Gemini 4 release date: the model has four of them, and only one has happened

Announced on September 30, open to almost nobody on October 4

Google announced Gemini 4 Argon on September 30, 2026, and that announcement is the only milestone the model has reached. Four days later, no public route to it exists. Google has published no API model ID, no model card and no knowledge cutoff, and a check of the public Gemini API catalog on October 1 found no Argon entry at all.

None of that is an accident or a failure. Google describes the rollout as staged for safety reasons, beginning with vetted cyber defenders in the Fairwind program. The practical consequence is the part worth keeping. Every number, every price and every capability claim about Gemini 4 comes from Google or from a partner, and nobody outside that group can reproduce any of it yet.

The facts in this article come from Google's Gemini 4 Argon announcement of September 30, 2026, from coverage published between September 30 and October 2, 2026, and from a check of Google's public API documentation on October 1, 2026.

A release date is four different events

Treating a launch as one event is where most of the confusion starts. A model can be announced, then appear in the consumer app, then be listed in the API catalog, and only then be reachable through a route you control. Those are four separate steps, and each one has its own proof.

EvoLink framed the distinction sharply when it noted that these are sequential states rather than conflicting reports. A model can be announced before the app receives it, reach the app before the API lists it, and be listed upstream before any gateway has verified a route. The table below is worth keeping, because it tells you what to look for instead of what to believe.

An announcement is therefore a real fact and an unusable one at the same time. It tells you where Google is going and nothing about what you can run this afternoon.

Ten weeks from one sentence to a launch

The road to September 30 began with a single line buried inside a smaller model's release note. On July 21, 2026, in the announcement of Gemini 3.6 Flash, Google wrote that it had started its most ambitious pre-training run yet, for Gemini 4. Logan Kilpatrick, who leads Google AI Studio, repeated the confirmation the same day.

Sundar Pichai extended it the next day on Alphabet's quarterly earnings call, describing Gemini 4 as a significantly larger frontier model and naming coding and agentic coding as the areas Google had to close. He also described a shift toward an almost monthly release cadence, with Gemini 4 as the baseline.

Reports of a stage change arrived in mid-September from unnamed sources. On September 24, Koray Kavukcuoglu said Gemini 4 had entered post-training and that he hoped for a release much earlier than the end of the year. Six days later the model shipped. The table keeps the milestones and, more usefully, what each statement left out.

• The first official mention of Gemini 4 appeared inside another model's release note.
• Ten weeks separate that sentence from the launch.
• The post-training statement on September 24 was followed by a launch six days later.

The rumors, checked against what shipped

Four things filled the gap between July and September, and all four are testable now. Checking them is the quickest way to calibrate how much of the next launch claim to trust.

The first was the description of the model. Through July the accurate statement was that Gemini 4 had no release date, no benchmarks and no price. That was correct then and wrong by September 30, when all three arrived at once.

The second was timing. The expectation of a late-2026 arrival came from reading Google's historical training cycles rather than from anything the company committed to, and the model landed roughly a quarter earlier.

The third and fourth concern scope and verification. A product name is not a request model ID, and a vendor table is not an independent result, which is exactly why the same Terminal-bench row produced two different leaders on the same day.

Read the fourth row beside the other three. Being early on a date is a harmless miss, while mistaking an announcement for an open model is the error that costs a planning cycle.

Who gets in, and in what order

Access runs in three stages and only the first has begun. Vetted cyber defenders arrive through the Fairwind program, and Google says internal teams receive a version with some cyber guardrails relaxed so they can reach the full defensive capability.

Paid API customers and Google AI Ultra subscribers are named as the next group, with no date attached. Developers, enterprises and consumers come last, promised as soon as possible and again left without a date. Nothing Google has published covers the free tier of the Gemini app or Google AI Pro.

The gate is deliberate rather than accidental. Google ties it to guardrail and monitoring work and to a voluntary pre-deployment review process with the United States government, which is the same reason the flagship reached a defender's desk before it reached yours.

How to check whether it has reached you

Because the announcement and the release are separate events, the only reliable test is your own account rather than any article about it, this one included. Four places carry a real confirmation, and each is a page you can open instead of a claim you have to weigh.

1. Open the Gemini API model list and search for a Gemini 4 entry that has a model ID, a rate limit and a changelog line dated after September 30, 2026.
2. Check the Vertex AI model garden for the same model under the same identifier.
3. Look through the Gemini app's own model picker for a Gemini 4 option you can actually select, not a label that only appears in screenshots.
4. Check the model list in your own workspace. On this site it still reads Gemini 3.8 Flash, which is the honest answer for this site.

• A keynote demo, an earnings-call phrase or a prediction article is not a confirmation.
• An anonymous arena alias, or an unlabelled model in an A/B test, is not a confirmation either.
• A gateway placeholder page carrying the model's name is not a confirmation on its own.

The same standard cuts both ways. A screenshot of a model answering is not proof that anyone can call it, and an empty catalog entry is a stronger signal than a rumour that access is coming.

What to use while the gate is closed

The generation that is actually callable did not stand still. Gemini 3.8 Live and 3.8 Live Extended Thinking arrived on September 15, 2026, Gemini 3.8 Flash is the newest Flash model, and Gemini 3.1 Pro still carries its preview suffix.

The wait works in your favour in one specific way. A Gemini 3.8 baseline built now is a fixed, reproducible record of what your prompts and your pipeline produce, and rerunning exactly that set on the day Gemini 4 opens is the only comparison that measures your own work rather than someone else's table.

One warning belongs here too. Gemini 3.5 Pro was expected in June 2026 and remains in partner testing, which is a reminder that a delayed model and an announced model can leave you waiting on the same morning.