Gemini 3.7 Flash Reaches Google Search AI Mode

Google has integrated Gemini 3.7 Flash into AI Mode in Google Search, giving paid Google AI Pro and Google AI Ultra subscribers access globally as of August 21, 2026. English is supported at launch, and users can select the model through the ”+” model selector in AI Mode.

The rollout follows Gemini 3.7 Flash’s introduction on August 13 and arrives roughly three weeks after Gemini 3.6 Flash. Google is positioning the new model for coding and agent workloads, with improved multi-step planning, instruction following, and intent recognition applied directly to conversational search.

Search Access and Platform Rollout

Gemini 3.7 Flash is available beyond Search AI Mode through Google AI Studio, Vertex AI, Gemini Enterprise, Google Antigravity, Gemini Spark, and the consumer Gemini app. The distribution gives developers several ways to compare the model’s behavior in a search interface, an API workflow, or an agent runtime.

Search users cannot treat the integration as a universal default. Model selection is explicit, and the initial rollout is limited to paid tiers and English. For teams evaluating AI agents, this makes controlled comparisons between the default Search experience and Gemini 3.7 Flash possible, especially for compound prompts that require multiple constraints.

Model Capabilities and Pricing

Gemini 3.7 Flash is natively multimodal, accepting text, images, audio, video, and PDF inputs. Its 1,048,576-token input context and maximum 65,536-token output support large document analysis and long-running agent interactions.

The model also exposes low, medium, and high thinking levels. Developers can trade reasoning depth for response speed instead of using one fixed inference profile across every request.

Item	Gemini 3.7 Flash
Input context	1,048,576 tokens
Maximum output	65,536 tokens
Thinking levels	Low, medium, high
Introductory input price through Dec. 31, 2026	$0.75 per 1M tokens
Introductory output price through Dec. 31, 2026	$3.75 per 1M tokens
Standard input price from Jan. 1, 2027	$1.50 per 1M tokens
Standard output price from Jan. 1, 2027	$7.50 per 1M tokens

The scheduled price increase matters for applications that use long contexts or generate extensive tool traces. If you are building a RAG system, context-window planning should account for both input volume and the higher 2027 rates rather than treating the one-million-token limit as free capacity.

Benchmark Results

Google reports gains over Gemini 3.6 Flash across software engineering, document comprehension, and workflow automation evaluations.

Benchmark	Gemini 3.7 Flash	Gemini 3.6 Flash
FrontierCode 1.1 Main	43.6%	34.4%
DeepSWE v1.1	65.3%	49.0%
Arena.ai WebDev Arena	1588 Elo	1538 Elo
GDP.pdf	34.0%	22.0%
AutomationBench	30.4%	17.0%

The largest relative improvements appear in long-horizon software engineering and enterprise automation. DeepSWE v1.1 rises from 49.0% to 65.3%, while AutomationBench increases from 17.0% to 30.4%. Those results align with Google’s emphasis on agents that must preserve instructions across several steps, although benchmark scores do not replace task-specific testing.

For production teams, the practical starting point is to test low, medium, and high thinking levels against representative prompts, tool calls, and failure cases. Track latency, token consumption, instruction adherence, and recovery behavior before routing live Search-like workloads to the model.