Analysis
Sundar Pichai confirmed Google has begun what he called its "most ambitious pretraining run yet" for Gemini 4 -- a training effort Google says is the largest single pretraining project in the company's history. The announcement came alongside a more uncomfortable admission: Gemini 3.5 Pro, the nearer-term model investors and developers have been waiting on, has now been delayed three times, and Pichai directly acknowledged Google currently trails rivals on AI coding capability, the specific weakness slowing that release. Google said it will ship Gemini 3.5 Pro "as soon as it's ready" rather than commit to a new date.
The framing Pichai used for Gemini 4 is the more interesting strategic signal. He said Google is building the model to compete at "the frontier level of where the frontier will be when Gemini 4 comes out" -- an explicit statement that Google isn't training to match OpenAI's GPT-5.6 lineup or Anthropic's Fable today, but is instead trying to anticipate where the competitive frontier sits by the time its own model actually ships. That's a materially different posture than the coding-gap admission around Gemini 3.5 Pro suggests: Google is conceding it's behind in the near term while betting its next flagship leapfrogs the field entirely once it arrives.
The competitive backdrop makes both admissions higher stakes than they'd otherwise be. OpenAI's GPT-5.6 shipped as three separate models -- Sol, Terra and Luna -- rather than a single release, and Moonshot AI's Kimi K3 has become the most capable open-weight model released to date, reportedly matching or beating closed frontier systems from both Anthropic and OpenAI on public benchmarks. A Google that publicly concedes a coding gap on its near-term model, while it's simultaneously the incumbent search and cloud giant, hands both established rivals and rapidly improving open-weight labs a concrete opening to court developers on that exact capability.
“The framing Pichai used for Gemini 4 is the more interesting strategic signal.”
The disclosure also can't be read apart from Alphabet's spending story. The same week Pichai discussed Gemini 4's ambitious pretraining scope, Alphabet raised its 2026 capex guidance to as much as $205 billion and reported its first-ever negative free cash flow, driven almost entirely by AI infrastructure buildout. Training "the largest pretraining run in company history" is not a free-standing R&D decision -- it is a direct claim on the same capital budget already drawing scrutiny from investors questioning whether AI infrastructure spend across the industry is converting into revenue fast enough.
For founders and investors building on top of Gemini's API and enterprise stack, the practical read-through is a near-term one: Gemini 3.5 Pro's delay means the coding-capability gap Pichai admitted to is a live constraint today, not a solved problem, and teams choosing a foundation model for coding-heavy products should weight that against Anthropic's and OpenAI's current offerings rather than assume Google's gap closes on any predictable timeline. Gemini 4's more ambitious, frontier-anticipating design is a longer-horizon bet that won't affect product decisions being made this quarter.
The bear case: publicly conceding you're behind on a specific capability, three delays deep on a model already promised to the market, is a real credibility cost regardless of how ambitious the next model's training run is described as being. "Ship when it's ready" is the same language companies use right before further delays, and Gemini 4's frontier-anticipating framing is unfalsifiable until the model actually exists and can be benchmarked.
Watch whether Gemini 3.5 Pro actually ships in the coming weeks and how its coding benchmarks compare to GPT-5.6 and Kimi K3 at release, whether Google discloses a rough timeline for Gemini 4 given the scale of the training run just disclosed, and whether Alphabet's next earnings call ties Gemini 4's training costs explicitly to the capex guidance already unsettling investors.
