As of 2026-09-03, Google has not announced a Gemini 4 release date, model ID, pricing, or benchmarks. The only statements you can treat as fact are two official ones: the 2026-07-21 Flash launch post, and Alphabet's Q2 earnings call the next day. Below: first split confirmed from unconfirmed, then why outsiders infer a year-end window, then: before the generation swap, the business JSON contract can be locked now.
"When will it be released" in the title is easy to write as a specific month. Google itself has not given a month. Treat inference as a schedule, and later it will not line up with the API catalog.
What Gemini 4 is, and whether you can use it now
Gemini 4 is the working name for Google DeepMind's next flagship base model. It has not entered the public catalog of the Gemini API, Vertex AI, or AI Studio. There is no preview ID, and no waitlist.
The only confirmed positioning is directional: a larger base than current Gemini, aimed at "the frontier when Gemini 4 actually lands," not today's leaderboard. On the earnings call, Pichai named the gaps to close as coding and agentic coding.
The production path is still 3.x: day-to-day on 3.6 Flash, high throughput on 3.5 Flash-Lite. Wait until a 4 model ID appears in the catalog before you switch the call. Don't rename fields early.
What official sources confirmed
The public record is short. Best to keep the original wording. Don't expand it into a roadmap.
| Date | Source | What the original wording says |
|---|---|---|
| 2026-07-21 | Gemini 3.6 Flash / 3.5 Flash-Lite / 3.5 Flash Cyber launch post | "We have started our most ambitious pre-training run yet, for Gemini 4" |
| Same day | Google AI Studio lead Logan Kilpatrick | Repeated the same milestone on X: pre-training has started |
| 2026-07-22 | Alphabet Q2 2026 earnings call, Sundar Pichai | Training Gemini 4; the next frontier needs a larger base; coding and agents are priorities; the Flash line is close to monthly releases |
The blog post put that sentence after 3.5 Pro's "we'll make it widely available when it's ready," as a closer, not a standalone launch. On the earnings call Pichai added two more layers: internal progress makes him confident people will be happy when they ship it; and what Google is aligning to is the frontier when Gemini 4 lands, not today's rival leaderboard.
Lock the boundary. Pre-training started ≠ a release is near. After pre-training come post-training, safety evals, partner testing, and quota. 3.5 Pro already showed: testing can sit for a long time.
What they have not said
The items below often get written up as "leaks" in search results. As of this piece, none of them have official numbers:
- Release date, quarter, or event name (I/O, Cloud Next, and a separate launch event were all not named).
- Parameter count, dense / MoE, context window, multimodal scope.
- Any benchmark scores, pricing, quota, or region.
- Whether it will ship with 4 Pro / 4 Flash / 4 Ultra from day one.
- Whether 3.5 Pro ships first, folds into 4, or gets skipped.
The November and December windows from prediction markets and tech press are bets on historical cadence, not an internal schedule leak.
Release-date predictions: why late 2026
Mark inference as inference. The usual basis is one line: Gemini generations land about every 6–9 months. Gemini 2.5 to Gemini 3 was about seven months; after 3 appeared in late 2025, the first half of 2026 was mostly 3.x Flash refreshes. Pre-training was confirmed in 2026-07. Extrapolate the old cadence, and you land on November–December 2026.
That window holds only if this "most ambitious pre-training" is not longer than before, and post-training and safety evals don't add extra delay. Google itself called this run the largest yet. Duration is more likely to run long than short.
A more stable observation point is not the month, but the catalog:
- A
gemini-4-*ID, or one with a preview suffix, appears in Gemini API / Vertex. - The next earnings call or developer event no longer only says "pre-training."
- Whether 3.5 Pro GAs first, or is replaced by a next-generation name.
Until those three happen, any "releases on this date" should be treated as a guess.
Dual track: 3.x Flash iterates monthly, 4 trains in the background
The more concrete commitment on the earnings call was not Gemini 4, but cadence: Gemini models are close to shipping "almost monthly." Combined with 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber landing together in July, Google is running dual tracks — cheap, fast Flash holds developers month by month; the larger flagship trains on a longer compute cycle.
For anyone integrating, that means two things: 3.x model IDs will keep changing; even if 4 appears at year-end, that does not mean the Flash line stops updating. Don't bind business fields to a single gemini-3.6-flash string. Bind them to a Schema.
The Gemini you can use now is not 4
1. 3.6 Flash / 3.5 Flash-Lite / Flash Cyber
In the same 2026-07-21 post, the models you can actually call are these three. Google's own numbers (launch materials, not your business acceptance line):
- 3.6 Flash: the workhorse. Versus 3.5 Flash, about 17% fewer output tokens on Artificial Analysis; DeepSWE 49% vs 37%, MLE-Bench 63.9% vs 49.7%, OSWorld-Verified 83.0% vs 78.4%. List price about $1.50 / $7.50 per million input / output tokens.
