Google's next Gemini Pro model has slipped from June into July, and the delay tells you exactly where the AI race now hurts: coding agents, long tasks, and token costs.
At Google I/O in May, Sundar Pichai asked developers for patience on Gemini 3.5 Pro. The message was simple enough: wait until next month. June came and went. The model still isn't broadly available.
You'd think a company with Google's compute budget could make that look routine. It hasn't. Business Insider reported in late June that Google had pushed Gemini 3.5 Pro into July while it gathered feedback from early users on Antigravity and LMArena. That is the verified story here. The louder claim, that Google scrapped the original base model and restarted pretraining from scratch, has been repeated in smaller tech reports, but Google hasn't confirmed it and the stronger wording doesn't deserve to be treated as fact.
That distinction matters. A delayed frontier model is normal now. A discarded pretraining run would be a much bigger admission, because it would mean months of compute and engineering time were effectively written off. If you're building on these APIs, you don't need drama. You need to know whether the thing will ship, what it will cost, and whether it will actually handle the work you want to give it.
Right now, Google has confirmed less than developers would like. AP reported from Google I/O that Gemini 3.5 Flash was announced for faster, code-heavy work and that Gemini Pro was expected the following month. The Verge also reported that Google made Gemini 3.5 Flash the default model in the Gemini app and Search while pushing deeper agentic features through Workspace, Android, Search, and AI Studio. Pro is the missing piece.
It isn't a cosmetic upgrade.
The delay is about agent work, not chatbot polish
The pressure on Gemini 3.5 Pro sits in coding, tool use, and long-horizon tasks. Business Insider's report said Google was collecting feedback around agent performance and token consumption, including lessons from Gemini 3.5 Flash. That is exactly where enterprise buyers now test models. They don't care much about a lovely one-paragraph answer if the same model burns through a budget while editing a codebase or loses the thread halfway through a multi-step workflow.
OpenAI has raised the bar there. Axios reported that GPT-5.6 launched on July 9, 2026, with a new ChatGPT Work product aimed squarely at professional use. Anthropic is charging more aggressively for Claude Fable 5, according to Wired, with usage-based fees now attached to its most capable model. You can dislike that pricing shift, but you can't ignore the signal. The best models are being sold as work engines, not chat toys.
Google needs Gemini 3.5 Pro to be credible in that market. Longer context windows, better coding, stronger reasoning, and cleaner tool use aren't nice extras when your customer is a developer team deciding what sits inside its product. They are the product.
Here's the thing: a July launch can still be fine. A weak July launch can't. If early testers found the model too expensive in tokens or too brittle on long tasks, delaying it is the right call. Shipping an undercooked Pro model just to satisfy a conference promise would be worse for Google than taking another few weeks and owning the slip.
DeepSeek gives developers another clock to watch
Google's timing is awkward because DeepSeek has its own transition running in the background. AP reported in April that DeepSeek rolled out V4 in Pro and Flash versions, with a 1 million token context window and deeper support for Huawei chips. The model is already part of the competitive frame, especially for teams watching price as closely as benchmark scores.
Several model trackers and technical summaries now point to a July 24, 2026 cutoff for older DeepSeek API names such as deepseek-chat and deepseek-reasoner. Treat that as an implementation risk, not a market anecdote. If your product depends on one of those names and the routing changes under you, your customers won't care that the AI industry is moving quickly. They'll just see failures, weaker reasoning, or higher bills.
That is the pressure.
Developers choosing an AI stack for the second half of 2026 are looking at Google, OpenAI, Anthropic, and DeepSeek with a simple question: which one gives me the least painful path from prototype to production? Google has distribution, cloud relationships, and the Gemini footprint across Search, Workspace, Android, and Vertex. But distribution doesn't answer an API call. The model does.
For now, the honest answer is to wait for the actual Gemini 3.5 Pro release notes and pricing. Don't build around leaked specs as if they're a contract, and don't assume a July 17 date is firm just because it's circulating. DeepSeek's July 24 migration matters too, if you're already on its API.
Google can still make this month work. But the next Gemini Pro has to arrive as a useful tool, not as a delayed promise with a bigger model number attached.
Also read: Meta Pulled Its Muse AI Tool But Your Instagram Photos Are Still In It • Amazon Has Cut 57,000 Jobs While It Pours $200 Billion Into AI • Anthropic's New Lens Shows What Claude Is Thinking Before It Speaks
