← Back to feed

Google DeepMind plans to ship Gemini 4 before the end of 2026, with no benchmarks, no architecture specs, and no pricing attached to that commitment.

Speaking publicly on September 23, Google DeepMind senior vice president Koray Kavukcuoglu confirmed that Gemini 4 has entered early post-training, the fine-tuning and alignment stage that follows an initial training run, and that Google intends to release a version of it “well before the year is out.”

That is the full extent of what Google confirmed. No parameter count. No benchmark comparisons. No price sheet. No specific launch date beyond Q4 2026.

What Post-Training Means for Timeline

Post-training is not the final stage of model development, but it is typically the last major one before public release. It covers supervised fine-tuning, reinforcement learning from human feedback, and the safety and alignment work that determines where the model lands on capability evaluations.

For Gemini 4 to ship before December 31, Google would need to complete post-training within roughly three months of this announcement. That is achievable — some post-training runs take weeks — but it leaves limited margin for the extended evaluation process that Anthropic and OpenAI have applied to their most capable recent releases. Frontier labs have typically taken weeks to months between the end of training and public availability on their most capable models.

A Different Kind of Announcement

Google has typically launched frontier models with dense benchmark tables. Gemini 3.1 Pro Preview shipped with SWE-bench, LiveCodeBench Pro, and Terminal-Bench numbers. Gemini 3 arrived with detailed comparisons across reasoning, coding, and mathematics.

Gemini 4’s announcement had none of that. Kavukcuoglu’s statement was a timeline signal, not a capability claim. The absence of benchmark numbers at announcement likely means either the post-training runs that produce them are not complete, or Google made a deliberate choice to lead with availability rather than performance claims at this stage.

The approach resembles how some model labs have moved from dense technical release posts to simpler launch announcements, letting third-party evaluators like Artificial Analysis and LiveBench supply the comparative context.

Competitive Position

Gemini 4 enters post-training against a significant gap at the frontier. On LiveBench (September 2026), the current top entries are Claude Fable 5.1 at 83.4 overall, Claude Opus 5.5 at 83.2, and GPT-6 Astra at 82.2. Google’s best publicly available model on the September run sits lower in the rankings.

Whether Gemini 4 closes that gap depends entirely on what the post-training run produced — information Google has not disclosed. The internal benchmark results from the training compute would normally be available to Google’s teams at this stage. The decision not to release them publicly alongside the post-training announcement is either a timing choice or a signal that the numbers are still coming in.

Google’s ability to ship a competitive frontier model before year-end would reset the competitive picture meaningfully. A Gemini 4 that matches or exceeds Claude Opus 5.5 on reasoning and coding benchmarks would break Anthropic’s current hold on the top of the quality leaderboard.