- 3.5 Flash-Lite: faster and cheaper, about 350 output tokens/s, about $0.30 / $2.50. High throughput, retrieval, and document processing go here.
- 3.5 Flash Cyber: a security-specialized model paired with CodeMender, limited to government and trusted partners — not a general API catalog item.
These scores belong to 3.6 Flash, not Gemini 4. 4 has no public leaderboard. 3.6 is already pushing coding and computer use, in the same direction Pichai set for 4 — that shows strategy continuity, not 4's scores.
2. 3.5 Pro is still in testing
Official wording: Gemini 3.5 Pro is testing with partners and will be widely available when ready. It is not 4, and it is not a flagship you can write into a production alias today. Public expectations around 3.5 Pro around I/O have already slipped more than once; coding ability was the named gap.
Until the relationship between 3.5 Pro and Gemini 4 is spelled out, don't default "the next Pro" to 4 on a schedule.
Don't confuse it with Gemma 4
Searching "Gemini 4" will mix in Gemma 4. Gemma is the open-source series you can download locally; on the same earnings call, Pichai said Gemma 4 has been downloaded more than 300 million times since its April 2026 release. That is a small model already shipping, not the closed-source flagship this piece is waiting for.
When you write monitors, changelogs, or a model registry, split family into gemini / gemma. Don't use the number 4 as a primary key.
A "facts list" that is enough to generate a Schema
Rather than write rumors into docs, fold confirmed / unconfirmed into a validatable JSON. Use YYYY-MM-DD strings for dates. Unknown fields use null, not "TBD" — later diffs and filters will be cheaper.
{
"family": "gemini",
"codename": "gemini-4",
"status": "pretraining",
"announced": "2026-07-21",
"release_date": null,
"public_model_id": null,
"confirmed": {
"pretraining_started": true,
"larger_base_model": true,
"priority_areas": ["coding", "agents"]
},
"not_confirmed": [
"parameter_count",
"context_window",
"pricing",
"benchmarks",
"tier_names"
]
}
With this site's inference rules, that sample yields:
| Field | Infer types | What to do after generate |
|---|---|---|
status / family |
string | After values stabilize, add an enum, e.g. pretraining / preview / ga |
release_date / public_model_id |
null | Write ["string", "null"] in the Schema. Don't treat it as a required date |
confirmed.priority_areas |
string[] | Confirm whether an empty array is valid; don't expand one article into a feature list |
not_confirmed |
string[] | This is a watch list, not a spec |
Once the sample is ready, open Structured Output to generate Gemini's responseSchema wrapper, then review required and additionalProperties by hand. When you move to Gemini 4, you usually only change the model ID; this list just needs a public_model_id.
When you actually run the business, the contract is still your orders, minutes, and tool I/O. 4 will not decide for you whether order_id is required.
How to accept it in JSON Toolbox
- Paste the facts list above or an existing Gemini reply into JSON Format and confirm it parses.
- Generate a first draft with Structured Output or JSON → Schema.
- Take a second set of real model output to Schema validation. The first sample is always optimistic.
- When 3.6 Flash and a later preview don't line up, use JSON Diff to see whether a field dropped or a type drifted.
- Convert to Zod when the frontend needs checks, and to OpenAPI 3.1 when you need API docs.
Everything stays in the browser and is never uploaded.
FAQ
When will Gemini 4 be released?
Google has not announced a date. 2026-07-21 only confirmed that pre-training has started. Outsiders infer November–December 2026 from historical generation gaps. That is pattern matching, not an official schedule.
Can you use Gemini 4 now?
No. There is no public model ID, API, pricing, or benchmarks. The production path is still shipping models such as 3.6 Flash and 3.5 Flash-Lite.
Are Gemini 4 and Gemma 4 the same thing?
No. Gemma 4 is the open-source small-model series released in April 2026. Gemini 4 is the next closed-source flagship and is still in pre-training.
Will Gemini 3.5 Pro ship before Gemini 4?
Google has not locked this in. Official wording only says 3.5 Pro is testing with partners and will be widely available when ready. The relationship between the two is unexplained.
Should you rewrite your Schema for Gemini 4 now?
No. First lock the business JSON with the current Gemini responseSchema / responseMimeType. A generation swap changes the model ID and wrapper, not the field contract.
Summary
Gemini 4 is confirmed in training. There is no release date. What you can write into docs: a larger base, coding and agents first, 3.x Flash keeps iterating monthly. The year-end window is an extrapolation of historical cadence. Wait for a public model ID before you change the call. Keep one core Schema. Validate, Zod, and OpenAPI all grow from it.