David Friedberg Calls Open Models Anthropic's Biggest Risk

On the All-In Podcast, David Friedberg put a single chart on screen and called it the biggest risk factor Anthropic faces. The chart came from Vercel's AI Gateway, which routes production AI traffic across the major model providers. It tracks one thing: the share of tokens running on models that publish their weights for download, against everything else.

In late June, closed models carried roughly seven of every ten tokens on that gateway. By mid-September the bars had flipped. The Vercel chart showed open-weight models handling 78.4 percent of the volume. Closed models were down to 21.6 percent. Twelve weeks.

"I don't think we've ever seen a tidal wave" like it in the history of technology markets, Friedberg said. He has watched a lot of technology markets.

Read the Chart Properly

A single day on one gateway is a noisy number, and that 78.4 percent was a record day, not an average. Vercel's own monthly AI Gateway Production Index gives the steadier read. In June, open-weight models carried 29 percent of tokens. In August, 56 percent. Vercel widened its definition of open-weight in the September edition, so part of that jump is accounting. Most of it is not.

The same Vercel index holds the line that matters most. In August, open-weight models took 56 percent of the tokens and 14 percent of the spend. Closed models still collected 86 cents of every dollar.

So two things are true at once. Volume has moved. Money has not, yet.

Friedberg's Real Argument Is a Split

The headline says Anthropic is in trouble. The argument underneath is more precise, and it is the part operators should take seriously.

Friedberg's read is a bifurcation. Most AI work, the summarising, classifying, drafting, routing and extraction that fills a normal business day, is moving to open-weight models because they are now good enough and cost a fraction of the price. And it isn't only text. Open weights are showing up in image and video generation and in the vision-language-action models that drive robots.

The premium models keep the narrow, hard end: serious engineering, mathematics, biology, problems where being slightly better is worth a great deal. For that work, he said, people will pay "close to infinity dollars." Anthropic's market isn't disappearing. It is shrinking to the work where quality is the whole point, and the everyday volume goes elsewhere.

The Top Five Open-Weight Models Right Now

Ranked by the Artificial Analysis Intelligence Index for open-weight models, as of 13 October 2026:

1. MiMo-V2.6-Pro (Xiaomi). Released 21 September. A mixture-of-experts model with 1 trillion parameters, 42 billion active per token, a 1 million token context window, and an MIT licence. It scores 46 on the Artificial Analysis index, first of 116 open models, at $0.43 per million input tokens.

2. GLM-5.3 (Z.ai). Released 18 August. 753 billion parameters, 40 billion active, 1 million token context. Artificial Analysis scores it 45. It runs under Z.ai's own licence, which allows commercial use with restrictions, and it is expensive for an open model at $1.40 per million input tokens.

3. Kimi K3 (Moonshot AI). Released 16 July. The largest of the five at 2.8 trillion parameters, 104 billion active, and it reads images as well as text. Artificial Analysis scores it 44. Moonshot's licence allows commercial use with restrictions. It is also the slowest here and the dearest at $3 per million input tokens.

4. GLM-5.3-Flash (Z.ai). Released 26 August. 320 billion parameters, 18 billion active, MIT licence. Artificial Analysis scores it 42 at $0.15 per million input tokens, which makes it the value pick of the list.

5. DeepSeek V4.1 Flash (DeepSeek). Released 10 September. 552 billion parameters, 16 billion active, MIT licence. Artificial Analysis scores it 39, and it runs at over 200 tokens a second, second-fastest of every open model tracked, for $0.30 per million input tokens.

Three things stand out. Four of the five come from Chinese labs. Three ship under MIT, which means you can download them, run them on your own hardware and build on them commercially without asking anyone. And the American open-weight names people still quote, Llama and Mistral, are not in the top ten.

The gap to the frontier is real: the closed leaders still score higher on the same index. But these models cost a fraction as much, and for most business work the difference doesn't show.

How We Run It

We already split our own stack the way Friedberg describes. The fast, frequent, low-stakes calls, the instant acknowledgement when a message lands and the semantic search across our entire knowledge base, run on a small open-weight model on our own Mac mini. The cost is zero and nothing leaves the building. The work that has to be right, writing code, reviewing it and making decisions inside client builds, runs on a premium model, and we pay for that gladly.

That split is the design question every operator now has to answer, task by task. Not "which AI do we use" but "which of our tasks are commodity and which are premium." Get it wrong one way and you pay premium prices for summaries. Get it wrong the other way and you put a cheap model on the work your reputation rides on.

One rule we hold without exception: an open model never comes in without rails. Downloading the weights is easy. The guardrails, the permissions and the checks that stop a model doing something it shouldn't live in your own Cystem, not in the model. Anyone who skips that step to save money is taking a risk they haven't priced.

The Takeaway

Friedberg is right about the direction and right to call it a risk for anyone whose business is selling tokens. For everyone else it is the opposite. The cost of good-enough intelligence just fell off a cliff, and the operators who sort their work into commodity and premium this quarter will run the same output for a fraction of the bill.

The chart flipped in twelve weeks. Your stack can flip in one.

Sources: David Friedberg, All-In Podcast (September 2026); Vercel AI Gateway open vs. closed token volume chart; Vercel AI Gateway Production Index, July and September 2026 editions; Artificial Analysis open-weights leaderboard and model pages, retrieved 13 October 2026.