Google released Gemini 3.7 Flash on 13 August 2026, calling it its most capable workhorse model yet for coding and agentic workflows. The drop lands just three weeks after Gemini 3.6 Flash — a cadence Google ties to developer feedback and algorithmic gains it says will also feed future models.
Availability is immediate across the Gemini API, Google AI Studio, Android Studio, and Gemini Enterprise Agent Platform. Google AI Pro and Ultra subscribers also get the upgrade inside Gemini Spark, the personal agent that runs across Workspace apps in more than 160 countries.
Confirmed
- Primary announcement: Google Blog — Introducing Gemini 3.7 Flash (13 Aug 2026).
- Positioning: workhorse Flash tier aimed at software engineering, web development, and knowledge-dense work (finance, law, biosciences).
- In Google’s published benchmark table vs Gemini 3.6 Flash: FrontierCode 1.1 Main 43.6% vs 34.4%; DeepSWE v1.1 65.3% vs 49.0%; WebDev Arena Elo 1,588 vs 1,538; AutomationBench 30.4% vs 17.0%; GDP.pdf 34.0% vs 22.0%. The same table also places 3.7 Flash near or ahead of Claude Sonnet 5 and GPT-5.6 Terra on several of those rows — still vendor-reported, not an independent bake-off.
- Context: up to 1,048,576 input tokens and 65,536 output tokens (per Google AI developer docs). Supported surfaces include function calling, code execution, computer use (preview), file search, structured outputs, search grounding, URL context, and thinking modes (low / medium / high). Image generation and Live API are not supported on this SKU.
- Introductory API list price through 31 December 2026: $0.75 / $3.75 per million input/output tokens. From 1 January 2027: $1.50 / $7.50 — i.e. the “half price” window is time-boxed, not the permanent Flash tier.
- Safety note in the launch post: updated safeguards for CBRN and cyber-offense misuse, with a DeepMind model card linked for detail.
Analysis
The interesting product move is not another Flash number — it is iteration speed plus a temporary price wedge. Shipping 3.7 three weeks after 3.6 signals Google is willing to burn the Flash line as a fast feedback loop for agent and coding workloads, while the intro rate undercuts the post-2026 list price enough to pull migrations and bake-offs into Q4.
Google’s own framing matters: 3.7 “thinks more diligently” on multi-step planning and tool calls, aiming to cut retries and human babysitting. If that holds outside Google’s harnesses, the AutomationBench and GDP.pdf jumps are more valuable than a single coding Elo tick — they point at document-heavy enterprise agents, not only autocomplete.
Treat the Sonnet 5 / Terra comparisons as orientation from Google’s chart. DeepSWE still shows GPT-5.6 Terra ahead in that same table (69.6% vs 65.3%), so this is not a blanket “beats everyone” claim even on vendor paper.
Unknown
- How much of the coding/agent lift reproduces under independent harnesses and real repos, not Google’s published rows.
- Whether “fewer retries” shows up as lower cost per completed task once thinking tokens, tool loops, and long context are billed in production.
- What sticky share of traffic stays on 3.7 Flash after 1 January 2027, when list rates double.
Our take
Believe the ship date and the price calendar first. Use the introductory window as a measured bake-off for coding and business agents — not as a forever Flash price. Discount the leaderboard until outsiders re-run it; the durable Google signal here is how fast Flash is iterating, and how openly the discount expires.
