Google's next flagship broke cover this month, and the more interesting story is the model that will never ship.

The first internal checkpoint for what people are calling Gemini 4 Pro has reportedly appeared, internally codenamed argon, with a leaked first output that took 2.4 minutes on high thinking effort. Community trackers expect a public release in October, with a Flash-Lite model and an updated image model arriving first.

Meanwhile Gemini 3.5 Pro, the flagship Sundar Pichai promised at I/O in May and said would follow the next month, appears to have been shelved. Reports say Google canceled it outright and redirected the compute into Gemini 4.

So Google spent four months missing deadlines on a model it has now apparently abandoned, in order to bet everything on the one after it. Here is what is actually known, what is extrapolated, and the naming trap that makes most Gemini 4 coverage unreliable.

Strip out the leaks and Google's official position is short.

Google confirmed in July that the Gemini 4 pre-training run is underway and that it will be a significantly larger model. On the Q2 earnings call, Pichai added that Gemini 4 requires a substantially larger base model to keep pace with the rest of the frontier field, and that the company is prioritizing coding and autonomous agent capability.

That is the entirety of it. No release date, no quarter, no month, no launch channel, and critically, no confirmation of any variant names.

Labs usually stay quiet until a model is close, so disclosing a pre-training run reads as a signal to investors and competitors that Google is scaling hard after a rough stretch. The emphasis on coding is not accidental either. Gemini 3.5 Pro's slip was reportedly driven by coding performance falling short of internal targets after a late-June training-data update, and Bloomberg reported as much at the time. A next-generation model that is not competitive on coding would be a serious problem for Google's position against Anthropic and OpenAI.

Here is the part worth sitting with.

Google announced the Gemini 3.5 series at I/O in May 2026 and said the Pro flagship would follow the next month. It never shipped. Google's public line became that 3.5 Pro was testing with partners and would launch as soon as it was ready.

Community reporting now says it was quietly canceled, with all compute redirected into Gemini 4. If accurate, that means the flagship Google promised on stage in May simply will not exist, and the company spent the summer shipping Flash-tier models while its Pro line sat empty.

It also explains the version numbering, which has confused everyone. Google's current lineup includes Gemini 3.1 Pro, 3 Deep Think, 3.8 Flash released September 2, and 3.5 Flash-Lite. A 3.8 Flash exists while a 3.5 Pro does not. That is what a shelved flagship looks like from the outside: the Flash line kept iterating past it.

Promoting the next generation before the current flagship reached general availability is exactly what unsettled parts of the developer community, and it is why the Gemini 4 tease is being read with more caution than a normal pre-announcement would get.

This is where the reporting gets unreliable, and it is worth knowing before you read any spec table.

The name is not confirmed. Tracker pages already exist for Gemini 4 Pro and Gemini 4 Flash-Lite. Those names are extrapolated from how the 2.5 and 3.x lineups were structured. Google has not confirmed a single Gemini 4 variant.

The leaked specs cite no Google source. The figures circulating, a 2-million-token context window, an adaptive high-effort reasoning core, and evaluations showing it beating Fable 5.1 on coding and GPT-6 Astra on multi-step reasoning, come from community posts. Copying a number does not turn a guess into a fact.

And there is a genuine naming trap. Gemma 4 is Google's open-weight model family, released in April 2026, downloadable and runnable on your own hardware. Gemini 4 is the unreleased proprietary frontier model. Benchmark numbers and specs that surface under a Gemini 4 search frequently belong to Gemma 4 instead. A Hacker News author had to publicly correct a post that confused the two.

There is also a useful framework buried in that coverage. For a team planning an evaluation, three separate gates have to open: the announcement, availability in the Gemini app, and a model ID appearing in the Gemini API or Vertex AI catalog. A model can be announced before the app gets it, and reach the app before the API lists it. All three gates are currently closed.

The dates worth holding loosely are October from community trackers and November or December from 9to5Google's read of past release patterns. Google has attached no date to any of it, so treat every month you see as an educated guess until a model ID shows up in the API catalog.

What you can actually use today is the current generation, and it is decent. Gemini 3.6 Flash runs $1.50 and $7.50 per million tokens with solid coding and agentic scores, and 3.7 Flash launched at half price through year-end. The Flash line has been Google's real product all year, which is the quiet admission underneath this whole story.

The strategic read is the interesting part. Google is the only frontier lab that spent 2026 without a competitive flagship, shipping Flash models every few weeks while its Pro tier went stale and then apparently got canceled. Gemini 4 is the recovery attempt, and Pichai has publicly tied it to needing a substantially larger base model to keep pace, which is an unusually direct admission that the current generation did not.

If it lands in October and the leaked coding results hold, Google is back in the race. If it slips the way 3.5 Pro did, the company will have gone more than a year without a flagship anyone chose on merit.

The stories moving fast beyond today's headlines:

Google confirmed in July that Gemini 4's pre-training run is underway and that it needs a substantially larger base model to keep pace with the frontier, with coding and autonomous agents as the priority. That is all Google has said. No date, no quarter, no launch channel, and no confirmed variant names.

The leaks say more. A first internal checkpoint reportedly appeared under the codename argon, with a first output taking 2.4 minutes on high thinking effort, a 2-million-token context, and evaluations beating Fable 5.1 on coding and GPT-6 Astra on multi-step reasoning. Community trackers expect October. None of it cites a Google source.

The bigger story is Gemini 3.5 Pro. Pichai announced it at I/O in May and said the Pro flagship would follow the next month. It never shipped, Bloomberg reported coding performance drove the slip, and reports now say Google canceled it entirely and redirected the compute into Gemini 4. The version numbers give it away: there is a 3.8 Flash but no 3.5 Pro.

Two traps to avoid. The names Gemini 4 Pro and Gemini 4 Flash-Lite are extrapolated from previous lineups rather than confirmed. And Gemma 4, Google's open-weight family from April, is a different product whose specs frequently get reported as Gemini 4's.

Watch the Vertex AI catalog rather than the rumors. A model ID appearing there is the only thing that settles it.

Stay building. 🤖