Google’s Gemini 3.5 Pro Is Stuck, and the Reason Is Uncomfortably Human

Back in May, on stage at I/O, Google told everyone Gemini 3.5 Pro was basically a month away. It is now the back half of July, the model still is not here, and the reason is the one thing Google would least like to admit out loud: its flagship AI cannot write good enough code yet.

That is the short version of a Bloomberg report, built on interviews with ten current and former employees, that has quietly turned into one of the more revealing stories in AI this year. Google tweaked the training data late last month to sharpen the model's coding chops. The results, per one person, were disappointing. So instead of shipping, Google is holding the line and reworking, which is the right call and also a slightly alarming one.

Coding is the whole ballgame now

Here is why this matters more than a normal slipped launch. Coding is not just one feature on a checklist anymore. It is the feature. Developers pay for the models that write and debug software best, and those same coding tasks are how you build the agents that can chain a dozen steps together without falling apart. If your model stumbles there, it does not matter how gracefully it writes a sonnet.

Google knows this. Sundar Pichai has already said out loud that the company was "a bit behind" on agentic coding, and there is a structural reason: Google has historically lacked the big developer-facing coding product that hands rivals a firehose of real-world training data. OpenAI and Anthropic have been eating well on exactly that. You can feel the internal frustration in the reporting, the sense of a company with world-class researchers watching smaller labs ship the thing everyone actually wants.

The stopgap: a faster Flash

So what does Google do while the Pro model cooks? It leans on the cheap seats. Google confirmed it is testing an upgraded Flash model with partners alongside the still-unreleased Pro. Flash is the small, fast, inexpensive tier, and right now Gemini 3.5 Flash is the only member of the 3.5 family that has actually shipped. Pushing out a better Flash keeps Google in the conversation, keeps the cost-per-token pitch alive, and buys time. It does not, however, answer the question people are really asking, which is whether the flagship can hang with the best.

Google's official framing leans hard into that value angle. "We're shipping quickly across a wide range of models while keeping them highly cost-effective for customers," the company said. Which is true, and also a very careful way of talking about the model you have not managed to release.

Give Google credit for not shipping

Here is the part worth sitting with. It would have been easy to push Gemini 3.5 Pro out the door in June, slap a benchmark chart on it, and let the marketing sort out the rest. Google did not do that. A model that hallucinates too often or falls over on multi-step tasks is worse than no model at all, because the trust you lose when an AI confidently botches someone's production code is expensive and slow to earn back. Holding a flawed flagship is the disciplined move.

But discipline and delay start to look the same from the outside, and the clock is not on Google's side. Every month Gemini 3.5 Pro stays in the oven is another month OpenAI and Anthropic get to define what a top-tier model feels like to use. The vision here is still enormous, and DeepMind's research bench is as deep as anyone's. The shipping product is what is late.

The real test is not the launch date. It is what shows up when the launch finally comes. If Gemini 3.5 Pro arrives and genuinely codes with the best of them, nobody will remember the wait. If it arrives merely fine, the wait will have cost Google far more than a few months. Either way, we will know soon enough, and "soon enough" is doing a lot of work in that sentence.