Google has officially put a name to its next major artificial intelligence model: Gemini 4.
The confirmation came during Alphabet’s second-quarter 2026 earnings call, where CEO Sundar Pichai revealed that Google has started what he described as the company’s most ambitious pre-training run yet.
That is the solid part. Gemini 4 exists, training has begun, and Google says it is already seeing encouraging progress at the frontier of AI development.
Beyond that, the picture becomes less complete. Google has not announced a release date, benchmark results, pricing, model sizes or a detailed feature list. A lot of what people expect from Gemini 4 comes from the direction Google is already taking with Gemini 3.5 and 3.6—not from a finished product announcement.
Google Has Officially Started Training Gemini 4
Pichai confirmed that Google is already building its next generation of Gemini models while discussing the company’s current AI roadmap.
“We have started our most ambitious pre-training run yet for Gemini 4,” he said during the July 22 earnings call, adding that the company is encouraged by the progress it is seeing.
That wording matters. Gemini 4 is not merely an internal concept or a name pulled from leaked code. Google has publicly acknowledged the model and confirmed that a major training run is underway.
Still, this sounds like a model deep in development rather than something close to general release. Pre-training is one of the earliest and most expensive stages of building a frontier model. Fine-tuning, safety testing, evaluation and product integration usually follow.
Google may have confirmed Gemini 4, but it has not promised that users will receive it soon.
Gemini 3.5 Pro Is Still Being Tested
The timing is slightly unusual because Google has not completed the Gemini 3.5 rollout.
Gemini 3.5 Flash arrived at Google I/O 2026, while Gemini 3.5 Pro remains in testing. Google DeepMind’s current model page still lists Gemini 3.5 Pro as “coming soon,” with Gemini 3.1 Pro remaining the available option for more complex work.
Google has also moved ahead with Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Its official model-card directory, updated on July 21, lists both of those releases but does not yet include a Gemini 4 model card.
So the model timeline is no longer a tidy sequence where one generation ends before the next begins. Google is training Gemini 4 while refining, testing and releasing several versions within the Gemini 3 family.
Messy? A little. It also reflects how quickly frontier AI development now moves.
Coding Looks Like an Obvious Target for Gemini 4
Google has not published a Gemini 4 capability sheet, although coding will almost certainly receive serious attention.
Pichai acknowledged during the earnings discussion that Google still has room to improve in coding and agentic coding. That admission came as developers increasingly compare Gemini with coding models and agents from OpenAI, Anthropic and other rivals.
The company has already made agentic coding central to Gemini 3.5. Google says its current models can reason through complex development tasks, maintain codebases and carry out longer workflows instead of producing isolated snippets.
Gemini 3.5 Flash, for example, was launched around the idea of combining intelligence with action. Google highlighted its performance on coding and agent benchmarks, including Terminal-Bench 2.1 and MCP Atlas.
Gemini 4 will probably push further in that direction. The real question is whether it can become more reliable over long coding sessions, where one incorrect decision can quietly break an entire project.
Generating code is no longer enough. The difficult part is planning, testing, debugging and knowing when something has gone wrong.
Agentic Workflows May Become More Important Than Chat
Google’s recent language around Gemini has shifted noticeably. The company talks less about a chatbot that answers questions and more about systems that can perform work.
Its Gemini 3.5 lineup already focuses on long-horizon tasks, multi-step problem-solving and workflows that use external tools. Google describes these models as capable of handling extended tasks rather than waiting for a fresh prompt at every step.
That gives us a reasonable clue about Gemini 4.
The next model may place more emphasis on planning actions, working across applications and completing complicated assignments with limited supervision. Think software development, research, data analysis, cybersecurity and enterprise operations—not only writing an email or summarizing a webpage.
None of this guarantees that Gemini 4 will work autonomously without mistakes. AI agents still struggle with consistency, tool failures and tasks that stretch over long periods. Google’s challenge will be making those systems dependable enough for real work.
A flashy demo is easy. Finishing the job correctly is harder.
Google’s Infrastructure Is Part of the Gemini 4 Story
Calling Gemini 4 Google’s “most ambitious” pre-training run suggests a large increase in the scale or complexity of the project.
Google has several advantages here. It develops its own Tensor Processing Units, operates enormous data centers and can distribute Gemini through Search, Android, Workspace, Google Cloud and the Gemini app.
That infrastructure gives the company somewhere to train a frontier model and somewhere to deploy it.
Alphabet has also been spending heavily to expand AI capacity. The company’s latest earnings discussion placed infrastructure near the center of its strategy as demand grows across consumer products and Google Cloud.
A larger model alone will not decide the race, though. Bigger training runs bring higher costs, slower experimentation and tougher deployment problems. Google still needs to make Gemini 4 fast enough and affordable enough to use across products serving hundreds of millions of people.
Cloud and Enterprise Use Will Likely Be Built In Early
Gemini 4 will not be designed only for the Gemini app.
Google has spent the past several years placing its models inside Vertex AI, Workspace and its broader enterprise platform. New Gemini releases now reach developers through Google AI Studio, Gemini APIs, Android Studio and Google’s agent-focused development tools.
The company is also showing how businesses can combine Gemini models with multiple subagents. Salesforce, for example, is integrating Gemini 3.5 Flash into Agentforce for complicated enterprise tasks involving context retention and multi-turn tool use.
Gemini 4 will probably enter that same ecosystem. Enterprise customers may become some of its earliest and most important users, especially where Google can bundle the model with cloud infrastructure, data services and workplace software.
That distribution could matter as much as the model itself.
Google Has Not Announced a Gemini 4 Release Date
No public launch date has been confirmed.
Google has not said whether Gemini 4 will appear in late 2026, sometime in 2027 or through a limited preview before a wider rollout. It has not announced whether Flash, Pro, Nano or Deep Think variants will accompany the main release either.
There are no official benchmark scores. No context-window details. No API pricing. No system card. No promise that Gemini 4 will immediately replace every model in the Gemini 3 family.
Those missing details are worth keeping in view. The phrase “Gemini 4” will attract speculation, but Google’s actual announcement was narrow: training has begun, the run is unusually ambitious, and the company likes the progress so far.
Everything else still needs confirmation.
Gemini 4 Shows Google Is Not Waiting for Gemini 3 to Finish
Google’s AI development cycle is becoming continuous.
Gemini 3.5 Pro is still being tested. Gemini 3.6 Flash has already arrived. Gemini 4 is now in pre-training. Different generations are overlapping, with smaller releases and specialized models appearing between the flagship launches.
That may feel confusing to users, but it could become normal across the industry. AI companies no longer have the luxury of releasing one polished model every year and leaving it untouched.
For Google, Gemini 4 is also a statement. The company wants investors, developers and competitors to know it is still willing to spend heavily on frontier training.
Whether that ambition turns into a better model is another matter.
The name is confirmed. The training run is real. The proof will come later.
