Gemini 4 is in post-training. Pin your model versions.

Google's next flagship model is past training and into the part where it gets shipped. Speaking at The Information's AI Agenda Live Summit this week in his first public appearance as head of Google DeepMind, Koray Kavukcuoglu said Gemini 4 has entered post-training and that he hopes to get it out "much earlier" than the end of 2026, according to 9to5Google. No date. Which is exactly why this is a planning problem and not a news item.

What actually happened

Kavukcuoglu described the intent plainly: "Our intention is to, like—as soon as possible—to release an early post-training output because we see the results and we are excited… [to] continue the fast-paced iterations." Post-training is the refinement stage — safety testing, guardrails, behavior tuning. Internal testing is already running with Gemini 4 powering Antigravity, Google's coding agent.

Context for the pace: the last major flagship was Gemini 3 Pro in November 2025, and the team dropped Gemini 3.5 Pro to push the Flash line instead. On competitive pressure from OpenAI and Anthropic, Kavukcuoglu said he has "the utmost trust in the team" and that "in my mind, it's a certainty that we are always gonna be at the frontier." The Information reported separately that Google is nearing release.

"An early post-training output" is the phrase to underline. That is not a polished GA launch. That is a preview model with a date-stamped name that behaves differently from the one before it.

Why model version pinning matters for your business

A new flagship is a behavior change, not an upgrade. Every prompt you tuned, every output parser that expects a particular shape, every agent whose tool-calling you nudged into working — all of it was fitted to a specific model. A smarter model can still break your pipeline by being verbose in a new place or refusing something it used to do. Treat a model swap like a dependency bump: staged, tested, reversible.

Name the model in config, never in code. If gemini-3-pro is hardcoded across six files, your migration is a search-and-replace across six files plus whatever you miss. One config value, one environment variable, and switching models is a deploy instead of a project. This is the cheapest thing on this list and the one most often skipped.

Preview models are for evaluation, not for load-bearing work. An early post-training release gets deprecated, renamed or superseded on Google's schedule, not yours. Run it against your eval set the week it lands. Do not put your invoicing agent on it.

Budget the re-test, not just the tokens. The real cost of a model release is a day of somebody re-running your hardest cases and reading the diffs. If you do not have that eval set stored and runnable, the cost becomes a month of noticing small wrongnesses in production. Build the harness before the model arrives — it is useful for every release after this one too.

Sources: 9to5Google, The Information.