Google DeepMind chief says Gemini 4 coming much sooner

2026-09-25 - ABikram Mondal

Google signals faster Gemini 4 timeline

Koray Kavukcuoglu, the new head of Google DeepMind, used his first public remarks to state that Gemini 4 has entered early post-training. The company wants an initial version out as soon as possible because internal results look strong. He spoke at The Information AI Agenda Live Summit on 24 September 2026.

Google last shipped a flagship model in the Gemini 3 series back in November 2025. Rivals moved ahead with GPT-6 from OpenAI and the Mythos line from Anthropic. Those releases set higher bars on several tasks.

Kavukcuoglu said the team plans to ship an early post-training output quickly and then keep iterating at a fast pace. He added that the goal is a launch much earlier than the end of 2026. No exact month or benchmark numbers were given.

The comments follow a leadership change. Demis Hassabis stepped down in August after noting progress on the model. Kavukcuoglu now leads daily operations and stressed trust in the team to stay at the frontier.

Earlier this year Google skipped a planned Gemini 3.5 Pro update that had been slated for June. Resources went instead to faster Flash variants. Training on Gemini 4 itself began as the most ambitious pre-training run yet, according to CEO Sundar Pichai in July.

What post-training means here

Post-training refines the base model for safer and more reliable behavior before public release. It covers alignment work, safety checks, and capability tuning. Google has not shared details on parameter count, context length, or training data sources for Gemini 4.

The approach of releasing an early version and iterating fast differs from Google past practice of polished big-bang launches. Kavukcuoglu noted excitement over current results and the intent to continue rapid updates afterward.

Competitors have adopted quicker cycles. OpenAI and Anthropic both pushed new models in recent weeks with price reductions on some variants. Google faces pressure to match that speed while closing performance gaps.

Investors reacted to the broader competitive picture. Alphabet shares fell nearly 5 percent earlier in the week on concerns tied to Meta hardware moves. The Gemini 4 update offers one data point on Google response.

No pricing, API access details, or specific capability claims have been released yet. Those will come closer to actual availability.

Context on recent Google releases

Gemini 3 Pro arrived in November 2025. A Gemini 3.1 Pro update followed in February 2026. Several Flash and smaller models filled the gap, including Gemini 3.8 Flash and Gemini 3.8 Live for voice agents.

At I/O in May the company highlighted Gemini as the layer for daily digital life. That positioning puts it directly against OpenAI and Anthropic offerings in agentic and multimodal work.

The skipped 3.5 Pro step showed internal trade-offs. The team chose quicker, lighter models over one larger update. Focus has now shifted back to the next full flagship.

Kavukcuoglu acknowledged the pause on 3.5 work. He said the team stepped back to prioritize Flash models before returning full attention to Gemini 4.

Who should watch this closely

Developers building on Google Cloud or Vertex AI will want early access when it arrives. The model could affect agent workflows, coding assistance, and multimodal tasks once benchmarks appear.

Enterprises comparing frontier models for cost and performance will track release timing and pricing. Google has not confirmed whether early versions will carry lower rates than full releases.

Competitors at OpenAI and Anthropic already ship frequent updates. A quicker Gemini 4 cadence could force further acceleration across the board.

Indian startups and researchers using Gemini via Google tools should note the timeline. Earlier availability means sooner integration testing for local applications.

General users of Gemini in search or Android will see incremental improvements first through smaller models. The flagship jump will matter more for heavy workloads.

What remains unknown

Specific numbers on context window, tool use, or benchmark scores are absent. Google has not confirmed whether Gemini 4 will include new architecture changes or simply scale previous work.

Safety testing details and any new alignment techniques have not been disclosed. Those usually appear in technical reports close to launch.

Availability on consumer apps versus API only is also unclear. Past flagships rolled out gradually across products.

The exact definition of much earlier than year end stays open. It could mean October or November or a December drop. Follow-up statements will narrow that.

ABikram Mondal builds automation for exactly this kind of problem at https://abikrammondal.com/services/automation.

Bottom line on timing

Google is signaling intent to close the release gap with rivals. The language of early post-training output and fast iterations shows a shift toward quicker shipping.

Whether the model actually delivers competitive gains will depend on numbers that are still weeks or months away. The announcement itself is the clearest public signal yet on the new timeline.

Watch for further updates from DeepMind or Google I/O style events. Those will likely include the first concrete details on capabilities and access.

Until then the story remains one of stated direction rather than delivered product. The next 30 to 60 days should clarify how much earlier the launch actually lands.

The short version. Google aims to release an early Gemini 4 version well before the end of 2026 after entering post-training.

Sources

Reported from the sources above on 2026-09-25. Figures are as published at the time of writing. If something here has moved on, the linked source is the one to trust.

From the desk of ABikram Mondal

If you got here because you are actually thinking about putting models like this to work inside a real business, wired into the tools a team already uses, that is the work I do. I build for founders and small teams who want the thing to exist and work, not a deck about it.

Automation services · everything I do · talk to me