Google Gemini 4: Pre-Training Has Officially Started — Here's What We Know

Google has confirmed that Gemini 4 is now in pre-training — the first and most compute-intensive stage of building a large language model. Logan Kilpatrick, Google's head of developer relations for AI, announced on July 21, 2026 that Google has "started our most ambitious pre-training run yet, for Gemini 4."

No specs, no benchmarks, no release date. But the timing and context tell us a lot about where Google is headed.

Why Is Google Starting Gemini 4 Now?

The announcement came alongside the launch of Gemini 3.6 Flash — and notably, without Gemini 3.5 Pro. That's the model everyone has been waiting for since Google I/O 2026 in May, when it was supposed to ship in June.

Gemini 3.5 Pro has been delayed three times due to coding performance gaps, hallucination issues, and a full architectural rebuild. According to 9to5Google, Google updated the training data in late June to improve coding skills, but the results were disappointing.

By announcing Gemini 4 pre-training while 3.5 Pro is still stuck in partner testing, Google is signaling two things:

1. They're not waiting for 3.5 Pro to fix itself. The next-generation model is already in motion.
2. Gemini 4 may be built on a completely new architecture, not an iteration of the 3.5 foundation that has caused delays.

What Has Google Actually Confirmed?

Here's the complete list of confirmed facts:

• Gemini 4 pre-training is officially underway (as of July 21, 2026)
• Google describes it as "the most ambitious pre-training run yet"
• The team is "excited by the progress"
• Gemini 3.5 Pro continues testing with partners in parallel
• No release date, benchmarks, model size, context window, or pricing have been shared

Everything beyond these points is industry speculation.

What Is Pre-Training and Why Does It Matter?

Pre-training is the foundational phase where an AI model learns patterns from massive datasets — text, code, images, audio, and video. This is where the model develops its core understanding of language, reasoning, and knowledge before any fine-tuning or safety alignment begins.

It's also the most expensive and time-consuming phase. Training a frontier model requires thousands of GPUs running continuously for months. Google's TPU infrastructure gives it a unique advantage here — the company designs its own training hardware, which can reduce costs and increase training speed compared to competitors relying on NVIDIA GPUs.

After pre-training, Google still needs to:

• Post-train the model for instruction following and helpfulness
• Fine-tune for specific tasks like coding, reasoning, and multimodal understanding
• Run safety evaluations and red-team testing
• Test with external partners (as they're currently doing with 3.5 Pro)
• Optimize inference costs before a public API launch

This pipeline typically takes 4-6 months from the start of pre-training to public availability.

When Could Gemini 4 Launch?

Google hasn't given any timeline, but we can estimate based on industry patterns:

These are rough estimates. Google could accelerate the timeline if pre-training goes well, or delay it if they encounter quality issues — as happened with Gemini 3.5 Pro.

Where Does Google Stand in the AI Race?

The Gemini 4 announcement comes at a critical moment. Here's how the major AI companies stack up in July 2026:

Google's current best production model is Gemini 3.6 Flash — a lightweight, cost-efficient model that competes with Claude Haiku and GPT-5.6 Mini, not the frontier flagships. Until 3.5 Pro ships or Gemini 4 arrives, Google is absent from the top tier of the AI leaderboard.

What Could Gemini 4 Bring?

Google hasn't revealed any technical specifications, so these are informed expectations based on competitive pressure and Google's own research publications:

• Stronger reasoning and coding — The areas where Gemini 3.5 Pro fell short
• Better multimodal understanding — Google has deep expertise in vision, audio, and video AI
• Longer context windows — Gemini models already support 1M+ tokens; Gemini 4 could push further
• Improved AI agent capabilities — Google's focus on agentic AI through Antigravity and Project Mariner
• Greater efficiency — Following the 3.6 Flash strategy of doing more with fewer tokens
• Competitive pricing — Google needs to match or beat the $10-15/M token range of Claude and GPT-5.6

What Should Developers Do Right Now?

If you're building with Google's AI stack, here's the practical advice:

1. Use Gemini 3.6 Flash for production workloads — It's available now, cost-effective, and improving
2. Don't wait for Gemini 4 — It's months away; build with what's available today
3. Design for model switching — Abstract your AI calls so you can swap models when better options arrive
4. Watch for Gemini 3.5 Pro — It could ship before Gemini 4 and fill the flagship gap in the short term

Official Announcement

For hands-on experience building with the latest AI models, check out our AI Tools for Professionals course — it covers practical workflows with Claude, ChatGPT, Gemini, and other leading AI tools. You can also explore our ChatGPT API Masterclass to learn API-level integration patterns that work across providers.

Google has laid the groundwork for its next-generation model. The questions that matter — how large, how capable, how fast, and how much — remain unanswered. But the race is on, and Gemini 4 is Google's biggest bet yet.