Meet Gemini 4, Google's Most Ambitious Pre-Training Run Yet

Gemini 4 is Google DeepMind's next-generation frontier large language model, confirmed to be in pre-training as of July 21, 2026 and described by Google as its "most ambitious pre-training run yet." The model has not been released, has no published benchmarks, no pricing, no context-window figure, and no confirmed launch date. What is known is narrow but official: pre-training is underway, CEO Sundar Pichai has called it "significantly larger" than any prior Gemini, and coding and autonomous agents are stated priorities.

Key Takeaways

• Google officially confirmed Gemini 4 pre-training on July 21, 2026, buried in the Gemini 3.6 Flash launch post on 9to5Google.
• The exact wording from the Gemini team: "our most ambitious pre-training run yet, for Gemini 4."
• Sundar Pichai told analysts on the July 23 earnings call that Gemini 4 is a "significantly larger" frontier model, per InfoWorld's coverage.
• Coding and autonomous agents are stated priorities for the Gemini 4 generation.
• Gemini 3.5 Pro remains stuck in partner testing after missing its June 2026 target — Gemini 4 is being trained in parallel, not as its replacement.
• No release date, model ID, price, benchmarks, or context window have been published.

What Is Gemini 4?

Gemini 4 is the working name for Google DeepMind's next frontier model generation, currently in the pre-training stage. Google confirmed the run in a single line at the end of its July 21, 2026 model announcement: "We have already started our most ambitious pre-training run yet, for Gemini 4, and can't wait to share more." Logan Kilpatrick, who leads Google AI Studio, repeated the confirmation on X the same day.

Two days later, Sundar Pichai extended the narrative on Alphabet's quarterly earnings call. He described Gemini 4 as a "significantly larger" frontier model and said Google plans to move to an "almost monthly" release cadence for Gemini models going forward, with Gemini 4 as the baseline. That is the entire confirmed product story.

Everything else — parameter counts, architecture, modality support, context window, pricing — remains unpublished. Speculation that Gemini 4 will ship by late 2026 is an inference from Google's historical six-month training cycles, not a Google commitment. As Andrew Curran read the announcement on X: "I take this to mean it will be the largest model they have ever trained."

Gemini 4 at a Glance

Field	Detail
Developer	Google DeepMind
Type	Frontier large language model
Status	Pre-training (confirmed July 21, 2026)
Positioning	"Significantly larger" than prior Gemini models
Modality	Not yet confirmed
Context window	Not yet confirmed
Parameters	Not yet confirmed
Priority workloads	Coding and autonomous agents (stated by Pichai)
Pricing	Not yet confirmed
Availability	None — no API, no preview, no waitlist
License	Expected proprietary (unconfirmed)
Release date	Not announced

Every "Not yet confirmed" here is deliberate. As of publication, Google has published no model card, no ID string, and no preview endpoint.

How Gemini 4 Works and What Makes It Different

The technical architecture of Gemini 4 has not been disclosed. Google has committed to three framing points that shape what to expect.

Most Ambitious Pre-Training Run. This is Google's own phrasing. Curran and other tier-A observers read it as a claim about scale — likely the largest compute budget and parameter count Google has ever committed to a Gemini training run. It does not necessarily mean the largest dense model; Google has shipped Mixture-of-Experts architectures in prior generations without disclosing them at launch.

Significantly Larger Frontier Model. Pichai's earnings-call phrasing is the second confirmed anchor. It signals that Gemini 4 is not a Flash-tier speed play but a top-of-stack reasoning model meant to compete with GPT-6 and Claude Opus successors.

Coding and Autonomous Agents. Both were named as explicit priorities. This aligns with the industry-wide shift toward agentic workloads — models that plan, call tools, and complete multi-step tasks rather than answering single prompts. Google's current Gemini 3.6 Flash already ships with computer-use capabilities scoring 83% on OSWorld-Verified, and Gemini 4 is expected to push that ceiling.

Emily, an ML observer, framed the ambition this way on X: "all the research in the last two years on world model, multi modalities, long horizon agents and new TPUs will come together for the first time." That is speculation, but it captures the assembled bet Gemini 4 represents.

What You Can Do With Gemini 4

Nothing, yet. Gemini 4 is not available to developers, enterprises, or consumers. There is no preview program, no waitlist, and no public model ID.

If you need Google frontier reasoning today, the practical options are the current-generation models: Gemini 3.1 Pro for advanced reasoning and Gemini 3.6 Flash for cost-efficient agentic workflows. Our analysis of the delayed Gemini 3.5 Pro covers what the intermediate flagship was supposed to deliver before it stalled in partner testing.

Expected use cases for Gemini 4 based on Google's stated priorities:

• Long-horizon agentic coding (multi-file refactors, autonomous PR generation)
• Computer-use and browser-control automation at higher reliability
• Complex reasoning over large multimodal contexts
• Enterprise workflows that today fall back to Claude Opus or GPT-6

None of these are validated on Gemini 4 itself — they are the workloads Google's public language suggests the model is being optimized for.

