate is explicitly commercial: faster iteration, recover coding leadership, get Gemini into agents, and organize TPU, Cloud, Workspace, and Android into one machine.
Brin is back in the room, pressing the core AI teams to catch the frontier. Meanwhile the flagship Gemini 3.5 Pro — announced for partner testing in July — still has no release date. The gap between a three-week Flash cadence and an absent Pro flagship is the clearest possible signal of where Google’s urgency is concentrated right now.
What this means for the market
First, the competitive frame is shifting from benchmark bragging rights to cost per completed task. OpenAI and Anthropic can still win any single leaderboard, but Google is competing on the unit economics of running agents at scale — the same playbook it used to commoditize cloud infrastructure pricing. Second, Flash models are now close enough to flagships (AA Intelligence Index: 56 vs GPT-5.6 Terra’s 57, Claude Sonnet 5’s 55) that many production workloads will never need the top tier. Third, the three-week cadence gives developers a reason to re-architect around models they can swap every few weeks.
None of this answers the open question of whether the next Pro flagship can actually lead. But the pressure is now visible: after the reorganization, DeepMind chose speed first. For everyone building on Gemini, that means a cheaper, faster-moving floor — and a price anchor the rest of the market has to respond to. See also our coverage of GPT-5.6 Sol’s 14x speed push and the agent funding winter for the broader cost-and-capex picture.
What developers should do now
- Benchmark 3.7 Flash on your own agent loops (Terminal-bench-style tasks, multi-file edits) — measure cost per completed task, not single-shot quality
- Model the December 31 promo window: lock in workloads before prices double in January
- Watch for Gemini 3.5 Pro: the Flash cadence is a signal, but the flagship is the real test
