Gemini 3.7 Flash Ships At Half Price, Bigger Coding Gains
Google cut Gemini 3.7 Flash token prices in half and posted double-digit benchmark gains over 3.6 Flash. Here is what changed, and what it costs after 2026.
Google shipped Gemini 3.7 Flash on Thursday, three weeks after Gemini 3.6 Flash, cutting the model's list price in half while lifting coding and agent benchmarks by double-digit margins. Google calls it "our most intelligent workhorse model yet for coding and agents," and the release notes point to a fast internal iteration cycle rather than a single big jump.
The benchmark gains, in numbers
Google's own release post lists several head-to-head scores against Gemini 3.6 Flash. On FrontierCode 1.1 Main, 3.7 Flash hits 43.6% versus 34.4% for 3.6 Flash. On DeepSWE v1.1, a coding-agent benchmark, it scores 65.3% versus 49.0%. On web development, Google says 3.7 Flash outperforms its predecessor on Arena.ai's WebDev Arena with an Elo score of 1588 versus 1538, and generates more functional layouts and feature-complete apps in fewer prompts.
Outside coding, Google reports gains on knowledge-heavy tasks: 34.0% versus 22.0% on the GDP.pdf document-processing eval, and 30.4% versus 17.0% on AutomationBench, which tests whether a model can complete real-world business workflows end to end. All of these figures come from Google's own benchmark reporting, not an independent third-party test.
| Benchmark | 3.6 Flash | 3.7 Flash |
|---|---|---|
| FrontierCode 1.1 Main | 34.4% | 43.6% |
| DeepSWE v1.1 | 49.0% | 65.3% |
| WebDev Arena (Elo) | 1538 | 1588 |
| GDP.pdf | 22.0% | 34.0% |
| AutomationBench | 17.0% | 30.4% |
Google also demonstrated a handful of showcase use cases alongside the numbers: generating a playable 3D game from a text prompt by pairing 3.7 Flash with its Nano Banana image model, orchestrating sub-agents to build interactive landing pages, and training a robotics model inside a three-agent graph loop. These are Google's own demo builds, not benchmarked results, and are worth treating as illustrations of what the model can be wired into rather than a claim about typical output quality.
Pricing: half off now, full price returns in 2027
Gemini 3.7 Flash is available through the end of 2026 at an introductory price of $0.75 per million input tokens and $3.75 per million output tokens — Google describes this as half the original 3.6 Flash cost per million tokens. That discount is temporary: starting January 1, 2027, the price rises to $1.50 per million input tokens and $7.50 per million output tokens.
Vercel's AI Gateway is running its own promotion on top of that: Gemini 3.7 Flash is listed at 50% off on the Gateway until December 31, 2026, with no platform markup on inference. To call it through the AI SDK, developers set the model string to google/gemini-3.7-flash — a configuration example, not a guaranteed price across every provider.
Fewer retries, according to Google
Google frames part of the upgrade as a developer-experience change rather than a raw benchmark one: the model "better adapts to roadblocks, clarifies intent when needed, and follows instructions with greater fidelity," and "thinks more diligently, putting in more effort into multi-step planning and tool calls." Google's stated conclusion is that this means less manual oversight and fewer retries across engineering workflows — a claim about workflow behavior that the benchmark table above doesn't directly measure, so treat it as Google's characterization rather than an independently verified number.
Where it's rolling out first
Google is routing its own consumer AI agent through the new model immediately: Gemini Spark, available to Google AI Pro and Ultra subscribers in more than 160 countries, started using 3.7 Flash the same day as the announcement. Google says the update makes Spark more efficient at Workspace tasks like consolidating files, drafting emails, and updating status documents.
On the developer side, Google points builders to Google Antigravity, the Gemini API in Google AI Studio, and Android Studio for direct access, plus the Gemini Enterprise Agent Platform for business deployments. Vercel added its own path the same day, folding the model into AI Gateway's coding-agent setup for Claude Code, Codex, OpenCode, and Pi.
Safety changes shipped alongside the model
Google says 3.7 Flash ships with updated safeguards against misuse in chemical, biological, radiological, and nuclear (CBRN) domains and cyber offense, framed as part of its ongoing Frontier Safety work rather than a response to a specific incident. The company did not publish new red-team numbers alongside this claim in the release post, and pointed instead to a separate 3.7 Flash model card for detail.
What this actually changes for builders
The practical story here is cost, not just capability: a model that's meaningfully better at agentic coding tasks, priced at half of what its predecessor cost per token, for a fixed six-month window. Teams already running 3.6 Flash in production get a straightforward incentive to test 3.7 Flash before the January 2027 price reset, since the coding and web-dev gains are large enough on paper to change which tasks are worth routing to a Flash-tier model versus a larger one. The open question the release doesn't answer is how these gains hold up outside Google's own benchmark suite — that's the test worth running before committing a production pipeline to the switch.
More from DangMua