Introducing Gemini 3.7 Flash

What it claims

Gemini 3.7 Flash replaces 3.6 Flash three weeks after that release. Google says it is better at coding, agentic tool use, web/UI generation and document-heavy knowledge work. Google attributes the gains to developer feedback and "algorithmic innovations" but does not describe them. The introductory price of $0.75/1M input and $3.75/1M output tokens is described as half of 3.6 Flash's original per-token cost. That price holds only through the end of the year, and a footnote (not included in the fetched text) qualifies it. As of today (2026-10-01), that leaves about three months at the introductory rate. The post does not state the price after that.

Evidence

All numbers are vendor-reported and compare only against 3.6 Flash:

The post gives no evaluation settings, such as thinking budget, scaffold, attempts or tools. It also gives no comparisons with other vendors' models or with larger Gemini tiers. "Early customer feedback" is cited without names or figures.

Qualitative claims:

• Better adaptation to roadblocks and clarification of intent.
• Higher instruction fidelity.
• UI generation that adheres to a reference screenshot, image or full design system.
• Fewer retries and less manual oversight.

How to use it

• Developers: Gemini API via Google AI Studio and Android Studio, plus agent workflows in Google Antigravity. A developer guide is linked.
• Enterprises: Gemini Enterprise Agent Platform and the Gemini Enterprise app.
• Consumers: Gemini Spark in the Gemini app, for Google AI Pro and Ultra subscribers in 160+ countries. It switches to 3.7 Flash from launch day.

The page does not give a model identifier string, context window, rate limits or a migration note. Get those from the developer guide or the model card.

Limits

• Cost per task is unknown. The post says the model "thinks more diligently, putting in more effort into multi-step planning and tool calls." It does not quantify the effect on output-token usage or latency. Measure tokens per task on your own workload before assuming the halved per-token price halves your bill.
• Plan for the price change. If you build a cost model on the introductory price, plan for a change after year-end.
• Stricter refusals are possible. The model ships with updated CBRN and cyber-offense misuse safeguards. Security-tooling workloads should test for new refusals. Details are in the 3.7 Flash model card, which is not in the fetched text.