Google ships Gemini 3.7 Flash three weeks after 3.6

A three-week model cycle

Google introduced Gemini 3.7 Flash on Thursday, only three weeks after 3.6 Flash. That is a short life for a model presented as a production workhorse.

The company says the update came from developer feedback and algorithmic changes. It is aimed at the jobs that now consume a large share of model traffic: writing and debugging code, working through business documents, using tools and keeping an agent moving across several steps.

This is not a new top-end Pro model. It is a faster-moving version of the model Google expects developers to use often. The quick release says something about the market too. Labs are no longer waiting for a grand model generation before replacing an everyday engine.

The gains are broad, and mostly Google's measurements

Google reports a FrontierCode 1.1 score of 43.6% for 3.7 Flash, up from 34.4% for 3.6 Flash. On DeepSWE v1.1, the model card lists 65.3% against 48.6%. It also reports a WebDev Arena Elo score of 1588, compared with 1538 for the previous model.

The same pattern appears outside code. Google lists 30.4% on AutomationBench, up from 17.0%, and 34.0% on GDP.pdf, up from 22.0%. The company describes those tests as measures of enterprise workflows and complex document understanding.

These numbers are useful signals, not a finished verdict. Some benchmarks are public, AutomationBench uses a private set, and the launch demos were built by Google. The model card does not provide a broad independent study of reliability in live customer systems.

It also records losses on a few tests. Chart reasoning was slightly lower without and with tools, for example. A newer model can improve the work Google cares about most without winning every column.

The price cut has an end date

Gemini 3.7 Flash costs $0.75 per million input tokens and $3.75 per million output tokens at launch. Batch and flexible processing cost half those amounts.

That headline price lasts through 31 December 2026. On 1 January 2027, the standard rates rise to $1.50 for input and $7.50 for output. Context caching also doubles. In other words, the launch offer is a 50% discount, not the permanent list price.

Google is giving 3.6 Flash the same temporary pricing, so developers do not save money simply by choosing 3.7 during the offer. The pitch is better work for the same current rate. After the promotion, the new model lands at the original 3.6 price.

Teams comparing models will still need to measure the cost of completed work. Output length, retries, tool calls and failed runs can matter more than the number printed beside a million tokens.

Google is putting it inside Spark now

The model is available through the Gemini API, Google AI Studio, Antigravity and Android Studio. Enterprise customers can use it through Google's agent platform and enterprise app.

It is also rolling into Gemini Spark for Google AI Pro and Ultra subscribers in more than 160 countries where Spark is available. Spark is Google's 24/7 personal agent, introduced at I/O, and can work across Gmail, Calendar and Docs under a user's direction.

That placement is more revealing than another leaderboard. Google is trusting the model with multi-step work inside products people already use, and doing so a day after launch rather than keeping it in a developer preview for months.

The open question is how much oversight those tasks still need. Google says 3.7 follows instructions more faithfully and needs fewer retries. It has not published independent evidence showing how often Spark completes real personal workflows cleanly.

What is confirmed, claimed and still open

Confirmed: Gemini 3.7 Flash was announced on 13 August 2026. It is based on 3.6 Flash, accepts text, images, audio and video, has a context window of up to one million tokens and produces text outputs of up to 64,000 tokens. Its access channels and dated pricing are public.

Google's claims: the coding, workflow, document and agent improvements come from the company's launch material and model card. Several results use external benchmarks, but Google selected the configurations and published the comparison.

Documented limits: the model can hallucinate and may occasionally be slow or time out. Google says it did not reach the tracked or critical capability levels in its frontier-safety assessment, though the supporting 3.7 report had not yet been published when the model launched.

Still open: independent performance on messy production work, the reliability of Spark over long-running tasks, and what happens to usage when the introductory price expires.

Three weeks is enough time for Google to ship another model. It is not enough time for the rest of the market to know how dependable that model is.