Google has started pre-training Gemini 4, a model it calls its most ambitious yet, even as the company’s current flagship, Gemini 3.5 Pro, remains stuck in limited testing due to stubborn coding performance issues. CEO Sundar Pichai disclosed the Gemini 4 effort during Alphabet’s July 22 earnings call, framing it as a foundation for future AI advances rather than an incremental upgrade. For Windows users and developers watching Google’s AI moves, the announcement signals both a bold long-term bet and a near-term reality: the model you can actually build with today is not the one promised for tomorrow.

Google’s Two-Track AI Strategy Takes Shape

The announcement came alongside a flurry of model updates on July 21, where Google released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite to general availability. These lighter-weight models are designed for cost-sensitive, high-volume workloads, with Gemini 3.6 Flash delivering improved coding, multimodal tasks, and token efficiency—Google claims a 17% drop in output-token usage and prices it at $1.50 per million input tokens, $7.50 per million output tokens.

Meanwhile, Gemini 3.5 Pro, originally teased at Google I/O in May with a promised June release, is still confined to partner testing. According to reporting by Bloomberg (via 9to5Google), the delay stems from disappointing results in complex coding benchmarks after an attempt to improve training data. Google officially says it will release the Pro model “as soon as it’s ready,” but has given no date.

This dual cadence—shipping Flash models while holding back the Pro tier—allows Google to maintain momentum without rushing a product that flunks internal quality bars. But it also highlights a persistent weakness: agentic coding.

Why Coding Became the Deciding Factor

Pichai has been unusually candid about the coding gap. In a May interview on the Hard Fork podcast, he admitted Google was “a bit behind” in agentic coding and tool use, areas where competitors like OpenAI and Anthropic have built strong feedback loops through products such as ChatGPT and Claude. Coding isn’t just another benchmark; it’s becoming the proving ground for AI that must plan, use tools, handle multi-step tasks, and recover from errors—capabilities critical for enterprise automation and real-world development.

The Flash models offer a partial answer. Gemini 3.6 Flash, for instance, boasts improvements in DeepSWE, OSWorld-Verified, and MLE Bench, along with fewer unnecessary code edits and execution loops. Yet these are Google’s own claims; Windows developers integrating these models should validate them against their own repositories and workflows.

What It Means for You

For Windows home users: The Gemini models you encounter inside Google AI Studio, Vertex AI, or via the Gemini API haven’t changed. Any app or service you use that relies on them continues to operate as before. The Flash releases may subtly improve performance behind the scenes, but your daily experience won’t shift dramatically. If you’re curious about AI coding assistants, note that Google’s internal tool, Antigravity, now has over 2.4 million weekly active users—a sign of growing internal momentum that may eventually feed into consumer-facing products.

For developers and power users: Your immediate actionable models are Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. They are production-ready, documented, and accessible via API. Gemini 3.6 Flash is the go-to for tasks where latency and cost matter as much as raw reasoning power. Gemini 3.5 Flash-Lite serves as a cheaper, faster option for simple automations. Avoid designing critical systems around Gemini 3.5 Pro until Google lifts the testing restriction and publishes a public preview.

For IT admins and enterprise architects: The Flash models make sense for pilot projects, especially in areas like document processing, ticket triage, or low-stakes chatbot interactions. Before deploying any model, run your own evaluations: measure task completion rates, total cost per task (including human review time), latency under load, and security constraints. With Google aiming for a near-monthly update cadence after Gemini 4, your infrastructure must expect version churn. Separate your business logic from the model layer, maintain prompt libraries, and build regression tests. Don’t commit to exclusive Google AI agreements yet if your long-term architecture demands flexibility; keep competitive options open.

How We Got Here: A Timeline of Ambition and Delays

Google’s AI release pace has been inconsistent. After the initial Gemini 1.0 series, the company shipped Ultra, Pro, and Nano, then followed with 1.5 Pro and Flash in early 2024. Gemini 2.0 arrived later that year, but the leap to 3.5 was supposed to be a generational jump. At I/O in May, Google announced Gemini 3.5 Flash first, then teased Pro for “next month.” When June passed with no release, whispers of coding troubles grew louder.

The Flash line moved forward: Gemini 3.5 Flash became generally available, and then the even more efficient 3.6 Flash landed on July 21, 2026. But Pro’s absence stuck out. The competitive landscape didn’t pause—OpenAI’s GPT-5 and Anthropic’s Claude 4 set high bars for agentic performance. Meanwhile, developers flocked to tools that delivered reliable coding help, leaving Google playing catch-up despite its massive infrastructure, which processes approximately 22 billion tokens per minute across more than 9 million developers.

Pichai’s earnings call remarks on July 22 acknowledged the problem head-on. He said reaching the “next real capability jump” required training larger base models, and that Gemini 4 would be built for the frontier “where it will be” at launch, not just to match current rivals. He also hinted at a faster release tempo once that foundation is laid. Internally, Antigravity’s rapid adoption—reportedly doubling weekly and compressing a two-year Chrome refactoring into three months—offers hope that Google can close the feedback gap.

What to Do Now: Practical Steps

  1. Experiment with available models immediately. Create a sandbox with Gemini 3.6 Flash and test it on tasks that mirror your typical workloads—document parsing, code generation, data extraction. Use the Gemini API’s changelog to stay updated on model versions and retirement notices.
  2. Benchmark real outcomes, not demo scores. Track whether the model completes multi-step tasks correctly and consistently. Calculate total tokens consumed per task, not just per prompt. Measure latency in your own network environment, especially if calling from Windows servers or Azure-based services.
  3. Design for portability. Write your application code so that the model endpoint, prompts, and evaluation logic are modular. This allows you to swap in future Gemini releases—or competitor models—without a rewrite. If you’re building a Windows desktop agent, keep the AI integration separate from the UI and automation scripts.
  4. Monitor Gemini 4 news without acting yet. Sign up for Google AI’s developer newsletter and follow official blogs. When Gemini 4 eventually enters public preview, you’ll have time to assess its coding and agentic improvements before anything becomes production-critical. Premature integration based on roadmap promises is a recipe for rework.
  5. Evaluate Flash-Lite for high-volume scenarios. If you manage thousands of daily interactions, Gemini 3.5 Flash-Lite’s lower cost and latency might surprise you. But verify that its accuracy holds up for your use case—cheap failures aren’t really cheap.

Outlook: A Pivotal Year for Google’s AI

The Gemini 4 pre-training run marks the start of a long journey. Full training, safety evaluations, and productization likely stretch into 2027. The more immediate question is whether Gemini 3.5 Pro will ever see a wide release, or if Google will quietly supersede it with a refined 3.6 Pro or jump directly to 4. Pichai’s vision of near-monthly updates suggests a future where “version numbers” blur into continuous improvement—a model AI consumers and enterprises will need to adapt to.

For Windows users, developers, and IT shops, the message is clear: build with what’s solid today, keep your architecture flexible, and watch Google’s next moves. The AI race isn’t a sprint, and Google’s willingness to invest in a foundation model while fixing its coding blind spot could eventually pay off—if the follow-through matches the ambition.