In February of this year, Google held the most powerful artificial intelligence model on earth. Its Gemini 3.1 Pro led the artificial analysis intelligence index, sitting four points ahead of Anthropic and six points clear of OpenAI, all while operating at less than half the running cost of its rivals. But seven months later, Google has yet to ship a single new Pro model. The flagship Gemini 3.5 Pro, promised in May for a June release, missed its window entirely and remains listed as ‘coming soon’ on DeepMind’s website. Meanwhile, Anthropic has shipped five flagship-tier models, OpenAI has delivered four generations, and a Chinese openweight model named Kim 3 has matched Google’s February benchmark.
The Timeline of a Stalled Frontier
The gridlock did not stem from a lack of capital or audience. Google spent up to $25 billion on AI infrastructure this year, and its Gemini app recently crossed one billion monthly users. Instead, the delay traces back to a series of strategic and organizational bottlenecks. The sequence began on February 19, when Gemini 3.1 Pro claimed the top index spot. At Google I/O on May 19, CEO Sundar Pichai announced Gemini 3.1 Flash alongside Gemini 3.5 Pro, stating the latter was showing strong internal improvements and would ship the following month. June came and went without the Pro model. By July 16, reports surfaced that Gemini 3.5 Pro was falling short of internal targets, particularly in coding capabilities, after a late-June training data fix yielded disappointing results.
On July 21, Google released several low-tier models—Gemini 3.6 Flash, 3.5 Flash Light, and a restricted 3.5 Flash Cyber—while slipping a quiet update into a blog post: ‘We have started our most ambitious pre-training run yet for Gemini 4.’ The next day, on an earnings call, Pichai shifted focus past the unreleased model, saying the company wanted to ‘compete at the frontier level of where the frontier will be when Gemini 4 comes out.’ Days later, Anthropic released Claude Opus 5, taking the top spot on the index with a score of 61, while Google’s newly launched 3.6 Flash scored 50, falling out of the top ten behind competitors like OpenAI, Meta, XAI, and Chinese labs. Though Google continued to release rapid updates, including Gemini 3.8 Flash which reached a score of 59, it remained behind Anthropic’s Fable 5.1 at 66 and OpenAI’s newly announced GPT6.
Three Interlocking Roadblocks
Industry analysts point to three simultaneous challenges that derailed Google’s momentum. First, the company underinvested in agentic coding—the primary capability driving enterprise adoption. Menlo Ventures reports that Anthropic commands 54% of enterprise spend on AI coding, followed by OpenAI at 21%, and Google at just 11%. Despite an April memo from co-founder Sergey Brin urging DeepMind to close the gap in agentic execution, DeepMind leader Demis Hassabis remained focused on broader artificial general intelligence (AGI) milestones rather than language model coding. While Google built ‘Anti-gravity,’ its agentic coding platform, the tool was designed as a harness that runs on whatever model a user plugs in, including Claude and GPT, rather than enhancing Gemini itself. Organizational redundancy further compounded this; Bloomberg reported that Google Cloud, DeepMind, and the Android team were simultaneously developing separate AI coding tools, duplicating work while competing for the same limited compute resources.
The second obstacle involves compute distribution. While Google possesses a structural advantage by manufacturing its own Tensor Processing Units (TPUs), it increasingly rents this infrastructure to direct competitors. In October, Anthropic signed to rent one million Google TPUs, followed by a May announcement that Anthropic committed $200 billion over five years to Google Cloud. This single contract represents roughly 40% of Google Cloud’s entire disclosed backlog. Meta also secured a multi-billion dollar TPU rental agreement. Inside Google, this created intense internal friction. Reuters reported in August that compute constraints and disagreements between Gemini’s project leaders starved critical areas like coding of resources, as engineers regularly hit capacity limits when trying to use AI to write code. While Google Cloud’s commercial metrics soared—growing 82% last quarter to $24.8 billion with a $514 billion backlog—the company essentially leased its primary hardware advantage to the competitors outstripping its own models.
