Google's most powerful next-generation model was supposed to be in developers' hands by now. Instead, Gemini 3.5 Pro is months behind schedule, held back by coding results that fell short of the company's own internal targets, according to a Bloomberg report that sent Alphabet's stock lower and reignited doubts about whether the search giant can keep pace in the fiercest corner of the AI race.
The delay, reported Thursday by Bloomberg and confirmed in broad strokes by a Google statement, is notable in part because of what it undoes. As recently as this week, industry watchers had expected Gemini 3.5 Pro to arrive imminently — some had pegged a launch as soon as Friday, July 17. Instead of a ship date, Google now has a reset training run and no firm timeline.
What Google said it would deliver — and when
At its I/O developer conference in mid-May, Google unveiled Gemini 3.5 Flash and told developers that the more capable Pro version would follow in June. On stage, the company said the model was "showing great improvements." That June deadline came and went without a release or an updated target.
According to Bloomberg, which spoke with roughly ten current and former employees, Google has been "taking time to try to improve" Gemini 3.5 Pro's capabilities, "particularly in coding." The model's performance on coding and complex, long-horizon reasoning tasks reportedly fell short of what the company expected internally. In late June, Bloomberg reported, Google updated the data used to train Gemini in an effort to lift its programming skills — but the results were "disappointing," pushing the model back into development.
Because this is reporting sourced to anonymous current and former employees rather than an official product announcement, the specifics of the training reset and the missed internal benchmarks should be read as Bloomberg's account, not a Google disclosure. Several of those people described real frustration inside the company, with some warning that Google risks a lasting loss of momentum against nimbler rivals.
Google's response
Google did not dispute that a delay is underway, but framed it around continued testing rather than failure. In a statement, the company said it is "currently testing 3.5 Pro, an upgraded Flash model, and other models with partners," and added:
> "We're shipping quickly across a wide range of models while keeping them highly cost-effective for customers."
That reference to "an upgraded Flash model" lines up with reporting that Google is eyeing a stopgap: rather than wait indefinitely for Pro, the company appears prepared to put a stronger Flash-class model in front of customers to stay competitive while Pro is reworked. The current flagship on the Pro track, Gemini 3.1 Pro, dates back to February — an eternity by the standards of this year's release cadence.
Market reaction
Investors did not take the news calmly. Alphabet shares fell roughly 3% intraday after the report and closed sharply lower on Thursday, with market coverage putting the decline at about 4.4% and the erased market value near $200 billion. For a company whose AI narrative has been a central pillar of its valuation, a stumble on the coding frontier — the capability most directly tied to enterprise adoption and developer loyalty — is exactly the kind of headline that moves the stock.
Analysis: the coding race is the whole race
Coding has become the sharpest edge of competition among frontier labs, and it is the one place Google can least afford to slip. OpenAI and Anthropic have built much of their enterprise momentum on models that developers trust to write, debug, and refactor real software, and Bloomberg noted that recent rival releases have outpaced Google's current offerings on generating code. Anthropic's Claude line and OpenAI's latest systems have set the bar that Gemini 3.5 Pro was explicitly measured against internally — and missed.
The pressure is compounded by Google's sheer breadth. Unlike a pure-play lab, Google has to ship models across a giant product portfolio — the Gemini app, Vertex AI on Cloud, Android Studio, AI Studio, Workspace — each with its own teams and its own coding-tool ambitions. That fragmentation is itself part of the story: employees told Bloomberg that an effort to "unite the company's internal artificial intelligence coding tools" is underway, and that some engineers have resisted AI-written code on the grounds that important code should be human-authored to meet Google's standards. Even so, the company has said 75% of new code at Google is now AI-generated and approved by engineers, up from 50% last fall — a sign of how much Google itself is betting on the exact capability its flagship model just failed to nail.
The timing sharpens the sting. The delay landed the same week that China's Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model that vaulted to the top of the Frontend Code Arena with 1,679 points, edging past closed flagships from Anthropic and OpenAI. An open model topping a coding leaderboard just as Google's closed flagship slips is a vivid illustration of how quickly the field is moving — and how little slack Google has if it wants to be seen as a leader rather than a fast follower.
What to watch next
The near-term question is whether Google ships the "upgraded Flash" stopgap in the coming weeks and how it performs on public coding benchmarks — a strong Flash release could blunt the competitive damage even without Pro. The longer-term question is whether the reset training run finally clears Google's internal bar, and when. Watch, too, for how Google talks about Gemini 3.5 Pro at its next public moment: a firm date would signal confidence, while continued silence would suggest the coding problem runs deeper than a single bad training batch. For now, the model that was supposed to headline Google's summer is instead a cautionary tale about how hard the last mile of frontier coding has become.
"We're shipping quickly across a wide range of models while keeping them highly cost-effective for customers."— Google, Company statement