Gemini 4 Launch Window 2026 DeepMind Post-Training Update

Published September 25, 2026 — Mountain View, Calif. Google DeepMind's Koray Kavukcuoglu said on Wednesday, September 23, at The Information's AI Agenda Live Summit that Gemini 4 has moved into early post-training. Google plans to ship the new flagship "much earlier" than year-end 2026 to close the gap with OpenAI's GPT-6 family and Anthropic's Claude 5.5 line.

Data last verified September 25, 2026 from The Information (via The Verge, Computerworld, Times of India), The Decoder, Google, and Startup Fortune.

At a glance

• Gemini 4 has entered early post-training, the refinement phase that turns a base model into a shippable product.
• DeepMind head Koray Kavukcuoglu told The Information on September 23 that Google wants to ship Gemini 4 "much earlier" than year-end.
• Gemini 4 would be Google's first flagship since the Gemini 3 series shipped in November 2025.
• No release date, API tier, price, or benchmark has been published; "much earlier" is a signal, not a committed date.

Quick Answer

Google DeepMind says Gemini 4 will launch "much earlier" than end of 2026, most likely in October or November 2026 based on the early post-training status, but as of September 25, 2026 there is no firm release date, public API tier, or pricing announced. The base model has cleared pre-training and is now in the same refinement phase OpenAI used to harden GPT-6 and Anthropic used for Claude Mythos 5.1.

What post-training actually means

Post-training is the phase that fine-tunes a base model for behaviour, safety, and tool use. It typically runs for weeks to months depending on the lab's safety bar and how many eval cycles the team runs before launch.

Why this matters for developers right now

If Gemini 4 ships with strong agentic coding and tool-use benchmarks, Vertex AI customers and AI Studio users should plan a multi-week evaluation cycle before switching any production workload. Most of the Gemini 3.6 Flash codepath already covers that test path today.

What Koray Kavukcuoglu said at the AI Agenda Live Summit

Kavukcuoglu used his first public appearance as DeepMind head to commit to an early Gemini 4 release window and to drop the AGI framing his predecessor favoured. "Our intention is to [...] as soon as possible [...] release an early post-training output because we see the results and are excited," he told The Information on September 23. He added that Google would "continue the fast-paced iterations" once the model is out (The Information, September 2026).

Kavukcuoglu took over operational leadership of DeepMind in August 2026 after Demis Hassabis stepped down as CEO. Hassabis now chairs the unit and serves as Alphabet's chief scientist, the same role Hassabis held throughout the Gemini 1 to 3 era (Startup Fortune, September 2026).

"The conversation of [whether] we achieved AGI or not is not the right conversation," Kavukcuoglu told the summit. "The conversation is more about are we able to build intelligent agents that we can trust" (The Information, September 2026). The shift dropped the existential-safety framing Hassabis used in every public appearance through 2024 and 2025.

Alphabet shares slid nearly 5% during the week of the talk on competitive pressure tied to Meta's new Muse AI hardware. The DeepMind chief's remarks landed with U.S. equity markets already volatile (Startup Fortune, September 2026).

Why Gemini 4 is still in post-training, not pre-training

Pre-training builds the raw model; post-training turns that model into one that behaves reliably for users and tools. Pre-training on the Gemini 4 dataset is already complete, which is why Sundar Pichai described the run as Google's "most ambitious pre-training run yet" in the Gemini 3.6 Flash announcement on July 21, 2026 (Google blog, July 2026).

Post-training includes supervised fine-tuning, reinforcement learning from human feedback, and tool-use alignment. Anthropic and OpenAI used the same pipeline to ship Mythos 5.1 and GPT-6 respectively.

Kavukcuoglu told the summit that the team "took a little bit of a step back" to ship faster, less powerful Flash models instead of forcing the Gemini 3.5 Pro release Pichai promised at Google I/O in May 2026. That promised model was never publicly released (The Information, September 2026).

How the Gemini 4 timeline compares to OpenAI and Anthropic

Google has not released a new flagship AI model since the Gemini 3 series shipped in November 2025, which gave OpenAI and Anthropic a free run of 2026. GPT-6 launched in March 2026, GPT-6 Astra on September 3, GPT-6 Sol and Luna on September 22. Anthropic shipped Claude Mythos 5.1 on September 1 and Claude Opus 5.5 on September 22.

