What shipped: Gemini 3.7 Flash, generally available 13 August 2026, three weeks after Gemini 3.6 Flash. Google calls it "our most intelligent workhorse model yet for coding and agents," with real gains on coding and agentic benchmarks.
The price cut is temporary. Introductory pricing of $0.75 per million input tokens / $3.75 per million output tokens — half of 3.6 Flash's launch price — runs only through 31 December 2026. From 1 January 2027 it becomes $1.50 / $7.50, doubling.
Where it lives today: Gemini Spark (Google AI Pro/Ultra subscribers), the Gemini API via AI Studio, Google Antigravity, Android Studio, Gemini Enterprise, and GitHub Copilot from day one. It also replaced 3.6 Flash as the default in the consumer Gemini app.
Where it does not (yet) live: Google's own materials don't mention Gemini Notebook anywhere. If the model behind your free-tier notebook chats matters to you, that hasn't been confirmed to have changed.
The numbers Google is leading with
From Google's own model announcement, the headline benchmark gains over 3.6 Flash:
Benchmark 3.6 Flash 3.7 Flash
------------------------------------------------
FrontierCode 1.1 Main 34.4% 43.6%
DeepSWE v1.1 49.0% 65.3%
WebDev Arena (Elo) 1538 1588
GDP.pdf (knowledge work) 22.0% 34.0%
AutomationBench 17.0% 30.4%
GDP.pdf is a document-comprehension benchmark across
finance, law and bioscience material. Google's own
figures also put 3.7 Flash ahead of Claude Sonnet 5
(28.0%) and GPT-5.6 Terra (24.7%) on the same test —
Google's benchmark, not independently reproduced.That GDP.pdf result is worth pausing on, because it's the one most relevant to this site's readers even though nothing here directly uses this model: it's a test of reading dense professional documents accurately, and Google's claimed 34% beats both Claude Sonnet 5 and GPT-5.6 Terra on the same benchmark. Treat it as a vendor's own number, not a neutral one — nobody outside Google has reproduced it yet.
The model card lists a 1-million-token input context with up to 64,000 output tokens, a knowledge cutoff of March 2026, and text, image, audio and video input (text-only output). Independent speed testing from Artificial Analysis clocked it at roughly 340 tokens per second, which they list as the fastest model they currently measure.
The price cut, and the part of it that expires
Google's introductory pricing is real: $0.75 per million input tokens and $3.75 per million output tokens, half of what 3.6 Flash launched at. It also has an expiry date that got far less coverage than the cut itself. From 1 January 2027, the price becomes $1.50 and $7.50 — back to roughly where 3.6 Flash started. If you're building something on this API for the long term, budget for the January price, not the launch price.
There's a second wrinkle beyond the sticker price. Independent reviewer eesel AI, testing the model for support and automation workloads, found that 3.7 Flash dropped the "minimal" reasoning tier that made 3.6 Flash usable for cheap, high-volume classification tasks — the new floor is "low," which thinks more by default. Thinking tokens also bill at the output rate even though the API only returns a summary of them; eesel's own sample showed 297 thinking tokens against 171 returned output tokens on one task. Their estimate is that heavier reasoning overhead adds roughly 40% more billed tokens per task versus 3.6 Flash, which erodes a meaningful share of the headline price cut for exactly the workloads Flash used to be the cheap option for. Their verdict: "If you picked Flash originally because it was the cheap tier, 3.7 is not your upgrade path."
Where it actually ships
Confirmed availability as of launch day: Gemini Spark, the agentic assistant inside the Gemini app, for Google AI Pro and Ultra subscribers in more than 160 countries; the Gemini API through AI Studio; Google Antigravity; Android Studio; the Gemini Enterprise Agent Platform and Gemini Enterprise app; and GitHub Copilot, live the same day across VS Code, Visual Studio, the Copilot CLI, JetBrains, Xcode and Eclipse for Pro, Pro+, Max, Business and Enterprise tiers (Business and Enterprise admins have to enable it). In the consumer Gemini app itself, 3.7 Flash replaced 3.6 Flash as the model-picker default across Android, iOS, web and macOS starting the same afternoon.
