Google rolled out Gemini 3 Ultra to all Workspace tiers on Thursday, including a 10 million token context window that the company claims is now the largest in any production-grade model. The move arrives roughly six weeks after Anthropic's Claude Opus 4.7 launch and feels timed to recapture the narrative around frontier capability after a quiet first quarter.
Ten million tokens is roughly equivalent to fifteen full-length novels, an entire mid-sized codebase, or about 20 hours of transcribed video. The practical question — and one Google's marketing has so far avoided answering directly — is what fraction of that context the model can actually reason over with high fidelity. Independent benchmarks from labs at Stanford and Anthropic's own Long-Context evaluation suite have historically shown sharp accuracy drops past the 200K mark, even on models that nominally support longer windows.
Why Google made it free
The most striking part of the announcement is not the technical claim but the pricing. Workspace customers — all 3 billion of them, across personal Gmail and enterprise — get full Gemini 3 Ultra access without an upcharge. Compare this to OpenAI's GPT-5, which gates the comparable feature behind a $200/month Pro tier, or Anthropic's Claude Opus 4.7, which is similarly metered.
The reason is straightforward: Google has the only AI distribution channel of its size, and is willing to operate the inference at a loss to keep users from forming habits with competitors. Internal estimates from a person familiar with the financial modeling put the cost of running Gemini 3 Ultra at scale across Workspace at roughly $1.4 billion per quarter — a figure that, in any other company, would be a board-level concern. At Google, it's a rounding error against ad revenue.
What this means for the market
Distribution-as-strategy is going to be increasingly difficult for OpenAI and Anthropic to counter. Both companies have built differentiated technology, but neither has the user base of Workspace, the device base of Android, or the enterprise relationships of Google Cloud. The likely response, already telegraphed by both labs, is to compete on quality and depth of integration rather than absolute capability — a framing that, if it holds, would represent a meaningful shift in how the AI industry organizes itself.
