Gemini 3.7 Flash: Three Weeks as the Newest Flash Model

Google published the Gemini 3.7 Flash model card on 13 August 2026. On 2 September it published the card for Gemini 3.8 Flash, which describes itself as built on 3.7. Google's API documentation now lists gemini-3.7-flash as deprecated: it is no longer selectable in AI Studio, and requests for it are automatically routed to gemini-3.8-flash.

So 3.7 Flash was Google's newest Flash model for about three weeks. That is short enough that it is tempting to skip it. It is worth a closer look, because what it introduced still shapes the current model, and because the way it was retired is a clean example of a risk every team building on hosted models carries.

What 3.7 changed over 3.6

The model card describes 3.7 Flash as the next iteration of the Gemini 3 family, based on 3.6 Flash, with "algorithmic improvements to its core reasoning foundation". It kept the same envelope: text, image, audio and video input, a 1M-token context window, up to 64K output tokens, and a knowledge cutoff of March 2026 for some domains and January 2025 for others. Thinking was exposed as low, medium and high; minimal returns an error.

Google's comparison against 3.6 Flash shows where the reasoning work landed:

• DeepSWE v1.1 (long-horizon software engineering): 65.3%, up from 48.6%
• Terminal-bench 2.1: 85.8%, up from 78.0%
• Terminal-bench 3.0: 14.9%, up from 5.4%
• AutomationBench (private set): 30.4%, up from 17.0%
• OSWorld-2.0 (computer use): 47.9%, up from 33.8%
• GDM-MRCR v2 at 128K: 97.0%, up from 91.8%
• Artificial Analysis Intelligence Index: 56, up from 52

Not everything moved forward. CharXiv chart reasoning slipped slightly, from 85.2% to 84.5% without tools and from 89.4% to 88.7% with them. And Terminal-bench 3.0 at 14.9% is a useful reality check: a harder benchmark revision puts every model in the table, including GPT-5.6 Terra at 20.8%, a long way from solved.

Pricing did not change. Google listed 3.7 Flash at the same introductory $0.75 per million input tokens and $3.75 output as 3.6 Flash, with a note that the introductory rate expires on 31 December 2026, after which it becomes $1.50 and $7.50. The :batch variant on OpenRouter is the same model on Google's asynchronous Batch API at half those rates.

Agentic video understanding

The feature most associated with 3.7 Flash arrived on 1 September, when Google introduced agentic video understanding. Normally a video is ingested at a fixed frame rate, one frame per second by default. In agentic mode the model decides what to watch, at what speed and through which channel, whether frames, audio or transcript, and loads only the segments it needs through an internal tool.

Google reported up to 88% fewer tokens, up to 66% lower analysis cost and up to 7% higher accuracy against static processing, with the largest gains on long videos. It is switched on with "processing": "agentic" on the video input and billed at normal token rates. Google named 3.7 Flash as the best quality-to-cost option among the three supported models, alongside 3.6 Flash and 3.5 Flash-Lite.

Note the timing: that announcement came one day before 3.8 Flash shipped, and 3.8 Flash was not in its list of supported models. If you depend on agentic video and your requests to 3.7 Flash are now being served by 3.8, confirm that the feature behaves as it did, or pin gemini-3.6-flash, which was on the list and remains supported.

The redirect is a model change you did not deploy

Automatic routing is friendly in one sense: nothing breaks. In another it is a change to your production system that never went through your review process.

The specific risk here is token spend. Google's Gemini 3.8 Flash guide says the model "can use more tokens on longer running and complex tasks, by design": it takes smaller reasoning steps, calls tools iteratively and verifies its work. Its default thinking level is medium. A pipeline sized around 3.7 Flash's token use, with output caps, timeouts and budgets tuned to it, is now running a model built to use more.

That may be fine. 3.8 Flash scores higher on almost every row of Google's comparison. But "higher scores, more tokens, same price per token" is a cost increase that arrives without an invoice line explaining it. The Gemini 3.8 Flash guide covers how to keep that in check with effort levels.

How to stop model IDs moving under you

Google shipped four numbered Flash releases between May and September 2026, and two of them now redirect. Treat that as the normal pace rather than an exception.

1. Pin a model string you have chosen, and record it next to the prompt version it was tested with. Agent prompt versioning covers treating the pair as one deployable unit.
2. Watch deprecation pages for every model you depend on. A redirect will not show up as an error in your logs.
3. Log the model that served each response, not just the one you requested, so a silent swap shows up in your metrics.
4. Keep a small regression set and run it whenever the serving model changes. Agent regression suites has the setup.
5. Alert on tokens per task, not just total spend. A redirect to a more thorough model shows up there first.

If you still have gemini-3.7-flash in your configuration, make the decision explicit: move to gemini-3.8-flash after measuring its token use on your workload, or move to gemini-3.6-flash if you prefer a supported model with more predictable spend. The Gemini 3.6 Flash guide covers that option.