This week is a full-on pileup in the AI world. Four major models are landing within the same stretch of days, but one of them is still stuck backstage: Gemini 3.5 Pro. And honestly, the story behind the delay is more interesting than the launch itself would have been.
What Google Promised
Back in May at I/O, Sundar Pichai stood on stage and told developers to “give us until next month.” That next month was June. June came and went, no model. Now the target is July 17, which is tomorrow. But let’s be clear about something: that date is not an official Google announcement. It’s coming from leaks and third-party reporting. As of this writing, Google’s official API pages don’t even list a gemini-3.5-pro model ID, just gemini-3.5-flash and gemini-3.1-pro-preview.
So I can’t write “it’s definitely dropping tomorrow.” But understanding why it’s this delayed makes the story worth telling either way.
The Real Reason for the Delay
Google finished the original Gemini 3.5 Pro, tested it, and then scrapped the whole thing. The reason is simple but painful: the model was breaking down in two critical areas.
First, it couldn’t hold structural consistency in complex SVG scene generation. Second, it fell apart in recursive tool-calling, the long chains where an AI agent calls a tool, takes the result, and calls another tool based on that, over and over. The second issue is the serious one, because Google is positioning this entire model generation around agentic coding, autonomous tasks that chain tool calls together. The crack showed up exactly where they planned to sell the thing.
So Google abandoned the 2.5 Pro architecture entirely and restarted pre-training from scratch. The result, according to leaks: a 2 million token context window (double the current model), a new reasoning layer called “Deep Think,” and autonomous workflow capabilities. Again, none of this is on an official model card yet. All leaked.
Three Rivals Shipped in the Same Week
Here’s the part that actually stings. While Google was rebuilding its model, its competitors weren’t sitting still.
GPT-5.6 launched on July 9. OpenAI came out with three tiers: Sol (flagship, $5 per million input tokens, $30 output), Terra (balanced, aimed at enterprise use), and Luna (lightweight, low latency). Sol claims new records on coding, biology, and cybersecurity benchmarks. One interesting detail: GPT-5.6’s public launch was originally planned for June too, but got delayed after the US government requested early access to frontier models, so it first opened as a limited preview to about 20 government-vetted organizations.
Grok 4.5 launched the same day, July 9. Built on xAI’s new V9 base model, 1.5 trillion parameters, roughly three times the size of its previous production model. It’s supplemented with training data from Cursor, the coding assistant, aiming squarely at the developer market. Elon Musk described it as “Opus-class but faster and more token-efficient.” Pricing is aggressive too: $2 per million input tokens, $6 output.
DeepSeek V4 moves to stable release in mid-July. This one matters for a practical reason: DeepSeek’s old model names (deepseek-chat, deepseek-reasoner) get shut off entirely on July 24. So for developers building on DeepSeek’s API, this isn’t optional, it’s a forced migration.
So in the nine-day window between July 9 and July 17, three out of four major labs shipped or are about to ship, and Google is still saying “soon.”
Google’s Play: Skip the Head-to-Head, Stay Cheap
Here’s the smarter part of the story. Instead of fighting OpenAI and Anthropic head-on in benchmark wars at the premium tier, Google appears to be playing a different card: being the “good enough but much cheaper” option for cost-conscious enterprise buyers.
That’s not a bad strategy. Google already has a massive distribution advantage through Search, Workspace, Android, and Vertex. It doesn’t need to be the best model, it needs to be the reasonable one, because it’s already going to land in front of millions of users automatically. Gemini 3.5 Flash is already a small proof of that strategy. It beat the previous Pro model on Terminal-Bench 2.1, and its price point still undercuts most of the competition.
Bottom Line
Google didn’t ship a model this week, but the fact that it didn’t ship is a story on its own. A lab correctly diagnosed its problem but underestimated how long the fix would take. Meanwhile its competitors kept moving, three major models landed back to back. If Gemini 3.5 Pro actually arrives on July 17, Google’s window to establish itself is narrow, developers are already settling into stacks built around GPT-5.6, Grok 4.5, and DeepSeek V4. If it slips again, the story becomes one of a lab that saw the problem clearly but got the timeline wrong.
We’ll find out tomorrow.
Veysel Okatan