In the month before this episode, the newest wave of companies setting up Vercel’s AI gateway chose Anthropic’s models twice as often as OpenAI’s. That’s not a market-share survey or a self-reported preference — it’s the checkout counter, according to Harpreet Arora, an AI product lead at Vercel, speaking on the MTS podcast. The gateway sits between the labs and the enterprises, so its traffic shows what companies actually buy, not what they say they’ll buy.
But the more interesting story isn’t the 2-to-1 figure. It’s what the same traffic reveals about how enterprises are rebuilding their AI stacks — and why the model war may not have a single winner at all.
Leadership keeps changing hands, and that’s the point
Arora is careful not to call Anthropic’s lead a settled outcome. In her telling, 2026 has been a year of swings: Anthropic’s Opus 4.5 release early in the year set off what she describes as a dominating streak through roughly April or May. OpenAI came back and took share. Anthropic had what she calls a lull in August and September. Then Opus 5.5 — “such a great model,” in her words — plus the follow-on Fable 5.2 brought new teams back to experimenting with Anthropic.
On the OpenAI side, Arora credits Astra with opening up new visual workflows — she cites work with Blender — and singles out the smaller GPT-6 Luna series (now 6.1) as a “sleeper hit” carrying serious volume. Her explanation for Luna’s traction has nothing to do with benchmark tables: developers building fast, snappy mobile apps pick it because it’s quick, cheap, and has a feel they like. It rarely comes up in the flagship-model conversation, but it moves units.
That detail matters for anyone watching the AI trade. Share of headlines and share of requests are two different markets.
The frontier model is a scout, not a workhorse
The most useful framework Arora offers is how a new frontier model actually spreads through a company. Engineers first probe what the new release makes possible — she describes developers trying something complicated in a sandbox they would never have attempted before. Once a workable plan emerges, they hand execution to something cheaper: an open-weight model, or a smaller model in the same family such as Haiku. Customer-support agents, go-to-market pipelines that scrape LinkedIn and scan social signals — anything that becomes repeatable gets pushed down-market once it’s understood.
“You do that a couple of times and then you say, okay, there’s a plan here, let me outsource that to a cheaper model — an open weight model, or a smaller model like Haiku — and that’s going to execute on the plan.”
This is why frontier releases and open-weight share aren’t zero-sum in her data. A strong Opus or Fable release expands the set of things enterprises attempt. Some of that work then migrates to cheaper models — and the cheap tier keeps growing even when the frontier stalls.

A quarter of requests now goes to “decision models”
If you haven’t heard the term “decision models,” you’re not alone — neither had anyone else in January. Arora’s number here is the episode’s most striking: a category that didn’t exist on anyone’s radar at the start of 2026 now handles roughly a quarter of the gateway’s requests.
“Decision models were not a thing on anyone’s radar in January of this year, and now it’s like doing a quarter of our requests.”
Open-weight models, meanwhile, take about a quarter of gateway spend. Both figures point the same direction: enterprises aren’t just swapping vendors, they’re re-architecting pipelines. Asking whether every step needs a general LLM, whether a step is really just a decision, whether part of the app should be a voice agent instead.
MetricSharePeriodNew teams choosing Anthropic vs. OpenAI2x AnthropicOct 2026Requests handled by decision models~25%2026Spend on open-weight models~25%2026Open-weight spend going to US providers~70%2026Teams using voice~1 in 92026
The open-weight money goes to America, even when the models don’t
Here’s a wrinkle that surprises people who assume enterprise buyers care most about which lab built the model. Arora says heavily regulated industries care about where the model actually runs — and that redirects the money.
“There’s also enough enterprises who want the runtime to be in the US, which is why two-thirds of the openweight models that we serve — the share is going to American providers. So yes they came out of China, but 70% of the spend went to American providers.”
In other words: most frontier open-weight models originate in China, but roughly 70% of the spend lands with US providers because that’s where enterprises want the inference to happen. (She uses “two-thirds” and “70%” interchangeably — treat them as the same figure.)

NVIDIA’s retreat is a real loss — and the quality curve has flattened
The episode was recorded on a day of unexplained news at NVIDIA, and Arora treats the company’s pullback from open-weight work as a genuine setback for the US side of that ecosystem. She declines to speculate on what happened. But she notes Jensen Huang had been a major voice for open-weight models at home and abroad, and NVIDIA had been pushing the frontier itself with its Nemotron model and by hiring talent from Poolside. She also points out NVIDIA is a multi-trillion-dollar company — $5 trillion in market value — that is “not going anywhere.” Others are still pushing: Elon Musk’s Grok models, and the rewrite of xAI’s stack from LangChain to its own inference engine.
Her deeper concern isn’t the number of companies. It’s the pace of quality improvement.
“Most of the frontier open-weight models are coming out of China… but we haven’t seen the same level of acceleration in model quality the way we were seeing before.”
She cites Kimi K3 and GLM 5.3 as genuine step changes in open-weight capability — then notes there’s been no equivalent step change in the couple of months since. Her explanation: some frontier labs have developed techniques that let them improve very quickly, and the open-weight side hasn’t reached that point yet. She flags this as her read, not confirmed fact, and names Reflection (whose open-weight model she says is doing well) and Thinking Machines as reasons for optimism about a diverse model market.
Voice is the fastest-moving modality; video is next
Arora calls audio the biggest change of the past month. Inside Vercel she sees engineers working with the command button held down, talking instead of typing, letting the model turn rambling into a coherent narrative. On the gateway, voice has gone from 0% at the start of 2026 to roughly one in nine teams.
Her forward call: video models can’t yet be trusted inside a repeated workflow, and she expects that to change within a year — with audio adoption continuing to climb. She also predicts more diversity among closed-weight models, naming Meta’s Muse models and Musk’s Grok as challengers to the current duopoly.
People don’t pick models on price alone — they pick on personality
Arora’s most quotable argument is that models have personalities, and developers gravitate to them the way people gravitate to colleagues. If agents are meant to stand in for digital knowledge workers, she asks, why would you want every knowledge worker to look and act the same?
“Models have personalities just like humans. If AI models and agents are supposed to represent digital knowledge workers, do you want every knowledge worker to look the same? No.”
That’s why Luna wins mobile-app builders on feel as much as cost, and why benchmark tables don’t fully predict which model a team ends up shipping on. Output style, speed, and how a model responds to the way you phrase things all shape adoption.
The tension worth tracking is between demand and supply in open weights. Arora’s data shows durable enterprise demand — a quarter of spend — and a clear preference for US-hosted runtimes. What she can’t show is a US supplier keeping pace on quality, especially with NVIDIA’s role in question. If the open-weight quality curve stays flat while Anthropic and OpenAI keep shipping, that quarter-of-spend figure could turn out to be a ceiling rather than a floor. Anthropic’s coming IPO — reportedly seeking a valuation north of $1.2 trillion on the Nasdaq under the ticker ANT — will be the market’s first chance to put a standalone public price on a frontier lab, and a weak reception could ripple through every company whose spending plans assume AI demand only goes up. The other number to watch is voice: one in nine teams is fast adoption for a modality that didn’t exist at the start of the year, and Arora expects video to follow the same path within twelve months.