Open models run 56% of AI tokens. Anthropic still gets 64%

The claim that "local models are largely irrelevant in business" is "completely wrong," @kimmonismus wrote this morning in a post about a Financial Times report on companies moving work off the big AI labs.

The FT numbers are hard to wave off. According to the report, mentions of open models on US earnings calls rose sixfold over the past year. AT&T now runs about 40% of its AI workloads on open models and wants that at 70% within a year. The post adds the detail that makes the math concrete: AT&T pushes about 45 billion tokens a day, so small per-token savings add up to real money.

Open-weight models are ones whose underlying files are published, so anyone can download them, run them on their own hardware, or pay a cheap host to run them. Think Kimi, Qwen, DeepSeek and their peers. Closed models, like Claude, GPT and Gemini, are only available through the company that built them.

The chart that started it

The FT's other headline number comes from Vercel, whose AI Gateway sits between thousands of apps and the model providers they call. Vercel's September production index found open-weight models handled 56% of gateway tokens in August. In December they handled 7%.

Vercel CEO Guillermo Rauch had already been watching the curve. On September 19, @rauchg posted that it looked like "a record day for token volume % of open models," with open models at 78.4% of traffic that day and closed models at 21.6%. That's one day, not a trend line, but it points the same way as the monthly data.

Most tokens, not most money

The same Vercel report has a number that gets less attention. Anthropic still took 64% of all spending on the gateway in August. Average token cost across the gateway fell 23.2% that month, the third monthly drop in a row.

So open models win on volume and closed models win on dollars, which makes sense once you think about how teams split the work.

My guess is that the high-volume, low-stakes jobs are the ones moving: sorting, tagging, summarizing, pulling fields out of documents, first drafts. The jobs where a wrong answer costs money (a tricky code change, a legal question, an agent that touches a customer's account) stay with the frontier labs and pay their prices. The kimmonismus post puts it well: customers are "testing how much of their work actually needs a frontier model."

I read this as routing more than defection. Companies seem to be sending the repetitive bulk of their traffic somewhere cheaper and keeping the expensive model for the hard parts. The FT's story about pressure on OpenAI and Anthropic holds up on volume. On spend, the labs still have the work people pay the most for.

What this means for a small business

You probably aren't pushing 45 billion tokens a day. The lesson still applies, just at a smaller scale.

Most small businesses that use AI today pay for one tool and send everything through it. That's fine when the bill is $20 a month. It stops being fine once you automate something that runs hundreds of times a day, like reading every inbound email, tagging every lead, or summarizing every call. A cheaper model can often handle that kind of job, and that's where the savings show up.

A few questions worth asking before your AI bill starts to climb:

• Which of your AI tasks are repetitive and easy to check, and which ones actually need judgment?
• If a cheaper model got one of those repetitive tasks wrong, would anyone notice before it cost you?
• Is your setup tied to one vendor, or could you swap the model behind a task without rebuilding it?

The last question is the one I'd push on. Vercel's report shows how fast the money moves: Opus 5 tripled its spend share in a month, and GPT-6 Astra took a third of OpenAI spending on the gateway within 48 hours of launch. If switching models means rebuilding your workflow, you'll stay on whatever you picked first, even after it stops being the best deal.

At New Face Design we build automations so the model can be swapped out without touching the rest. If you want to know which of your processes are worth automating and what model each one actually needs, our free process audit is a good place to start.