Analysis

Open-Weight Models Hit 56% of Gateway Tokens but 14% of Spend: What the Gap Means

Vercel data shows open models winning on volume while closed labs keep the revenue. The gap says a lot about where AI money is going.

Elena Vasquez

Former ML Researcher, Industry Analysis Lead

Published 7 min read
Close-up of a stock market trading chart with indicators for financial analysis.
In this story 8 sections

Open-weight models processed 56% of production tokens on Vercel's AI Gateway in August 2026 but only about 14% of estimated spending. The gap exists because open tokens cost far less than closed ones. Volume is moving to cheaper models faster than revenue is, and that shapes who profits from AI.

This analysis is for AI buyers, investors, and builders who want to understand where inference money is going. Emergent Wire covers the AI industry, and we read the Vercel numbers alongside two outside analyses to see what holds up.

What Do the Vercel Numbers Show?

They show a rapid shift in volume. Techstrong's analysis of Vercel's August data says open-weight models handled 56% of production tokens, up from 36% in July and under 10% in December 2025. Those same models accounted for only 14% of estimated spending.

August 22 was a record day at 62% of traffic, according to Vercel CEO Guillermo Rauch. Constellation Research adds earlier points: about 11% in April and 29% in June, when open models produced under 4% of spend.

A blurred view of a contemporary office hallway showcasing modern architectural design and lighting.

The trend line matters more than any single day. Share has risen in nearly every period reported, and Constellation notes that over the latest two-month period open weights made up about half of all token volume.

The price side moved too. Techstrong reports that the average price per token on the gateway fell 23.2% in August. Among teams using more than 10 million tokens in both July and August, the median price per token dropped 7.6%, versus a 2.9% decrease in July. In plain terms, teams are getting more inference from the same budget, which is how Vercel's own report put it.

Why Is Volume Rising So Much Faster Than Revenue?

Price is the main reason. Techstrong estimates that closed-weight tokens cost about 7.8 times as much as open-weight tokens on average. The author labels that a back-of-the-envelope figure, and it is his own calculation from Vercel's data, not a number Vercel published.

Agents are a second driver. Constellation points out that agents make many model calls per task, so small per-call cost differences compound quickly. A coding agent that makes 50 calls to finish a task will feel a price gap far more than a chatbot that makes one.

A third driver is multi-model use. Constellation reports that teams handling more than 10 million requests a month averaged 35 models. Large users already route work across many models, and adding a cheap open model for routine steps is a small change.

There is a subtle point inside these numbers. A lower price per token does not always mean lower total spending. If cheaper tokens let teams run more agents, call models more often, and process larger documents, total token use can rise faster than prices fall. Both effects can be true in the same quarter, which is why overall AI spending can grow even as unit prices drop.

Who Is Still Making the Money?

Closed labs, and especially Anthropic. Techstrong reports that Anthropic received 64% of estimated August spending on the gateway, and that it has taken at least 61 cents of every dollar spent through the gateway every month since December.

Inside Anthropic's lineup, buyers traded down. Techstrong says Fable 5's share of gateway spending fell from 13.2% in July to 4.9% in August, while Opus 5 rose to 22.5% of spending. Vercel says Opus costs about half as much per token as Fable, and that nine in 10 teams using Fable reduced their usage.

Desk setup showing calculator, cash, coins, and financial notes for budgeting.

That tells us something important. Buyers are not only moving to open models. They are also moving to cheaper closed ones. The pressure is on price at every level of the market, which fits what we saw in our look at why AI model prices keep falling.

Constellation offers a useful earlier benchmark for the concentration of money. In June, it says, Anthropic had 32% of token volume but 61% of spend, while open-weight models had 29% of volume and under 4% of spend. By August the volume share of open models had nearly doubled and Anthropic's spending share had edged higher, which shows how sticky premium revenue can be.

How Close Are Open Models to Closed Ones in Quality?

Close on many tasks, but not equal everywhere. Techstrong cites Mozilla's State of Open-Source AI 2026 report, which puts the capability gap between leading open-weight and closed models at about 3.3%. It says open models are at or near parity on coding, instruction-following, and general knowledge.

Constellation names Qwen, DeepSeek, GLM, and Kimi as the open-weight families driving growth, and notes that Chinese labs are setting much of the pace, often under permissive licenses. It also reports more than 151,000 downstream Qwen derivatives on Hugging Face.

A small quality gap can still matter. For a hard reasoning task, a few points of accuracy can decide whether an agent finishes a job or fails halfway. That is why many teams keep a frontier model for the hardest steps. For a deeper comparison, see our overview of open-weight versus closed models in 2026.

How Much Should You Trust This Data?

Trust it as a direction, not a market census. Techstrong notes that Vercel's data covers anonymized aggregate traffic through its own gateway only. It does not capture direct API use, private self-hosted deployments, cloud-provider endpoints, or all enterprise contracts.

