Google Confirms Gemini 4 Release Is Coming Soon

Google has said the Gemini 4 release is coming, but not when. If you own a delivery plan, that is the worst kind of news: real enough to worry about, vague enough that you can't schedule it.

What Google actually said about the Gemini 4 release

Koray Kavukcuoglu, who leads Google DeepMind, spoke at The Information's AI Agenda Live Summit in the week of 24 September 2026. 9to5Google's report says the model has entered post-training. That is the stage after the base model is trained, where it gets tuned, safety tested, given guardrails and tried out internally, including in Google's Antigravity tooling.

His intention, in his words, was to release "an early post-training output" "as soon as possible" because Google likes the results. He hopes the launch lands "much earlier" than the end of 2026.

That is the whole confirmed picture. Here is what has not been announced:

• A release date or even a month.
• Pricing.
• Context window, modalities or benchmark numbers.
• Which surfaces get it first (the Gemini API, Vertex AI, the Gemini app).

Note the wording "early post-training output". It suggests a first release may be a preview rather than a finished, stable model. That is my reading, not a Google statement, so don't plan around it either way.

What you can build on today

The DeepMind models page does not list Gemini 4. It currently lists Gemini 3.8 Flash, described as best for complex agentic tasks at scale, plus variants such as Gemini 3.8 Live. It also shows Gemini Omni 1.1 Flash, Gemini Robotics 2 and Gemini Embedding. Gemini 3.8 Flash was Google's most recent model release, in September 2026.

I found no pricing for Gemini 4 and did not verify current Gemini 3.8 Flash prices, so check the official pricing page before you put numbers in a budget.

How to plan around an unannounced model

Treat Gemini 4 as an option on your roadmap, not a dependency. Three habits keep that cheap.

Keep the model name out of your code

If swapping models means a code change and a release, you have made every launch a project. Put the model ID in config:

# llm-config.yaml
llm:
  provider: google
  model: <current-approved-model-id>   # set from the vendor docs, not from memory
  fallback_model: <previous-model-id>
  eval_suite: evals/regression-v3
  promote_only_if_eval_score_gte: 0.92  # your own threshold

The placeholders are yours to fill. The point is that a model change becomes a config change gated by your own eval score.

Write the evaluation before the model exists

A regression suite built from your real customer inputs tells you within a day whether a new model is better for you. Without one, the debate becomes opinions and launch-day benchmarks. Build it now, while nothing is on fire.

Budget for both outcomes

Give the roadmap two scenarios: stay on the current model, or migrate. For the migration scenario, reserve a fixed slice of capacity (a few days of eval, prompt retuning and rollout) rather than a launch date. Pricing is unknown, so state cost as a range and mark it as an assumption to revisit on announcement day.

What to do on Monday morning

1. Ask engineering where the model ID is set today, and move it to config if it is hardcoded.
2. Confirm an eval suite exists, and add ten real failure cases if it is thin.
3. Add a line to the risk register: "Gemini 4 timing unannounced, no commitments depend on it."
4. Do not promise a customer a feature that needs Gemini 4.
5. Set a review trigger: when Google publishes a model card and pricing, run the eval and decide.

And when a dated announcement does arrive, remember that a preview label usually means limits and changes ahead. Check the release notes before you move production traffic.

A confirmed direction without a date is enough to prepare for, and not enough to depend on: teams that isolate the model behind config and evals can adopt Gemini 4 in days, while everyone else will be replanning.