Stripe paid $7B for a router. Glean is selling the choice not to call a model at all

Stripe has bought OpenRouter for more than $7B ^[1], and at Glean the model-selection feature customers actually use is the automatic one, chosen mostly for economic reasons ^[6][7]. Both facts point at the same layer: the thing that decides which model handles a request, or whether a model is needed at all, is where enterprise AI spend is now controlled ^[8].

Arvind Jain, Glean's co-founder and CEO and a former Google Distinguished Engineer ^[4], gave Latent Space the arithmetic behind that. The newest frontier models cost double or quadruple the per-token rate of their predecessors, and because they can run longer tasks, users run longer tasks, so per-user spend can land at 10 to 20 times last year's level ^[9]. Nothing about that curve is fixed by picking a cheaper vendor. It is fixed, if at all, by not sending every request to the expensive tier.

Glean exposes three levels of control: employees can pick a model explicitly, administrators can restrict models or impose usage limits, and automatic mode selects per task ^[6]. "Why are people talking about model routing? Why are they excited about it? It's mostly because of cost," Jain said ^[5]. The sharper version of the same idea is the null route. "A big goal of Glean is to avoid using LLMs for tasks where we don't need them," Jain said, citing queries where someone is adding or multiplying two numbers and could have used a calculator ^[8].

The commercial claim attached to this is single-sourced and worth reading carefully. Glean engineering lead Tony Gentilcore recently said the product is "4x more cost-effective" than Claude Code, averaging $0.45 per task against $1.84 for Claude Cowork, which he attributed to Glean's harness and routing ^[10]. Those two figures divide out to about 4.1x ^[17]. No task mix, sample size, or quality comparison was given in the source, and a per-task average says nothing about how many tasks a routed system needs to finish the same work.

What is harder to copy is the observation data. Glean reports $300M in ARR this year, a three-fold increase over 15 months ^[2], which implies roughly $100M 15 months earlier ^[18]. The company was last valued at $7.2B after a $150M Series F ^[3], meaning Stripe paid about what the private market last put on Glean as a whole ^[20]. Deployment breadth is the point: Zillow reports 80% adoption across 7,000 employees ^[11], roughly 5,600 people ^[19], and Booking.com describes Glean as its first company-wide AI platform ^[12]. Jain says that gives Glean a view of which model people reach for first and when they escalate after an unsatisfying answer, which feeds back into routing ^[13].

Latent Space frames the demand as a product of frontier-model competition plus increasingly capable open-weight models such as Kimi K3 and Qwen3.8-Max ^[14]. Jain's own pitch is a meta-harness: "You can think of Glean today as a superset of ChatGPT, Claude, Gemini, Grok" ^[15]. The architecture also includes a model called Waldo ^[16].

Watch whether any routing vendor publishes quality-held-constant cost data rather than per-task averages, and whether Stripe's OpenRouter price shows up as pass-through pricing pressure on the enterprise routers.