Google released Gemini 3.7 Flash on August 13, an AI model aimed squarely at coding and agent workloads[1]. It lands just three weeks after its predecessor, 3.6 Flash. Alongside broad benchmark gains, Google set the per-million-token price at half that of 3.6 Flash for the remainder of the year.

A workhorse model replaced in three weeks

The Flash line sits at the volume end of the Gemini family, built around speed and price. Google describes 3.7 Flash as its most intelligent workhorse model yet for coding and agents[1].

What stands out is the pace. Only three weeks separate it from 3.6 Flash. Google attributes the compressed cycle to developer feedback and algorithmic improvements[1]. Read another way, the company is being pushed to update its smaller models faster as it competes with OpenAI and Anthropic in coding assistance and agentic workflows[2].

Coding and web development scores move up across the board

Google's published benchmark figures show a clear gap over 3.6 Flash[1].

On FrontierCode 1.1 Main, which measures coding ability, 3.7 Flash scored 43.6 percent against 34.4 percent for 3.6 Flash. On DeepSWE v1.1, an evaluation of software engineering problem solving, it jumped to 65.3 percent from 49.0 percent. In web development, it posted an Elo score of 1588 on Arena.ai's WebDev Arena, ahead of 1538 for 3.6 Flash. Google also claims high design fidelity when the model is handed a screenshot, an image, or a full design system as reference.

Knowledge-dense work improved as well. On GDP.pdf, which tests whether a model can process complex documents, it reached 34.0 percent against 22.0 percent for 3.6 Flash. On AutomationBench, which measures completion of real business workflows, it scored 30.4 percent against 17.0 percent. These target fields such as finance, law, and biosciences.

The margins are eye-catching, but every figure comes from Google itself and has not been independently verified[2]. Better benchmark numbers do not automatically translate into production reliability, latency, or actual cost.

Behavior tuned to reduce rework

Google says it improved the day-to-day developer experience, not just the scores[1]. The model adapts more readily when it hits a roadblock, asks for clarification when intent is ambiguous, and follows instructions more faithfully. It also puts more effort into multi-step planning and tool calls.

For anyone building agents, this is the unglamorous part that matters. A model that wanders off course requires a human watching over it and retrying every failure. More disciplined execution means less oversight and fewer retries.

Half price this year, double from 2027

The introductory price is 0.75 USD per million input tokens (about 120 yen) and 3.75 USD per million output tokens (about 600 yen)[1]. That is exactly half the original cost of 3.6 Flash.

The catch is that the rate only holds through December 31, 2026. Google states that from January 1, 2027, pricing rises to 1.50 USD for input (about 240 yen) and 7.50 USD for output (about 1,200 yen)[1]. For teams running token-hungry agents around the clock, the 2027 rate is the more realistic basis for operating cost. It is worth checking whether the numbers still work at that level before locking in an architecture around the introductory price.

※1 USD = 159 JPY (based on the August 14, 2026 New York close)

Gemini Spark switched over the same day

This is not only a developer story. Gemini Spark, the personal AI agent available to Google AI Pro and Ultra subscribers in more than 160 countries, moved to 3.7 Flash on the same day[1].

Spark is positioned as an agent that runs around the clock under the user's direction and takes action on their behalf. Google says the model update improves tool use across Google Workspace apps, raising accuracy and output quality for complex, multi-skill workflows. Consolidating files, drafting emails, and updating status documents were the examples given.

On safety, Google says the model shipped with updated safeguards against misuse in chemical, biological, radiological, and nuclear (CBRN) domains and in cyber offense[1].

Access splits three ways. Developers use the Gemini API via Google AI Studio and Android Studio, plus Google Antigravity. Enterprises get it through the Gemini Enterprise Agent Platform and the Gemini Enterprise app. Individuals reach it through Spark in the Gemini app[1].

Summary

Gemini 3.7 Flash is a practical coding and agent model that arrived just three weeks after 3.6 Flash. Google's own figures put it clearly ahead of the previous generation, at 43.6 percent on FrontierCode 1.1 Main and 65.3 percent on DeepSWE v1.1. Pricing holds at half rate, 0.75 USD input and 3.75 USD output, only through the end of 2026, and is set to double on January 1, 2027. Gemini Spark, the consumer-facing agent, switched to the new model on the same day.

Source [1]: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/

Source [2]: https://www.eweek.com/news/google-gemini-3-7-flash-coding-agents-pricing/