Gemini 3.5 Pro Is Months Late and DeepMind Is Rattled

Bloomberg reported hours ago that Google's Gemini 3.5 Pro, the company's flagship AI model, is months behind schedule because Google keeps going back to improve it, particularly at coding. The more interesting part of the report is not the slipped date. It is who is worried: ten current and former employees told Bloomberg that engineers, researchers and managers inside Google fear the company is losing its edge while Anthropic and OpenAI ship models that beat Gemini.

• The delay is real and it is measured in months. Gemini 3.5 Pro was shown at Google I/O on May 19 with a June availability target. That window closed with no public launch.
• Coding is the specific bottleneck. Bloomberg's sources point at coding capability as the thing Google keeps returning to fix, which matters because coding is the use case rivals monetize hardest.
• The complaint is organizational, not just technical. Bloomberg describes multiple layers of stakeholders who prepare each model for release across Search, Maps and YouTube. DeepMind staff also said Google lacks a clear commercial product for businesses building AI coding tools.
• Most of what you have read this week is unconfirmed. The July 17 launch date, the 2 million token context window and the leaked benchmark scores all trace to unnamed sources and aggregator posts. Google has published no model card.

What did Bloomberg actually report?

The core claim is narrow and worth stating precisely. Gemini 3.5 Pro is months behind schedule, and the reason is that Google has been taking time to improve its capabilities, with coding called out specifically. Bloomberg's account rests on ten current and former employees, who describe frustration among engineers, AI researchers and managers, many of whom worry the company risks losing its edge as Anthropic and OpenAI produce models that exceed Gemini's capabilities.

RelatedKAT-Coder-Pro V2.5 Lands Second on SWE-Bench Pro

Two structural details do more explanatory work than the delay itself. First, Bloomberg reports that Google has multiple layers of stakeholders involved in preparing models for release, because the company is weaving AI across a vast product portfolio including Search, Maps and YouTube. Second, staff inside DeepMind raised concerns that Google lacks a clear commercial product for businesses building AI coding tools. That second point is the one to sit with. It is not a complaint about model quality at all. It is a complaint that even a great model has nowhere obvious to land.

Why is coding the thing holding Gemini back?

Coding is where frontier models stopped being a demo and started being a business. It is the workload with the clearest willingness to pay, the most legible benchmarks, and the tightest feedback loop between capability and revenue. It is also the hardest thing to fake. A model can look strong on a single file completion and fall apart across the long, multi step tool chains that a real agent runs, where one wrong call at step nine poisons everything downstream.

That difficulty compounds with Google's structure. A coding model is judged by developers who switch providers in an afternoon, which means Gemini 3.5 Pro cannot ship as merely competent. It has to clear whatever Anthropic and OpenAI shipped last month, and that bar moves while Google's release path is still collecting approvals. The delay and the missing commercial coding product are the same problem viewed from two ends.

What is confirmed, and what is just the rumor mill?

This story has attracted an unusual volume of confident, specific, unsourced claims, and most coverage blends them with the reported facts. Here is the split, because it changes how much weight any of it deserves.

As of this week there is no model card, no pricing page and no gemini-3.5-pro listing in the public Gemini API documentation. Treat every number in the bottom half of that table as speculation until one appears.

How did Google get here?

1. May 19Gemini 3.5 Pro shown at Google I/O June availability targeted on stage
2. JuneTarget window closes with no launch no model card published
3. June 18Noam Shazeer leaves for OpenAI Transformer co-author, Gemini co-lead
4. JuneJohn Jumper leaves for Anthropic Nobel laureate, confirmed by both labs
5. June 22Alphabet has its worst day in over a year shares close roughly 6% lower
6. July 16Bloomberg reports the model is months late 10 current and former employees
7. TBDGemini 3.5 Pro general availability no date confirmed by Google

What it means for the market

Alphabet (GOOGL) already took the hit once. On June 22, after Shazeer's and Jumper's exits landed within days of each other, the stock closed roughly 6% lower in its worst session in about a year, erasing well over $200 billion in market value. Analysts split on whether it mattered: D.A. Davidson's Gil Luria said the departures raised the concern that Google is losing the war for talent at the frontier, while Jefferies' Brent Thill called them noise and kept a Buy rating.

The signal for investors is not the delay. Shipping windows slip constantly and nobody reprices a trillion dollar company over four weeks. The signal is Bloomberg's structural claim: if the drag is a review path wired through Search, Maps and YouTube, it does not resolve when 3.5 Pro finally ships, because the same path gates whatever comes next. Watch two things instead of the launch date. Watch whether Google announces an actual enterprise coding product rather than another model, since that is the gap its own staff named. And watch whether the departures continue, because talent leaving is the market's cheapest available proxy for whether insiders think the structural problem is fixable. This is analysis, not investment advice.

RelatedCursor Builds 'Sand' Agent to Rival Claude Cowork

Who is actually affected?

Less than the headlines suggest, in the short term. If you build on Gemini today, nothing broke: the shipping Gemini models are live and unaffected by one that has not launched. The real exposure is planning risk. Anyone who sketched a roadmap around a June launch has now been wrong twice, and the rational response is to stop treating unannounced Google models as scheduled infrastructure.

The group with genuine exposure is enterprises building coding tools who wanted a Google option. Bloomberg's reporting says that option does not clearly exist, and the delay does not create it. For them the honest read is that this is a two horse market right now, and the third horse is still in the paddock waiting on approvals.

Our take

The most quotable line in Bloomberg's report is about frustrated engineers, and it is the least important one. Every large lab has frustrated engineers. The line that should worry Alphabet shareholders is the boring one about multiple layers of stakeholders, because that is a description of a company whose greatest asset has become its release bottleneck. Google's distribution across Search, Maps and YouTube is exactly why a Gemini launch has to satisfy a dozen constituencies, and exactly why a competitor with one product and no legacy surface can ship on Tuesday.

That is not a problem a better model solves. It is the tax Google pays for being Google, and it is the first time the tax has been legible from outside. We would take the reported organizational critique more seriously than the leaked benchmark numbers, and we would take Google's silence on a launch date more seriously than either. When a company is confident, it announces. When it is months late on its flagship and its own engineers are talking to Bloomberg, it does not.

Original analysis by GenZTech, based on Bloomberg's July 16 report. Disclosure: this piece concerns competition between Google, OpenAI and Anthropic; claims about relative model quality are attributed to Bloomberg's sources and are not our own assessment.