How Gemini 4 Compares

A direct comparison table is not yet meaningful — Gemini 4 has no published numbers. What can be said is where it sits in the competitive frame Google itself has invoked.

Model	Status	Vendor positioning
Gemini 4	Pre-training	"Most ambitious pre-training run yet"
GPT-6	Rumored imminent	Successor to GPT-5.6, per pre-release reporting
Claude Opus 4.8	Available	Anthropic's current top-tier reasoning model
Gemini 3.5 Pro	Partner testing	Delayed since June 2026

Our first look at GPT-6 covers the closest competitor Gemini 4 will face at launch. The competitive read from independent observers is skeptical of Google's timing: pre-training just started while GPT-6 is reportedly weeks from release.

Availability: How to Access Gemini 4

There is no way to access Gemini 4 today. No API endpoint, no AI Studio preview, no consumer app rollout, no partner testing program. Pre-training is a compute milestone — the model does not yet exist as a usable artifact.

After pre-training completes, Gemini 4 will still need post-training (instruction tuning, RLHF), reasoning optimization, red-team safety evaluation, and external testing before any public release. Historically this stage has taken Google two to six months per model.

When Gemini 4 does ship, expect the standard Google surface distribution: Gemini API and Vertex AI for developers, the Gemini app for consumers, Google Workspace integrations for enterprise, and likely a tiered rollout (partner preview → paid preview → GA). No official confirmation of any of this exists.

What We Don't Know Yet

The list of open questions is longer than the list of confirmed facts:

• Exact parameter count and whether Gemini 4 uses a Mixture-of-Experts architecture
• Context window (Gemini 3.1 Pro is 1M tokens; Gemini 4 could match or exceed)
• Modality support — text-only, multimodal, or native video/audio
• Whether Gemini 4 ships as a family (Pro/Flash/Ultra) or a single model at launch
• Pricing tier and API model ID string
• Concrete release timeline
• Whether Gemini 3.5 Pro ever ships or gets quietly skipped, as Mark Kretschmann and others have publicly speculated

TestingCatalog's estimate of a year-end 2026 arrival is an inference from six-month training cycles, not a Google commitment. Treat every dated prediction you see elsewhere as speculation until Google publishes a model card.

Frequently Asked Questions

What is Gemini 4? Gemini 4 is Google DeepMind's next-generation frontier large language model, confirmed to be in pre-training as of July 21, 2026. Google describes it as its most ambitious pre-training run to date. No specifications, benchmarks, or release date have been published.

When will Gemini 4 be released? Google has not announced a Gemini 4 release date. Based on Google's typical six-month training cycles, some observers speculate a late-2026 launch, but this is unconfirmed. Pre-training is only one stage before post-training, safety testing, and public release.

Is Gemini 4 available now? No. Gemini 4 is not available in any Google product, API, or preview. There is no gemini-4 model ID in the Gemini API or Vertex AI catalogs, and Google has not opened a waitlist. Only pre-training has begun.

How is Gemini 4 different from Gemini 3.5 Pro? Gemini 4 and Gemini 3.5 Pro are separate models developed in parallel. Gemini 3.5 Pro is still in partner testing after missing its June 2026 target, while Gemini 4 is a new pre-training run built for the next generation and described as significantly larger.

How much will Gemini 4 cost? Gemini 4 pricing has not been announced. For reference, Google's current models on the API are priced from $0.30 per million input tokens for Gemini 3.5 Flash-Lite up to $2 per million input tokens for Gemini 3.1 Pro.

Will Gemini 4 be open source? No indication of open weights. Google has released weights only for its smaller Gemma family, while flagship Gemini models remain proprietary and API-only. Gemini 4 is expected to follow the same closed-weights pattern.

How does Gemini 4 compare to GPT-6? A direct comparison is not yet possible because Gemini 4 has no published benchmarks, specs, or release. Both are positioned as next-generation frontier models from their respective labs and will likely compete on reasoning, coding, and agentic workloads once Gemini 4 ships.

What to Watch Next

Three signals will move this page forward. First, any Google DeepMind publication naming a Gemini 4 variant (Pro, Flash, or Ultra) with even a partial spec sheet — that flips the model from "training milestone" to "product." Second, the fate of Gemini 3.5 Pro: a public launch, a rebrand to Gemini 3.6 Pro, or a quiet cancellation each shifts what Gemini 4 needs to be at launch. Third, watch for a gemini-4 model ID appearing in the Gemini API or Vertex AI catalog change logs, which historically precede preview access by days.

Building similar frontier-scale chat and agent workloads? On kie.ai you can try Gemini 3.1 Pro, Gemini 3.6 Flash, and Claude Opus 4.8.