Model Routing: Why Cost, Not Capability, Is Driving Enterprise AI Strategy in 2026

Key takeaways

• Frontier model prices are climbing "double or quadruple" versus prior generations on a per-token basis, according to Glean CEO Arvind Jain, and users are running longer tasks on top of that.
• One enterprise AI vendor claims routing cuts its per-task cost to $0.45 versus $1.84 for a comparable frontier-model-only workflow, roughly a 4x gap.
• Enterprises are increasingly refusing to rely on a single model provider, and interest in open-weight models has surged in the last three months as a hedge against frontier pricing.
• A pre-routing filtering layer, deciding what context an LLM even needs, can cut both latency and token spend before a model is ever called.
• For SaaS builders, the practical takeaway is the same one enterprises just learned the hard way: your model mix, not your model choice, determines your margin.

Why is model routing suddenly a big deal?

Model routing is not a new idea, but the economics behind it changed quickly. Glean CEO Arvind Jain told Latent Space that "AI models have been getting expensive... on a per token basis, they're more expensive, sometimes double or quadruple the rates of the previous models." Combine that with users running longer, more complex tasks on the newest models, and per-user AI spend can climb 10 to 20x year over year even before headcount grows.

That's the same dynamic every SaaS founder pricing an AI feature has to model: token cost per request is only half the equation. Task length and model choice multiply on top of it, so a single unrouted "always use the frontier model" policy can quietly wreck a feature's margin as usage matures.

How does routing actually save money?

In Glean's case, routing happens on three levels: end users can pick a model, admins can restrict or cap model access, and an automatic mode routes each task to the model likely to handle it best for the lowest cost. Glean's engineering lead has claimed the system runs at roughly $0.45 per task versus $1.84 for a comparable Claude Code-style workflow, a claim illustrative of the kind of gap routing can create rather than a universal number every team should expect.

The bigger lever, though, sits upstream of routing: Glean's Waldo layer reportedly assembles the context a query needs before any model is called, which the company says cuts latency by half and tokens by a quarter. In other words, the cheapest token is the one you never spend. A smaller model with the right context, in this framing, can beat a frontier model loaded with irrelevant data.

Why are open-weight models suddenly part of enterprise strategy?

According to Jain, interest in open-weight models inside large organizations was "minuscule" as recently as last year, held back partly by the perception that many of the strongest open models were developed outside the US. That changed within the last three months, as frontier pricing made maintaining a frontier-only AI budget "untenable" for some businesses. Open-weight models are, in Jain's words, "an order of magnitude cheaper" for many tasks, and no enterprise wants to depend on one or two providers anymore.

For anyone modeling unit economics on an AI product, that's a signal worth taking seriously: provider diversification isn't just a resilience play, it's becoming a cost play.

What does this mean if you're pricing an AI product?

If you're a founder shipping an AI feature, the enterprise routing story is really a preview of a decision you'll eventually face too: do you hard-code one model per feature, or build in the flexibility to route by task, user tier, or cost ceiling? Modeling that trade-off is exactly the kind of scenario worth running before you commit to a pricing tier, since the gap between "always frontier" and "routed" can be the difference between a healthy margin and a subsidized feature.

Model routing is no longer an infrastructure nicety, it's a direct response to a cost curve that keeps climbing faster than most product roadmaps account for.