Google Ships Gemini 3.7 Flash With Adjustable Reasoning Tiers

Google DeepMind has released Gemini 3.7 Flash, a high-efficiency model optimized for software engineering and agentic workflows. Arriving just three weeks after its predecessor, the model introduces tunable thinking configurations that allow developers to adjust reasoning depth based on cost and latency constraints. Google has designated this release as its primary engine for coding and autonomous agents.

Developers building multi-agent systems can route workloads through low, medium, or high thinking levels. The minimal setting remains unsupported. The model maintains a 1M token context window and a 64k token output limit, with a knowledge cutoff of March 2026 for most domains.

Benchmark Performance

Gemini 3.7 Flash demonstrates substantial gains over version 3.6, particularly in coding and enterprise automation tasks. The algorithmic improvements to its reasoning foundation translate directly to higher resolution rates in complex workflows.

Benchmark	Metric	Gemini 3.7 Flash	Gemini 3.6 Flash
DeepSWE v1.1	Long-horizon SWE	65.3%	49.0%
FrontierCode 1.1 Main	Production code	43.6%	34.4%
WebDev Arena	Elo	1588	1538
AutomationBench	Enterprise workflow	30.4%	17.0%
GDP.pdf	PDF comprehension	34.0%	22.0%

These results position the model as a strong candidate for developers evaluating and testing AI agents in production environments, where long-horizon reliability remains a bottleneck.

Pricing and Availability

The model is available immediately via the Gemini API, Google AI Studio, Android Studio, and the Gemini Enterprise Agent Platform. It also powers Gemini Spark for AI Pro and Ultra subscribers.

Google has launched the model with temporary introductory pricing, representing a 50% reduction from the 3.6 Flash launch cost.

• Through Dec 31, 2026: $0.75 per 1M input tokens / $3.75 per 1M output tokens.
• Starting Jan 1, 2027: $1.50 per 1M input tokens / $7.50 per 1M output tokens.

This aggressive pricing strategy coincides with leadership changes at DeepMind, where Koray Kavukcuoglu has taken over from Demis Hassabis. The rapid iteration of Flash models suggests a shift toward high-frequency optimization while the flagship Gemini 3.5 Pro remains unreleased.

If you rely on the Gemini API for high-volume automated reasoning, you should benchmark the new tunable thinking levels against your current workloads. The ability to trade latency for reasoning depth at the API level requires a reevaluation of how you budget tokens across complex task pipelines.