Third, Google experienced an unprecedented talent drain at the highest levels. Within one week in June, Gemini co-lead Noam Chazir departed for OpenAI, and Nobel laureate John Jumper left for Anthropic, followed shortly by two other Gemini researchers. This executive exodus contributed to Alphabet stock falling approximately 5% on June 22, erasing a quarter of a trillion dollars in market value. On August 5, chief scientist Jeff Dean—a 27-year Google veteran—departed alongside Sanjay Gowat, Gemini co-lead Oriel Vignyels, and Google Brain founder Quaca to launch a new venture, Discovery Loop. On the same day, Demis Hassabis stepped back from the CEO role at DeepMind to become chair, handing day-to-day control of Gemini to Kuray Kavuklu, who reports directly to Pichai. DeepMind’s hiring-to-departure ratio plummeted from 12:1 in 2023 to 2:1 this quarter, while Anthropic maintained a 22:1 ratio. Departed staff cited burnout from 60-hour workweeks, pre-IPO equity incentives at rivals, and internal friction over a London union bid triggered by a Pentagon contract that placed Gemini on classified networks.
Distribution Versus Deliberate Choice
Despite lagging behind on index benchmarks, Google has maintained massive user metrics. The Gemini app expanded from 400 million users in May to over one billion in August, making it the fastest-growing product in Google history. AI overviews in search reach 2.5 billion monthly users, and Apple’s upcoming Siri features in iOS 27 are powered by models Apple co-developed with Gemini. Furthermore, Google has secured over 8 million paid Gemini Enterprise seats, and its API systems process 22 billion tokens per minute. For many enterprises surveyed by Andreessen Horowitz (a16z), current models are ‘good enough,’ allowing Google’s highly economical Flash models, such as Gemini 3.8 Flash, to dominate on price.
However, the metrics for deliberate choice show a different trend. On Open Router, a platform where developers choose which models to run, the market share of major US labs, including Google, fell from 70% to roughly 30% over the past year. Menlo Ventures’ enterprise API spend tracking shows Anthropic at 40%, OpenAI at 27%, and Google at 21%. Meanwhile, AI text detector Pangram reported that Google’s share of analyzed text collapsed from 12% to 1.9% this July. While distribution wins casual users, developers and enterprise engineers are increasingly embedding competitor models into their workflows—a highly sticky layer where only one in ten builders switch vendors annually.
The Business Model Dilemma
This cycle of falling behind and catching up is not new for Google. In November 2022, the launch of ChatGPT prompted Google to declare a ‘code red.’ Its initial rapid responses failed: Bard, launched in February 2023, contained a factual error in its demo that wiped out $100 billion in market value in a single day, and Gemini 1 in December 2023 required its image generator to be pulled within two months. However, when Google took its time, it succeeded. Gemini 2.5 Pro claimed the top spot on LM Marina in March 2025, and the viral success of Nano Banana and Gemini 3 in November sent Alphabet shares up 8%, prompting OpenAI’s Sam Altman to declare his own ‘code red’ on December 2.
The underlying delay points to a fundamental business model trap. For startups like Anthropic and OpenAI (with annualized revenues of $65 billion and $40 billion respectively), the model is the sole business, making frontier performance an existential necessity. For Alphabet, the model is merely an input. Search represents 54% of Google’s revenue, and advertising accounts for roughly 70%. A highly advanced AI that directly answers queries reduces search result page views and, consequently, ad clicks. While a better model is pure financial upside for startup labs, it represents a direct trade-off for Google’s core business. This dynamic risks creating a comfortable drift, similar to Nokia’s position in 2007 when strong distribution numbers masked a product line that was falling behind the market. As the industry awaits the launch of Gemini 4, the real test of Google’s rebuild will not just be its position on the leaderboard, but whether it has resolved its internal battles over compute allocation, coding development, and key talent retention.