Sundar Pichai's own framing, posted the day of launch: "Our Flash models are workhorses that offer performance at a great price. So, we're shipping updates fast to get them in developers' hands. And now just 3 weeks after launching 3.6 Flash, 3.7 Flash shows significant gains including in coding and agentic work."
What it does not confirm: Gemini Notebook
This is worth stating plainly rather than assuming continuity, because it's the question our readers will actually have. Nothing in Google's announcement, the model card, or the two dedicated write-ups on this launch mentions Gemini Notebook. The confirmed consumer-facing footprint is Gemini Spark and the general Gemini app model picker — a separate surface from Gemini Notebook, which has its own model behind the scenes and its own release cadence. If a faster, cheaper Flash model reaching your free-tier notebook chats or Audio Overviews would matter to you, that hasn't happened yet, and Google hasn't said when or whether it will.
Why the cheap model is shipping fast while the flagship stalls
The three-week gap between 3.6 and 3.7 Flash isn't happening in isolation. Gemini 3.5 Pro — the flagship Google promised "next month" at I/O on 19 May 2026 — is now in its fourth consecutive month of delay, and on Google's most recent earnings call Pichai reportedly sidestepped direct questions about its release.
That delay is unfolding alongside a real leadership shakeup at DeepMind. On 5 August 2026, Demis Hassabis stepped down as DeepMind's CEO to become Alphabet's chief scientist and DeepMind's chairman, citing more time for Isomorphic Labs, his drug-discovery venture; Koray Kavukcuoglu took over day-to-day operations as SVP. Multiple senior figures, including longtime Google engineer Jeff Dean, also moved roles or departed around the same period, and an all-hands the following day revealed some teams moving out of DeepMind into corporate Google. Alphabet's stock dropped roughly 4% the day the changes were announced.
None of that is confirmed as the cause of the Pro delay — Google hasn't said so directly. But the pattern is consistent: the model Google can ship fast and cheap keeps shipping fast and cheap, while the one that needs to be the best in the world keeps slipping, during the same months its research organization is being restructured.
People also ask
Is Gemini 3.7 Flash actually cheaper than 3.6 Flash?
Yes, but only through 31 December 2026. Introductory pricing is $0.75/$3.75 per million input/output tokens, half of 3.6 Flash's launch price. From 1 January 2027 it rises to $1.50/$7.50 — back near where 3.6 Flash started, not a permanent cut.
Does Gemini 3.7 Flash power Gemini Notebook?
Not confirmed. Google's launch materials list Gemini Spark, the Gemini API, Antigravity, Android Studio, Gemini Enterprise and GitHub Copilot as where it ships. Gemini Notebook isn't mentioned anywhere in the announcement.
Is Gemini 3.7 Flash better than Claude or GPT for coding?
On Google's own benchmarks, yes, by a real margin on FrontierCode and DeepSWE. Those are Google's figures, not independently reproduced, and at least one independent reviewer (eesel AI) found that once you account for 3.7 Flash's heavier reasoning-token overhead, GPT-5.6 Terra still wins on long-horizon agentic coding tasks specifically.
Why did Gemini 3.7 Flash ship so soon after 3.6 Flash?
Google frames it as an accelerated update cadence for its Flash line. It's also shipping while Gemini 3.5 Pro, the flagship model promised for release in mid-2026, remains delayed into its fourth month past that promise, during a period of leadership changes at DeepMind — Google hasn't linked the two directly.
Is 3.7 Flash still the cheap option for high-volume tasks?
Less clearly than 3.6 Flash was. It dropped the "minimal" reasoning tier that made 3.6 Flash viable for cheap classification-style work, and thinking tokens bill at the output rate. One independent review estimated roughly 40% more billed tokens per task as a result, which eats into the headline price cut for exactly that use case.