Gemini 4 enters post-training as the frontier race compresses into one week

Google DeepMind chief Koray Kavukcuoglu confirmed on September 24 at The Information's AI Agenda Live Summit that Gemini 4 has entered post-training, with the company aiming to ship "much earlier" than a year-end target. The confirmation lands 48 hours after Anthropic and OpenAI each cut frontier prices by 40–50% on the same afternoon, and days after xAI's Grok 4.7 shipped at $2/$6 input/output per million tokens. Four labs, one week, one capability tier.

Kavukcuoglu's framing was less roadmap than urgency signal. "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," he told the room. No specific date, per AI Weekly, which added the necessary caution: "This is a statement of intent, not a shipped product; if the 'early' Gemini 4 slips past year-end, the narrative flips from Google closing on OpenAI and Anthropic to Google still trailing them."

The trailing is real. On Intelligence Index v4.3.2, per Forkast, Gemini 3.6 Flash sits at 34.0 against Anthropic's Opus 5.5 at 57.6, GPT-6 Astra at 52.7, and GPT-6 Sol at 47.5. A 24-point gap between Google's shipping frontier model and the leader isn't the kind of deficit closed by a Flash refresh at the same $0.75/M price. It's why Gemini 3.5 Pro has been stuck in partner testing since June without shipping, and why Kavukcuoglu is signaling early instead of waiting.

The pre-training run, announced July 21, took roughly two months. Antigravity, an internal Google test product, is already running on Gemini 4.

The economic story underneath the horse race matters more than any single benchmark. OpenAI cut GPT-6 Sol pricing 50% versus GPT-5.6; Anthropic's Dianne Penn confirmed Claude Opus 5.5 lands around 40% below Opus 5; xAI's Grok 4.7 held the same $2/$6 line while raising capability. A fourth entrant is now visibly loading into that same tier, from an Alphabet subsidiary that spent August reorganizing.

For small-business owners, the practical read-through isn't which model wins. It's that capable AI inference has a permanently lower cost floor. Continuously-running services that research a company, scan lead sources, and prepare outreach, the economics of LemonLime at $999/month, depend on exactly this compression. Four labs undercutting each other every week is the supply side of that math.

Gemini 4 shipping on schedule would confirm the pattern. Slipping past year-end would confirm a different one.

Sources

• https://9to5google.com/2026/09/24/google-says-gemini-4-release-is-coming-as-soon-as-possible/
• https://www.cnbc.com/2026/09/22/anthropic-openai-cheaper-ai-models.html
• https://aiweekly.co/alerts/deepminds-kavukcuoglu-targets-pre-year-end-gemini-4-ship
• https://forkast.news/googles-gemini-4-enters-post-training-and-the-three-way-frontier-race-just-compressed/
• https://startupfortune.com/google-deepminds-new-chief-says-gemini-4-could-ship-much-sooner-than-planned/