Developers spotted checkpoints on LMArena they believe are an early Gemini 4 Pro build, and flooded X with 3D physics and code tests. Google has confirmed nothing. The only official word came on 24 September: Koray Kavukcuoglu, head of Google DeepMind, told The Information that Gemini 4 is in early post-training.
Key takeaways
- Kavukcuoglu has run Google DeepMind since August 2026, after Demis Hassabis stepped down as CEO
- Officially confirmed: Gemini 4 in post-training, release “much earlier” than the end of 2026
- Unconfirmed: arena checkpoints, one visible as “gemini-3.8-flash”, codename “barium-b”
- Every circulating 3D demo comes from X posts, not Google material
- Claude Opus 5.5 and GPT-6 launched 22 September, a week before the leak wave
What Google actually said
Kavukcuoglu, previously DeepMind's CTO, publicly addressed the Gemini 4 timeline for the first time.
Our intention is to, as soon as possible, release an early post-training output.
Koray Kavukcuoglu, head of Google DeepMind, speaking to The Information.
Post-training tunes a base model for predictable behaviour, and safety testing continues. Kavukcuoglu confirmed that Google engineers already use Gemini 4 internally in Antigravity, and that new models shape the design of the “next two or three generations” of TPU chips.
On Gemini 3.5 Pro, teased in summer 2026 and never released, he said the company “took a little bit of a step back” and prioritised Flash models.
What Google did not confirm
The rest is community observation. A model labelled “gemini-3.8-flash” appeared on LMSYS Chatbot Arena and, testers say, far outperformed the production Gemini 3.8 Flash. 36Kr describes that checkpoint?Checkpoint: A saved snapshot of a model taken during training, tested before any production release. as “barium-b” or “Checkpoint 2”.
The demos, from a floatplane takeoff compared against Opus 5.5 to an interactive ISS model in Three.js, are one-off X clips with no repeatable methodology. The same applies to leaked specs, including a roughly 10M-token context window and aggressive pricing. None of it is confirmed by Google.
| Officially confirmed | Unconfirmed leak |
|---|---|
| Gemini 4 is in early post-training | Arena checkpoints taken to be an early Gemini 4 Pro |
| Release “much earlier” than the end of 2026 | Codename “barium-b” and the “Checkpoint 2” label |
| Google engineers use the model internally in Antigravity | 3D physics and code-generation demos from X posts |
| New models shape the design of 2–3 further TPU generations | A roughly 10M-token context window and aggressive pricing |
| On Gemini 3.5 Pro the company “took a step back” in favour of Flash models | “gemini-3.8-flash” outperforming the production Gemini 3.8 Flash |
Why it matters
Google has not shipped a flagship since Gemini 3, while Anthropic and OpenAI shipped theirs on 22 September. A public timeline from the new DeepMind chief is a signal aimed at the market and investors, not developers.
At the same time the narrative around the model is being written entirely outside the company, on anonymous arena checkpoints. When the launch comes, Google will be answering expectations it never set.
What next?
- An early post-training build of Gemini 4 “much earlier” than the end of 2026, no date given
- Gemini 3.5 Pro's status remains open, with no formal cancellation announced
- Leaked specs verifiable only once the official model card and API pricing appear
Sources
- The Information — Google Nears Release of Flagship Gemini 4 AI Model
- 9to5Google — Google says Gemini 4 release is coming 'as soon as possible'
- The Decoder — Deepmind was built to chase AGI, but its new chief just wants Gemini 4 out the door
- Nokiapoweruser — Google's Secret Gemini 4 Pro Test Leaks: Mind-Blowing 3D Code and Benchmarks Surface
- 36Kr — 刚刚，Gemini 4 Pro全新曝光，实测碾压Opus 5.5