Source: benchlm.ai model release log, official provider announcements, September 2026.

What the post-training phase signals about timing

OpenAI moved GPT-6 from initial post-training to general availability in roughly six weeks. If Google follows a similar pattern, Gemini 4 could appear in late October or early November 2026. No Google employee has confirmed a date.

What Gemini 4 needs to ship with at launch

Developers and enterprise buyers will judge Gemini 4 on coding benchmarks, agentic tool use, and long-context recall, not on press release claims. Pichai told Alphabet's Q2 earnings call in July 2026 that coding and agentic coding are the two areas where Google needs to close the gap with OpenAI and Anthropic (Alphabet earnings, July 2026).

Expect Google to release Gemini 4 alongside refreshed Gemini API tiers and a Vertex AI rollout. AI Studio access usually follows within days of an internal launch, depending on safety review outcomes for regulated industries.

If Google wants to match OpenAI's roughly monthly release cadence, the first post-training version will likely be a smaller Gemini 4 variant, with the full Gemini 4 Pro arriving later (The Decoder, September 2026).

How enterprise teams should plan for Gemini 4 right now

Plan for Gemini 4 without making it the default until benchmarks and Vertex AI pricing land. Lock-in risk drops sharply when a new flagship is inside a release window, and most enterprise AI teams treat "much earlier" as a heads-up, not a date.

• Delay long model lock-in contracts until Gemini 4 benchmarks are public.
• Capture a baseline by testing current Gemini 3.8 Flash workloads on Vertex AI.
• Run identical coding and reasoning tasks on Gemini 3.8 Flash, GPT-6 Sol, and Claude Opus 5.5.
• Track Gemini API pricing for Vertex AI before any production switch.

Until Google publishes prices and benchmarks, treat the "much earlier" statement as a signal that the DevDay-week news cycle will be busy (The Information, September 2026).

FAQ

Six short answers on the Gemini 4 launch window in 2026.

When is Gemini 4 launching?

Google has not set a public date. DeepMind head Koray Kavukcuoglu said on September 23, 2026 that Google wants to ship Gemini 4 "much earlier" than end of 2026. Expect an early preview first, then a wider rollout (The Information, September 2026).

Who is the new head of Google DeepMind?

Koray Kavukcuoglu, previously DeepMind's CTO, took over operational leadership in August 2026. Founder Demis Hassabis now chairs the unit and serves as Alphabet's chief scientist (Startup Fortune, September 2026).

Why did the Gemini 3.5 Pro promised in May never ship?

Pichai teased a Gemini 3.5 Pro update at Google I/O in May 2026. Kavukcuoglu told The Information the team "took a little bit of a step back" to ship faster Flash models instead. The model was never publicly released (The Information, September 2026).

What did Gemini 3 leave out that Gemini 4 will fix?

Coding and agentic tool use were the two areas Pichai named on Alphabet's Q2 call. Gemini 4 is widely expected to improve on Gemini 3 series results on benchmarks that test long-horizon agentic coding (Alphabet earnings, July 2026).

Will Gemini 4 be in Vertex AI and AI Studio on day one?

Google has not confirmed. Recent flagship launches reached Vertex AI within days of public release, but safety review can delay AI Studio access for regulated industries (Google Cloud blog, September 2026).

How does Kavukcuoglu's plan differ from Hassabis's?

Hassabis framed DeepMind around AGI and existential safety. Kavukcuoglu is steering the team toward "intelligent agents that we can trust" and faster release cycles (The Information, September 2026).

What coding benchmarks matter most for Gemini 4?

Watch SWE-bench Verified for agentic coding and MMLU-Pro plus GPQA Diamond for reasoning. A clean win on either would change the conversation in DevDay week, and OpenAI's response will follow (Stanford CRFM, 2026).

Is Gemini 4 an open-weights model like some Meta launches?

No. Google has not signalled any open-weights plan for Gemini 4. Meta's recent Muse launch is the open-weights counterweight in the September 2026 model race (Meta AI Research, September 2026).

Verify current model availability on the official Google AI and Vertex AI documentation before any production switch. This post is dated and may age within days.