Breaking

Google DeepMind Says Gemini 4 Is in Post-Training as It Targets an Earlier Release

September 25, 2026 at 10:30 EDT

Foundation Models
AI Agents

Google is moving its next flagship AI model, Gemini 4, into the final refinement stage, with Google DeepMind chief Koray Kavukcuoglu saying the company wants to release an early post-training version as soon as possible and well before year-end. The update signals that Google is trying to accelerate its frontier-model cadence as Anthropic and OpenAI continue to compete aggressively on coding, agentic workflows and enterprise AI.

Sep. 2026 · Google DeepMind

Gemini 4 enters refinement

Google says Gemini 4 is in early post-training and is being prepared for an earlier release, as frontier-model competition increasingly centers on coding, agents and multimodal work.

$0.75

Gemini 3.8 Flash input price per 1M tokens during the introductory period

73.7%

Gemini 3.8 Flash on DeepSWE v1.1 long-horizon software engineering

89.4%

Gemini 3.8 Flash on Terminal-bench 2.1 agentic terminal coding

Published frontier-model coding scores

73.7%

Gemini 3.8 Flash

74.0%

Claude Opus 5

72.7%

GPT-5.6 Sol

DeepSWE v1.1 scores are tightly clustered, underscoring why Gemini 4's post-training phase matters.

What developers want

An earlier preview could help teams test coding, tool use, document work and multimodal workflows sooner.

What remains uncertain

Early post-training can still change reliability, safety behavior and final benchmark performance.

Release path

Pre-training run→Early post-training→Early release target

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.