Token minimization is the new platform skill

The lead AI story in the NYT today is a quiet one. “Tech Workers Maxed Out Their A.I. Use. Now They’re Trying to Minimize It.” The summary in one line: the companies that spent 2025 wiring AI into everything are now spending 2026 figuring out how to do the same work with fewer tokens, smaller models, and tighter context windows. The bill came due.

The most useful counterpart write-up landed on Hacker News this morning. Alex Ellis’s “Local Qwen isn’t a worse Opus, it’s a different tool” is a careful, receipts-included account from a small software business. The honest read: local open-weights models are no longer toys, and they pay for themselves quickly on a real subset of workloads. They are not a drop-in replacement for the frontier, but they are an excellent fit for the everyday work that doesn’t need a 0.5T-parameter giant.

Put the two together and the picture for any team shipping AI in 2026 is the same picture. The throughput-per-dollar curve has finally caught up with the capability curve, and the next round of competitive advantage will come from the teams that route the right work to the right model at the right size, not the teams that pipe everything to the most expensive endpoint.

This is exactly the lever Microsoft’s Agent Optimizer in Foundry Agent Service is built to pull. The model-target optimization mode explicitly searches across deployment options for the best quality-cost tradeoff, runs the candidate against your eval set, and gives you per-task breakdowns with token costs before you promote a winner. That is a real cost-control surface, not a vibes one. Pair it with Foundry Local for the workloads that should never touch a remote frontier model in the first place, and you have a credible cost story.

Three practical reads for anyone running an agent program right now.

• Cost-per-outcome is the metric, not cost-per-token. Two agents can serve the same workload at radically different unit economics depending on how often they escalate to frontier models, how much context they reload per turn, and how aggressively they cache. Instrument the outcome, then optimize the path to it.
• Treat model routing as a first-class architectural decision. A workload that uses a small local model for triage, a mid-tier hosted model for synthesis, and a frontier model only for the genuinely hard turns is going to beat a single-model architecture on cost and resilience. Build for it now; retrofitting it later under board pressure is no fun.
• For SI and ISV partners, packaged cost-optimization offers are a defensible product. A consultancy that can land in a customer’s existing Foundry agent, run an optimization cycle against the customer’s evals, and ship a documented “you spent $X on tokens last month, here’s a $0.6X version with the same quality” engagement has a much better story than “we’ll write you a smarter system prompt.”

The flashy AI work will keep being model launches. The actual money is going to come from the boring engineering work of using less of what’s already there. Worth getting fluent now.