Why Google shipped four Gemini Flash models in four months

Why Google shipped four Gemini Flash models in four months

Pro is still Gemini 3.1. Flash is on 3.8. Gemini 4 is a pretraining run, not a model you can call.

From Google I/O on May 19 to the September 2 launch of Gemini 3.8 Flash, Google put four numbered Flash models into production: 3.5, 3.6, 3.7, and 3.8. The same stretch added 3.5 Flash-Lite, 3.5 Flash Cyber, and 3.8 Flash Cyber. The public Pro checkpoint is still Gemini 3.1 Pro.

That is the whole story of the summer. Flash is the model Google can revise and ship. Pro is the model it has not been willing to replace. Gemini 4 is the next base model, and Google has said it is in pretraining — nothing more.

Four Flashes, one job each

3.5 Flash (May 19) was the I/O model. Google called it frontier intelligence with action and said it beat Gemini 3.1 Pro on Terminal-Bench 2.1, GDPval-AA, and MCP Atlas. It became the default in the Gemini app, AI Mode in Search, Antigravity, and Spark. 3.5 Pro, they said, was already in internal use and would roll out the next month.

3.6 Flash (July 21) was the cost patch. Google said it used fewer output tokens, took fewer reasoning steps, and cut the output price. The same post shipped 3.5 Flash-Lite and 3.5 Flash Cyber, and it is where Google first said the Gemini 4 pretraining run had started.

3.7 Flash (August 13) was the completion patch, 23 days later. The model card says it is based on 3.6 — algorithmic changes to the reasoning stack, which is how a three-week gap is possible. We wrote up that stretch in why three Flashes landed in 12 weeks.

3.8 Flash (September 2) is the same pattern one more time. Google called it the third Flash release in six weeks and its best reasoning and coding model yet, at 3.7's speed and introductory price: $0.75 / $3.75 per million input and output tokens through December 31, 2026, then $1.50 / $7.50. The 3.8 model card says it is based on 3.7. On DeepSWE v1.1, Google's table shows 73.7% versus 65.3% for 3.7, next to Claude Opus 5 at 74.0%.

Same line. Four version numbers. Each one after 3.5 is a revision of the previous public checkpoint.

Why Flash can move this fast

A new pretrain takes months. A post-training patch on a frozen base can ship in weeks. 3.7 is based on 3.6. 3.8 is based on 3.7. The version number goes up because the behavior on agents changed, and because developers will not switch if the old name still looks current.

Flash also has somewhere to go. It is already the default in Search, the Gemini app, Antigravity, and the API, so the next patch can aim at the failure those products already showed: too many tokens in July, too many abandoned tool loops in August, another DeepSWE climb in September.

The competitive reason sits next to that. While the flagship waits, Flash is the Gemini Google can put on a scoreboard beside Claude, GPT, and Grok. A workhorse that ties a much more expensive model on a coding agent benchmark is a launch. A delayed Pro is a footnote.

Why Pro is still 3.1

The longer version of this decision is why Google is still on Gemini 3.1 Pro.

At I/O, Sundar Pichai told developers 3.5 Pro was coming in June. On July 21 Google said it was still in partner testing and would ship when it was ready. As of mid-September, DeepMind's public pages still describe 3.5 Pro as coming soon, and the API catalog's Pro model is Gemini 3.1 Pro Preview.

Reporting after the June miss — including Decrypt's account of Bloomberg — said the larger model fell short internally, especially on coding, and a late-June data refresh disappointed. That lines up with the product Google actually shipped. 3.5 Flash already beat 3.1 Pro on the agent benchmarks Google chose to highlight. A Pro that loses to the company's own Flash on coding is a hard model to launch.

So 3.1 Pro stays in the catalog as the documented Pro checkpoint: longer-context pricing at $2 / $12 per million tokens up to 200k, and $4 / $18 above that. The work you feel getting better every few weeks is Flash.

What to expect from Gemini 4

Google's only on-record statement is the July 21 line: it has started its most ambitious pretraining run yet, for Gemini 4, and is encouraged by the progress. There is no release date, no model size, no benchmark, no price, and no API id.

Two tracks are easy to mash together. 3.5 Pro is a delayed model inside the Gemini 3 family, still described as partner testing. Gemini 4 is a new pretraining run announced in the same July post. Google has not said that 4 replaces 3.5 Pro, and it has not said that 3.5 Pro is cancelled.

What the summer does tell you:

• Gemini 4 will not arrive on the Flash cadence. 3.6, 3.7, and 3.8 could land weeks apart because they share a base. A new pretrain is the thing Google has been unwilling to rush. 3.5 Pro was "next month" in May and still not generally available in September.
• It has to clear 3.8 Flash on coding. The public workhorse is already at 73.7% on DeepSWE v1.1, beside Opus 5. A Gemini 4 that shows up weaker than the Flash already in Search repeats the 3.5 Pro problem.
• The first usable shape may still be a workhorse. This year's lesson is that Google ships the model it can put in products. A Pro-branded Gemini 4 that sits in partner testing for a quarter would match the 3.5 Pro pattern. A Flash-branded Gemini 4 that becomes the API default would match everything since I/O. Until a model card and an API name exist, Gemini 4 is a research status. Late-2026 dates in circulation assume a training cycle. Google has not named a quarter.

What to run on pull requests now

The model that moved is the one to review with.

• Use 3.8 Flash for the hard diff. It is the current end of the chain, at 3.7's introductory price, and it is in ThinkReview Model selection. The 3.8 announcement covers how to pick it.
• Keep 3.7 or Flash-Lite when the meter matters more than the last point of DeepSWE. Google still supports 3.7 at the same intro price.
• Treat 3.1 Pro as the stable Pro checkpoint. Use it when you want that specific model. The summer's coding gains were published on Flash.
• Ignore Gemini 4 until it has a model id. Switching a review default to a rumor is how teams miss the model that already shipped.

ThinkReview runs on the merge request you have open, on GitHub, GitLab, Azure DevOps, and Bitbucket. Pick the Flash that exists. Leave Gemini 4 for the day Google gives it a name.

Sources: Gemini 3.5, 3.6 Flash, 3.7 Flash, 3.8 Flash, and the 3.7 and 3.8 model cards. Review the next PR with the model that shipped: Install ThinkReview.