Your Model Choice Just Became a Budget Decision

For a few weeks this summer, the whole industry watched the frontier. Which lab shipped the smartest model. Who Washington would let use it. Meanwhile, the people building actual products walked away from that fight and started running on cheaper stuff. Let me catch you up on what shifted while everyone stared at the top of the leaderboard.

The smart model is not always the one running your app

The volume moved. Vercel's data shows open-weight models handled nearly a third of AI requests on the platform in June. On OpenRouter, the top six models are all open ones from Chinese firms, with Anthropic's Claude Opus 4.7 sitting in seventh. Chinese open-weight models made up 41% of Hugging Face downloads this spring, passing U.S. models.

Here is the pattern underneath those numbers. Closed frontier models are becoming the premium layer, used for hard, high-value tasks. Open models are soaking up the everyday, high-volume work. As Hugging Face CEO Clem Delangue puts it, in a few years the frontier models may be for experimenting and a few high-value jobs, while most production workloads run on private or open source models.

So the question for your roadmap is not "which model is smartest." It is "which model is smart enough for this specific task, at a price I can afford to run all day."

The bill showed up

The cost pressure is real and it is hitting big companies first. Meta shut down an internal token spend leaderboard after AI costs put the company on track for billions in 2026. Uber blew through its 2026 AI coding budget by April. Microsoft canceled Claude Code licenses and pushed engineers onto its own Copilot tool.

Instagram head Adam Mosseri said the plain part out loud on Lenny's Podcast. In a year or two, he said, "the burn rate of a strong engineer might be the same as their salary," and in that world "you're going to probably need to put in some caps." He treats tokens like any other resource he has to allocate, next to payroll and GPUs.

Meta has no caps today. But the direction is set. Token spend is becoming a line item you manage, not a free experiment you cheer on.

Renting versus owning

The reason open models keep winning volume comes down to control. Delangue says his customers increasingly want to own their models instead of renting them through a black box API, a shift that hit home once the bills for scaling closed models arrived. A new repository shows up on Hugging Face every seven seconds, and half of the Fortune 500 use the platform to deploy their own private and open models.

There is a risk angle too. Last month the Trump administration pressured Anthropic and OpenAI to restrict their most powerful models, which raised a nervous point: access to a closed model can be pulled overnight. Microsoft's Satya Nadella warned against single-provider lock-in and argued that control of your own data should come first.

Owning a model is more work than calling an API. But it protects you from a price hike, a policy change, or a provider deciding your use case no longer fits.

The open players are real companies now

This is not a hobbyist movement anymore. DeepSeek is in talks to raise about $1.5 billion at a $71 billion valuation and eyeing a 2027 IPO, a month after raising $7 billion. In June it processed nearly 23% of the tokens flowing through Vercel, against Anthropic's 32%. Chinese labs keep shipping capable open weights, like Z.ai's GLM-5.2, which competes with Anthropic on agentic coding and finding security holes.

U.S. open players are getting funded too. Reflection AI, valued at $8 billion, just signed a $1 billion compute deal with Nebius weeks after a similar deal with SpaceX. The point is simple: open models will keep coming, keep improving, and keep getting cheaper.

The deep cut

Stop treating model selection as one big bet. The teams handling this well pick per task. Frontier model for the hard 5%. Cheaper open model for the high-volume 95%. Mosseri hinted at the discipline: a token budget should be sized to how much you trust a person to spend it in an ROI-positive way, because "it's not that hard to build a token incinerator."

So before your next review, put a number on it. Know your token spend per team, per feature, maybe per engineer. Know which workloads could move to a cheaper open model with no user-facing hit. That is a routing and budgeting exercise you can start Monday, and it will save you more than waiting for the next frontier release.

Three questions for your team

1. Which of our production workloads actually need a frontier model, and which are running on one out of habit? Put a token cost next to each.
2. If a closed provider raised prices or cut our access tomorrow, what breaks, and what is our fallback open model for each critical feature?
3. Do we know our token spend by team or feature yet? If not, who owns getting that number before the next budget cycle?