Google released Gemini 3.7 Flash on August 13, 2026, as a generally available model for coding and agent workloads. The production model ID is gemini-3.7-flash, and Google has made it available through the Gemini API and Google AI Studio, alongside integrations including Android Studio.
The release is not simply a new model target. Its launch price expires at the end of 2026, while migration from some earlier Gemini models requires configuration changes. Google also reports sizable benchmark gains, but those results remain vendor-produced evidence rather than independent proof of performance in a reader’s workload.
Gemini 3.7 Flash is a production model, not a preview
Google’s Gemini API guide lists Gemini 3.7 Flash as generally available and ready for production use.
The model supports a 1-million-token context window, a maximum output of 64,000 tokens and the same built-in tool suite as Gemini 3.6 Flash. Those limits make it suitable for testing long-context coding and multi-step agent workloads, but the size of the window does not guarantee that a model will use every part of a large input equally well.
For production adoption, teams still need task-specific evaluations covering answer quality, latency, token consumption, tool-call accuracy and failure recovery.
The introductory price ends on December 31
Through December 31, 2026, Google is charging $0.75 per million input tokens and $3.75 per million output tokens for Gemini 3.7 Flash. The same introductory rate also applies to Gemini 3.6 Flash.
On January 1, 2027, the listed rates rise to $1.50 per million input tokens and $7.50 per million output tokens. Both prices therefore double at the start of 2027.
That date matters for any service expected to remain on the model into next year. A cost estimate based only on the launch rate will understate the scheduled 2027 bill unless the deployment moves to another model or Google changes the pricing again.
Teams evaluating the release should model both price periods, particularly for agents that make repeated model or tool calls within one user task.
Thinking levels trade cost and latency for more reasoning
Gemini 3.7 Flash supports low, medium and high thinking levels, with medium as the default.
Google describes low as the latency-oriented option, medium as the general setting for complex code and agent tasks, and high as the setting for the most difficult reasoning and coding work. The company also warns that high thinking can consume more tokens and increase cost.
Developers should not assume that the highest setting is automatically the best production choice. A useful evaluation should compare the levels on the same workload and record task success, latency, output tokens, reasoning-token use and downstream tool calls. The scheduled price change makes that measurement more important for long-running agents.
Migration depends on the model you use today
Moving from Gemini 3.6 Flash is mainly a model-ID change because that release had already removed several older request patterns.
Applications moving from Gemini 3.5 Flash, Gemini 3 Flash Preview or Gemini 3.1 Pro need a closer review. Google’s migration checklist tells developers to remove temperature, top_p and top_k, replace the older thinking_budget setting with thinking_level, remove unsupported candidate_count configurations and stop sending prefilled model turns.
The guide also documents function-calling and conversation-history requirements that may affect applications using the GenerateContent API or multi-turn tool flows.
A safe migration therefore needs more than changing the model string. Teams should canary the new target, replay representative requests, check tool-call schemas and compare token and latency distributions before increasing production traffic.
Google’s benchmarks are encouraging, not independent validation
In its launch announcement, Google reports that Gemini 3.7 Flash scored 43.6% against 34.4% for Gemini 3.6 Flash on FrontierCode, 65.3% against 49.0% on DeepSWE and 1,588 Elo against 1,538 on WebDev Arena.
Google also reports 34.0% against 22.0% on GDP.pdf and 30.4% against 17.0% on AutomationBench.
These figures support Google’s claim that the new model improves on its predecessor under the company’s reported evaluation conditions. They do not establish the same improvement for every repository, agent harness, prompt, tool set or latency target.
The practical decision is to treat the benchmarks as a reason to test, not as a substitute for testing. Google’s model card provides additional vendor-authored evaluation and limitation context, while production teams remain responsible for measuring the model against their own acceptance criteria.
