The Model That Keeps Not Shipping: Google's Gemini 3.5 Pro Slips Past Another Deadline

A launch date that came and went, quietly

The most consequential AI story of the past 48 hours is not a model that shipped. It's a model that didn't.

On July 16, Bloomberg reported that Google is months behind schedule on Gemini 3.5 Pro, the flagship model Sundar Pichai told developers at May's I/O conference was coming "next month." June came and went. So did the middle of July. As of this writing, Gemini 3.5 Pro has not been released.

Google did not dispute the delay. A spokesperson told Bloomberg: "We're currently testing 3.5 Pro, an upgraded Flash model, and other models with partners, and we're productively engaged with the U.S. government on model testing and broader frameworks." That is a confirmation dressed as a progress update. It contains no launch date.

The stated reason is narrow and specific, which makes it more interesting than a generic slip. Per Bloomberg, the model's coding capabilities fell short of internal expectations. Google updated the training data late last month specifically to fix this, and — in the words of one person cited in the reporting — "the results were disappointing."

Why coding is the metric that decided this

It is worth pausing on what Google chose to hold a flagship model back over.

Not safety evaluations. Not multimodal reasoning. Not context length. Coding. In 2026, code generation has become the load-bearing benchmark for frontier models, because it is the capability that converts most directly into enterprise revenue and developer lock-in. It is also, unlike most benchmarks, brutally verifiable: the code runs or it doesn't, the tests pass or they don't. You can market your way around a soft eval. You cannot market your way around an agent that writes a patch that breaks the build.

Pichai has already conceded the ground publicly. As Search Engine Journal notes, he previously acknowledged Google was "a bit behind" on agentic coding relative to competitors. The delay reads as Google declining to ship a flagship that would make that admission measurable.

There's a defensible reading here, and I want to give it fair weight: shipping a model you know underperforms on the capability everyone will immediately test is worse than shipping late. Google has been through this. A rushed launch that gets picked apart within 72 hours does more durable brand damage than a delay that gets one bad news cycle.

The organizational story underneath the technical one

The more damaging detail in Bloomberg's reporting isn't the benchmark shortfall. It's the explanation for why fixing it is hard.

The reporting describes structural complexity inside Google — DeepMind, Google Cloud, Android, and Search all developing overlapping AI coding tools simultaneously — with shifting priorities and competing responsibilities across teams slowing execution. PYMNTS summarized the same dynamic as multiple internal stakeholder layers involved in preparing a model release, plus competing factions building separate coding tools.

Bloomberg's account was sourced to current and former employees — Search Engine Journal puts the number at ten — and describes the delay as a source of frustration among Google engineers, researchers, and managers, many of whom worry the company risks losing its edge as Anthropic and OpenAI ship models exceeding Gemini's capabilities.

That last clause is the part Google cannot fix with more training data. A capability gap is an engineering problem. A capability gap that ten insiders are willing to describe to Bloomberg is a morale problem, and morale problems in AI labs express themselves as departures, which express themselves as further capability gaps.

What Google actually has in market right now

Here is the state of play, sticking strictly to what's confirmed.

Gemini 3.5 Flash — announced alongside the Pro promise in mid-May — is the only released model in the 3.5 line. It is not a placeholder in any meaningful sense: per PYMNTS, it became the default across the Gemini app and Search's AI Mode, serving an app with over 900 million monthly users across 230 countries. 9to5Google notes that Gemini 3.1 Pro shipped back in February 2026, so the Pro tier is not absent, just stale.

This matters for calibrating the alarm. Google's consumer AI surface is not degraded by this delay — Flash is carrying it, at enormous scale. Search's AI features are unaffected. What's affected is the top of the developer funnel: the model you reach for when the task is hard, when you're building an agent, when you're deciding which lab's API your company standardizes on. That's the segment where being nine months past your last Pro release is a real cost.

An upgraded Flash model is confirmed to be in partner testing. Secondary outlets have reported that Google has registered names suggesting stopgap releases, and have characterized this as a third missed deadline with ongoing hallucination problems. Those specifics are not confirmed by Bloomberg's reporting or by Google, and should be treated as unverified.

Hype versus real

Three claims are circulating that deserve separating.

Real: Gemini 3.5 Pro is late, the reason is coding performance, Google has confirmed it's still in testing, and there is no announced date. All of this is on the record.

Real but easily overstated: the market reaction. CNBC reported that Alphabet shares fell on the news. Single-day moves on a model-delay headline are sentiment, not a verdict on Google's AI position — a company with 900 million monthly Gemini users and its own TPU stack is not fragile because one release slipped.

Overstated: framing this as Google losing the AI race. The evidence supports a specific, bounded claim — Google is behind on frontier coding and knows it — not a general collapse. Anyone extrapolating from "the Pro model is late" to "DeepMind is broken" is running well past what the sources say.

There's a genuinely underappreciated angle, though. 9to5Google's report also references the widely-cited figure that as of April 2026, roughly 75% of new code at Google is AI-generated and engineer-approved. If Google's own model is falling short on coding, and Google's own engineering is increasingly downstream of AI-generated code, the delay isn't just a product problem. It's a bet on the company's internal velocity that hasn't paid off yet.

The takeaway

The clarifying fact of this story is that Google chose the delay. A company under this much competitive pressure, with a CEO who publicly promised a June date, decided that shipping a Pro model with sub-target coding performance was worse than the humiliation of missing three months and counting. That is a judgment about how thoroughly the market now tests coding capability — and how quickly a weak flagship gets exposed.

Whether it was the right call depends on something none of the reporting can tell us: whether the next attempt clears the bar. If Gemini 3.5 Pro eventually ships and is genuinely competitive on agentic coding, this week reads as discipline. If it ships late and merely adequate, Google will have paid the delay cost and gotten nothing for it.

The detail I'd watch isn't the launch date. It's whether the organizational fragmentation Bloomberg's sources described — four divisions building overlapping coding tools — gets addressed. A model can be retrained. A structure that produces late models tends to produce more of them.