Why Google shipped three Gemini Flash models in 12 weeks

Why Google shipped three Gemini Flash models in 12 weeks

On August 13, Google introduced Gemini 3.7 Flash — 23 days after 3.6 Flash. Zoom out one more step. Gemini 3.5 Flash launched at Google I/O on May 19. That is three workhorse Flash models in 86 days.

Frontier labs used to treat numbered releases like research milestones. This summer Google treated Flash like production software: ship the I/O model, then patch cost, then patch the retries. The interesting part is not the version numbers. It is that 3.5 Pro never arrived, and Flash had to do its job anyway.

Three Flashes, three jobs

3.5 Flash (May 19) was the I/O statement. Google called it frontier intelligence with action — a Flash model that outperformed Gemini 3.1 Pro on Terminal-Bench 2.1 (76.2%), GDPval-AA, and MCP Atlas, at about 4× the output speed of other frontier models. It became the default in the Gemini app, AI Mode in Search, Antigravity, and Spark. Koray Kavukcuoglu, Jeff Dean, Oriol Vinyals, and Noam Shazeer put their names on it. 3.5 Pro, they said, was already in internal use and would roll out next month.

3.6 Flash (July 21) was the unit-economics patch, 63 days later. Once 3.5 was running on billions of users and production agents, the complaint was not "make it smarter" — it was tokens, latency, and tool-call waste. Google said 3.6 used 17% fewer output tokens, took fewer reasoning steps, and cut output price from $9 to $7.50 per million. Independent tests were harsher on the headline that moves leaderboards: Artificial Analysis scored 3.6 the same as 3.5 at matched high-reasoning settings. Faster. Cheaper. Roughly 0% smarter.

3.7 Flash (August 13) was the completion-rate patch, 23 days after that. A cheap agent that loops is still expensive. Google's own copy is the tell: 3.7 "better adapts to roadblocks," "clarifies intent," and "thinks more diligently" on multi-step planning. The model card says it is based on 3.6 — algorithmic improvements to the reasoning foundation, not a new pretrain. That is how a 23-day gap is even possible.

Same family. Three constraints. Capability at I/O, cost in July, retries in August.

They could iterate this fast because Flash is the public product

Ars Technica asked the obvious question: are the 3.6-to-3.7 deltas enough to justify a new model in three weeks, or is this also about looking like Google is still shipping? Both can be true.

At I/O, Sundar Pichai told developers 3.5 Pro was coming in June. It did not. At the 3.6 launch Google said Pro was still in partner testing, with Gemini 4 pre-training already underway. Reporting after the miss — including Decrypt's summary of Bloomberg's account — said the larger model fell short internally, especially on coding, and a late-June training-data refresh disappointed.

That leaves Flash as the only Gemini Google is willing to put on a scoreboard this summer. While Pro waits, Google can:

• Keep developers on a current checkpoint instead of a two-month-old I/O model
• Collect production traces from Spark, Antigravity, and the API
• Avoid putting a delayed Pro next to GPT-5.6 Terra, Claude Sonnet 5, and Grok 4.6
• Cut list price when competitors undercut. 3.7's intro rate is $0.75 / $3.75 through December 31 — half of original 3.6 — then back to $1.50 / $7.50. OpenAI's GPT-5.6 Luna sits at $0.20 / $1.20. The promo is how Flash stays in the agent budget until January.

The cadence is not three new brains. It is one workhorse line under continuous revision because the flagship cannot ship.

The gains are where agents actually die

Google's 3.7 numbers concentrate in software engineering and multi-step workflows:

Those are Google-reported. The model card also shows Artificial Analysis Intelligence Index 56 vs 52, Terminal-bench 2.1 at 85.8% vs 78.0%, and OSWorld-2.0 at 47.9% vs 33.8%. Chart reasoning (CharXiv) slipped a little. The patch is targeted.

That targeting matches how coding agents fail. First-pass code that almost compiles is not the expensive part. The expensive part is the fifth tool call that undoes the third, the ambiguous instruction the model guessed instead of asking about, and the review comment that sounds confident about a bug that is not there.

Pull request review lives in the same failure mode. A model that retries less is more useful on a messy, multi-file diff than a model that merely streams tokens faster — which is exactly the 3.5 → 3.6 → 3.7 arc.

What this means for code review

Three Flash models in 12 weeks is good news for anyone reviewing AI-generated code, with one caveat: the interesting number is no longer tokens per second.

• Expect the workhorse to move. 3.5 was the I/O default. 3.6 optimized the meter. 3.7 optimized the retry. Your review stack should switch models as fast as Google ships them.
• Prefer models that finish the job. A 16-point DeepSWE jump in three weeks is a statement about long-horizon engineering, not chat fluency.
• Watch retries, not list price. 3.6 and 3.7 cost the same during the promo. The cheaper review is the one that does not need a second pass to walk back a hallucinated finding.
• Keep a human on the loop. Treat severity labels as input. Pair any model with repository-level context when the real bug lives outside the hunk.

That is the job ThinkReview is built for: a copilot on GitHub, GitLab, Azure DevOps, and Bitbucket, not a bot that rubber-stamps the PR. Google just showed it can revise a production coding model three times in a summer. Your review workflow should keep up.

Pick the model in Model selection. Run it on a real diff. Keep the judgment.

Details reference Gemini 3.5, 3.6 Flash, the 3.7 Flash announcement, and the 3.7 model card. Ready to review the next PR with a model that matches the work? Install ThinkReview.