Gemini 3.7 Flash: Google's New Coding & Agent Model Explained

What is Gemini 3.7 Flash?

Gemini 3.7 Flash is the primary agentic workhorse in the Gemini 3 family, sitting between the deep-reasoning Pro models and the high-throughput Flash-Lite tier. It is multimodal, accepting text, images, audio, and video, with a 1 million token context window and up to 64K output tokens. Its knowledge cutoff is March 2026.

Google positions it as its most intelligent workhorse model yet for coding and agents, optimized for multi-step orchestration, full-stack code refactoring, and general reasoning. It launched alongside a notable gap in Google's lineup: the more powerful Gemini 3.5 Pro remained delayed at the time of release, putting extra weight on the Flash tier as Google's day-to-day model.

What's new in Gemini 3.7 Flash?

The headline change is control over the reasoning effort. Gemini 3.7 Flash exposes a thinking level you can tune to trade quality against cost and latency, with supported values of LOW, MEDIUM (the default), and HIGH. Unlike some earlier models, MINIMAL is not available and will return an API validation error.

• Better coding. It generates more deployable, production-ready code on the first try and can complete app builds in fewer prompts.
• More disciplined agents. It adapts to roadblocks, clarifies intent when needed, and puts more effort into multi-step planning and tool calls, which means fewer retries and less manual oversight.
• Stronger document work. It handles complex documents and real-world business workflows markedly better than 3.6 Flash.
• Improved Workspace tool use. It powers Google's Gemini Spark agent for more efficient knowledge work across Google Workspace apps.

Gemini 3.7 Flash benchmarks

The reported gains over 3.6 Flash are concentrated in coding, agents, and document processing:

• FrontierCode: 43.6% versus 34.4%
• DeepSWE: 65.3% versus 48.6%
• WebDev Arena: 1588 Elo
• GDP.pdf (complex document processing): 34.0% versus 22.0%
• AutomationBench (real-world business workflows): 30.4% versus 17.0%

The gains are real but incremental, in line with a refinement rather than a new base model. On some evals the picture is mixed: rival GPT-5.6 Terra still leads on terminal and computer-use agents, and CharXiv shows a small regression against 3.6 Flash.

Gemini 3.7 Flash pricing

Price is the strongest part of the story. Gemini 3.7 Flash lists at $0.75 per 1M input tokens and $3.75 per 1M output tokens, roughly half the original 3.6 Flash list rate. That rate is introductory and expires on December 31, 2026; from January 1, 2027 it rises to $1.50 input and $7.50 output.

For context, in Google's own comparison Claude Sonnet 5 sits at $2.00 / $10.00 and GPT-5.6 Terra at $2.00 / $12.00 per 1M tokens. At a typical 80/20 input-output mix, that puts Gemini 3.7 Flash near a blended $1.35 per 1M tokens today, well under both rivals. For teams running agents at volume, that intelligence-per-dollar gap is the main reason to evaluate it, more than any single benchmark win.

How to access and use Gemini 3.7 Flash

The model is broadly available across Google's developer and enterprise surfaces, including the Gemini app (rolling out via Spark, which requires an AI Pro or Ultra subscription), Google Antigravity, AI Studio, Android Studio, the Gemini Enterprise Agent Platform, and the Gemini Enterprise app.

For the API, target the model string:

gemini-3.7-flash

If you are migrating from an older configuration, note a few Gemini 3.x requirements: strip temperature, top_p, and top_k from generation configs, replace thinking_budget with the thinking_level string enum, and remove candidate_count, which is unsupported in Gemini 3.x. Prefilled model turns should also be removed. Always confirm details against Google's official documentation, since API behavior can change.

Gemini 3.7 Flash vs 3.6 Flash and rivals

Against 3.6 Flash, version 3.7 is a cheaper, sharper iteration: clear coding and document gains, better agentic discipline, and roughly half the price, shipped only three weeks later. One partner reported the 3.7 Flash agent ran about 35% cheaper than 3.6 Flash with a higher prompt-cache hit rate and fewer tool errors.

Against competitors, the trade-off is clearer on price than on peak capability. Gemini 3.7 Flash undercuts Claude Sonnet 5 and GPT-5.6 Terra substantially on tokens, but GPT-5.6 Terra still holds an edge on terminal and computer-use agent tasks. The right choice depends on whether your workload is cost-sensitive at scale or leans on the specific agent capabilities where rivals lead.

Should you switch to Gemini 3.7 Flash?

Gemini 3.7 Flash is a strong default for teams building coding assistants, document-processing pipelines, and tool-using agents where cost per token matters. The tunable thinking level lets you dial latency down for real-time tasks or quality up for harder reasoning, and the low introductory price makes high-volume agent workloads far cheaper to run.

Two caveats are worth planning around. First, the introductory pricing doubles on January 1, 2027, so build your cost model on the post-introductory rate if you are committing long term. Second, it is a refinement rather than a leap, and rivals still lead on some agentic tasks, so validate it on your own workload before switching production systems. For deterministic, workflow-controlled agent pipelines, its improved instruction-following and fewer tool errors are the most practically useful gains.