Gemini 3.7 Flash: Pricing, Benchmarks & 1M Context (2026)

Gemini 3.7 Flash is the model Google took to general availability (GA) on August 13, 2026. Its model ID is gemini-3.7-flash, with a context window of 1,048,576 tokens (roughly 1M) and a maximum output of 65,536 tokens (64k). Google positions it as "the most capable workhorse model for coding and agentic use," and it is now the default model behind Antigravity, Google's agentic development environment. Introductory pricing is set at $0.75 per 1M input tokens and $3.75 per 1M output tokens (through December 31, 2026).

Compared with the previous generation covered in Gemini 3.6 Flash pricing and benchmarks, 3.7 Flash brings three notable changes. First, an explicit thinking level control — low, medium (default), or high — for tuning how much reasoning effort the model applies. Second, clear gains on coding and automation benchmarks. Third, a standout result on raw throughput: Artificial Analysis measured it as the fastest of 186 models at 340.1 tokens per second. Its Intelligence Index is 56 (at the high thinking setting), a reminder that the speed advantage doesn't necessarily translate into the top intelligence tier.

Pricing runs in two tiers — introductory and standard. Approximate yen figures below assume 1 USD ≈ 150 JPY.

The introductory price is exactly half the standard price, so it's worth designing systems on the assumption that costs will double in 2027. Output tokens are priced five times higher than input tokens, meaning output-heavy workloads — code generation, long-form reports — will tend to dominate the total bill.

Every metric improved, but the absolute levels deserve a level-headed read. An AutomationBench score of 30.4% means roughly seven out of ten multi-step automation tasks can still fail. Rather than handing agentic automation entirely to Gemini 3.7 Flash unsupervised, design around verification steps and human review.

- Choosing a thinking level: use low for latency-sensitive chatbots and real-time responses, medium (the default) for general coding assistance and document processing, and high for complex reasoning or multi-step tool-using agent workflows. Higher thinking levels consume more thinking tokens, which increases both latency and cost
- Long-context practicalities: a 1M-token window doesn't mean you should always fill it. Cost and latency scale roughly linearly with the tokens you send, and stuffing in irrelevant context can bury the signal and hurt accuracy. Pair the large window with deliberate context selection — retrieval, chunking — rather than dumping everything in
- Verification-first for agent use: given the 30.4% AutomationBench score, avoid designing for unattended multi-step automation. Build in approval checkpoints and rollback paths

On price, cheaper alternatives exist: DeepSeek-V4-Pro-0813 and GPT-5.6 Luna both undercut Gemini 3.7 Flash. Claude Haiku 4.5, by contrast, sits at a somewhat higher $1.00/$5.00. For teams chasing raw latency along a different axis, OpenAI's ultrafast mode (GPT-5.6 / Cerebras) is worth a look. Rather than ranking models purely on benchmark deltas, it's worth testing against your actual workload — cost-sensitive, latency-sensitive, or long-context-heavy.

Gemini 3.7 Flash is available through Google AI Studio, Android Studio, and various enterprise platforms. On the consumer side, it powers the Spark feature in the Gemini app for AI Pro/Ultra subscribers. It is not currently available in the EEA, UK, Switzerland, or Nigeria.