Google has released a new mid-tier Gemini model and cut the price developers pay to run it. According to Google’s Gemini 3.7 Flash announcement, dated August 13, 2026, the model arrives three weeks after Gemini 3.6 Flash at an introductory rate of $0.75 per million input tokens and $3.75 per million output tokens — half the original 3.6 Flash cost per million tokens, and available at that rate through the end of the year.
What Google says changed
The pitch is coding and agent work rather than raw scale. Google reports that 3.7 Flash scores 43.6% on FrontierCode 1.1 Main against 34.4% for 3.6 Flash, and 65.3% versus 49.0% on DeepSWE v1.1, per the same announcement. On web development, Google puts the model at an Elo of 1588 on Arena.ai’s WebDev Arena compared with 1538 for its predecessor, and on document-heavy work it claims 34.0% versus 22.0% on the GDP.pdf benchmark and 30.4% versus 17.0% on AutomationBench (Google). All of those figures are vendor-reported and self-graded; no independent replication is published alongside them.
Two of the yardsticks are worth knowing before reading the numbers. FrontierCode, built by Cognition, grades whether a maintainer would actually merge a pull request, using maintainer-authored tasks, rubric plus unit-test grading, and rules that zero out runs which consult solution-bearing sources such as the original PR (Cognition’s FrontierCode methodology). AutomationBench comes from Zapier and scores agents on the final state of a realistic business environment — a CRM with live data, an inbox, a calendar with conflicts — with deterministic checks and no LLM judge, across six domains including sales, support, and finance (Zapier’s AutomationBench post). Both are outcome-style evaluations, so movement on them maps more closely to shipped work than a trivia score does.
Where it shows up
Google says developers can reach 3.7 Flash through the Gemini API in Google AI Studio, Android Studio, and Antigravity, while enterprises get it in the Gemini Enterprise Agent Platform and app. Gemini Spark, the always-on personal agent for Google AI Pro and Ultra subscribers in over 160 countries, switches to 3.7 Flash from launch day, with Google citing better tool use across Workspace apps (Google). The company also states the release ships with updated safeguards covering CBRN and cyber-offense misuse.
The practical read
For anyone already running Flash in production, the decision is mostly arithmetic: the same workload class at half the introductory token price, and Google says in its own 3.7 Flash announcement that more disciplined execution means fewer retries and less manual oversight. The caveat is the word introductory — the rate is fixed only through the end of 2026, so cost models built on it need a renewal assumption. The three-week gap since 3.6 Flash also signals a faster refresh cadence, which matters if your evaluation harness is slower than Google’s release schedule. The model is already turning up in third-party tooling as well, including Gemini 3.7 Flash’s arrival in GitHub Copilot.
Teams evaluating a switch should re-run their own task set rather than trust the leaderboard deltas. Cognition’s and Zapier’s benchmarks are designed to resist gaming, but a self-reported score on either still describes Google’s harness, not yours.