The spending figures are also estimates. Techstrong says they are calculated from published list prices, not from invoices, so negotiated discounts and caching are not reflected. Constellation says the data should not be treated as overall market share.

The 62% record also needs care. One day is not a trend, and weekends can skew toward experimentation. The monthly figure of 56% is the sturdier number, and it still tells a clear story.

Close-up of diverse hands shaking indoors, symbolizing partnership and agreement.

Another question is who the gateway's customers are. A gateway such as Vercel's tends to attract developers building applications on the web, who value flexibility and easy switching between models. Large enterprises with direct contracts, private clouds, or on-premises systems may behave differently. Their usage would not show up here, so the open-weight share across all buyers could be lower or higher.

What Does the Gap Mean for AI Labs?

It means volume leadership does not equal profit leadership. Constellation says open models may win volume before they win significant revenue, and that economic value is spread across model developers, inference providers, cloud providers, hardware vendors, gateways, and firms that specialize models.

Managed inference providers are one beneficiary. Constellation names Together AI and Fireworks as services that let customers use open models without running GPU infrastructure. A company can capture revenue by hosting open models well, even if it did not train them.

Closed labs face a squeeze. They must keep a quality lead large enough to justify a price that is many times higher. Our look at AI inference cost per million tokens in 2026 shows how wide that price spread has become.

What Should Buyers Do With This?

Buyers should route by task. Send hard reasoning to a frontier model and send routine work to a cheaper one. Constellation lists classification, extraction, repetitive agent steps, and coding as candidates for cheaper models where they are capable enough.

Start with a short plan:

  1. List your AI workloads and tag each as hard, medium, or routine.
  2. Test a cheaper model on the routine tier with your own prompts.
  3. Measure quality, latency, and total cost per finished task, not per token.
  4. Consider a managed provider before you self-host an open model.
  5. Re-run the test each quarter, because prices and models keep changing.

Cost per task is the number that matters. A cheap model that needs three tries can cost more than an expensive one that succeeds once. Emergent Wire's editors would test on real workloads before moving any production traffic.

Detailed close-up of gold CPU pins, showcasing technology and design on a circuit board.

There is also a policy angle worth a brief mention. Governments and large companies care where models come from, and Constellation notes that Chinese labs are setting much of the pace in open weights. Some buyers will accept that for low-risk work and avoid it for sensitive data. Expect procurement rules, not just price, to shape how the market splits over the next year, and watch for buyers who ask vendors to disclose which models sit behind a product.

Conclusion

Open-weight models now carry most of the token traffic on one major gateway, yet they earn a small slice of the spending. That gap shows a market splitting into a cheap, high-volume tier and a premium, high-margin tier. Both are growing, and buyers are learning to use each one.

Your next step is to tag your own workloads and test a cheaper model on the routine ones. Emergent Wire will follow the next Vercel report to see whether the gap narrows.

What share of AI tokens do open-weight models handle?
On Vercel's AI Gateway, open-weight models processed 56% of production tokens in August 2026, up from 36% in July and under 10% in December 2025, according to Techstrong. August 22 set a record day at 62%. This is one gateway's data, so it should not be read as overall market share.
Why do open-weight models earn so little despite high volume?
Because they cost far less per token. Techstrong reports that closed-weight tokens cost about 7.8 times as much as open-weight tokens on average, which is the author's own estimate. Open models therefore handle routine, high-volume work, while closed models keep harder, higher-priced tasks, and that explains why volume and revenue diverge.
Does this mean open models are as good as closed ones?
Not entirely, but the gap is small on many tasks. Techstrong cites Mozilla's State of Open-Source AI 2026 report, which puts the capability gap between leading open and closed models at about 3.3%, with near parity on coding and instruction-following. Hard reasoning and frontier tasks still favor closed models.
What should companies do with this data?
Match models to workloads. Constellation Research suggests sending hard reasoning to frontier models and routine work, such as classification, extraction, and repetitive agent steps, to cheaper models that are capable enough. It also suggests trying managed inference providers before self-hosting, since open models can be bought as a service.

Written by

Elena Vasquez

Former ML Researcher, Industry Analysis Lead

9+ years experience

Elena Vasquez holds a PhD in computer science with a research focus on model evaluation methodology, and spent seven years split between an academic lab and a mid-size AI startup's research team before moving into journalism full time. She joined Emergent Wire to bring a practitioner's skepticism to industry coverage — she has built and broken enough benchmarks herself to recognize when a vendor's claimed number is doing more marketing work than measurement work. Her beat covers model capability claims, evaluation methodology, and the gap between published benchmarks and production performance.

Credentials

  • PhD, Computer Science (Machine Learning)
  • Former research scientist, applied AI startup (4 years)
  • Published research on model evaluation methodology