Vercel gateway data shows open models at a record 78.4% of tokens

Vercel's AI Gateway data claims open-weight models hit a record 78.4% of token volume, and X spent the week treating that number like a trophy. The claim is real and comes from Vercel's own published index. Whether one gateway's sample proves anything about the market is the argument underneath it.

Key facts

• Vercel's AI Gateway Production Index reports open-weight models at a record 78.4% of token volume, with the snapshot corroborated by an independent recap roughly a day old.
• The same index claims combined spend on Moonshot AI, DeepSeek, and Z.ai models passed OpenAI through that gateway — a claim amplified by @Jason.
• The driving post from @rauchg drew over 1.32 million views, 1.3k+ likes, and 170+ quotes, which is what turned a routine index update into a discourse event.
• The recap's counterweight: on the same data, Anthropic still captured a reported 64% of spend, because volume is not revenue.

What's driving the conversation

Guillermo Rauch (@rauchg) started it by posting the index snapshot, and the victory-lap reading spread fast: open models are not a developer-toy segment anymore, they are where production tokens run. The amplifiers mostly took the number at face value. @Jason framed it as a market-share milestone — open-lab spend overtaking OpenAI — and accounts like @sovereignbrah and @AIHegemonyMemes ran it as a bigger narrative about closed labs losing the volume war.

The pushback came from people reading the same chart more carefully. Token volume measures how many tokens moved, not how much money changed hands, and cheap models generate volume almost by definition: if your tokens cost a fraction of a frontier model's, your share of volume inflates even at a small share of spend. The daily.dev recap carries the same caveat — most tokens, but Anthropic still capturing the majority of revenue through the same gateway. There is also a sampling question: Vercel's gateway serves a particular customer base, heavy on US web-development workloads, so the index describes Vercel's traffic before it describes the market.

The substance

The verifiable part: Vercel publishes its gateway index as a monthly report, and the 78.4% figure is in Vercel's own post. This is company-published data, not an independent audit, and it describes one gateway's traffic. Within that frame, it is a real data point: on Vercel's infrastructure, open-weight models carry most of the tokens.

Everything beyond that frame is interpretation. "Open models won" is not what the data says; it says that on one US-skewed gateway, tokens are cheap and open. The spend crossover (Moonshot + DeepSeek + Z.ai > OpenAI through this gateway) is a claim from the same source, and revenue share tells the opposite story on the same page. Both can be true at once: open models dominate volume, closed labs dominate revenue. That split is the actual story, and the discourse mostly flattened it.

Why builders are watching

For anyone deciding where to run workloads, the useful signal is narrower than the headlines: production traffic through one major gateway is increasingly open-weight, which means tooling, evals, and operational maturity around open models keep improving. Our open-model API roundup tracks where those models are servable, and the September price index tracks what they cost across providers. Watch how the volume-versus-revenue gap evolves in Vercel's next index, and whether other gateways publish comparable numbers — right now the market-share narrative rests on one company's self-report.