Open models process nearly a third of AI tokens
Vercel sees open-weight models entering production rapidly, but closed models account for most spending.
Open-weight models processed 29 per cent of the tokens that passed through Vercel’s AI Gateway in June 2026. According to the same dataset, less than 4 per cent of spending went to these models, pointing to high volume at relatively low cost.
Vercel bases its monthly Production Index on anonymised routing data from AI Gateway. The service connects production applications with various AI providers. It is therefore not a complete measurement of the global AI market, but of traffic passing through a single platform.
According to Vercel, the share of open-weight models rose from 11 per cent in April to 29 per cent of token volume in June. DeepSeek supplied 22.6 per cent of all tokens on the platform, putting it close to Google. Vercel classifies models as open-weight when their model weights are available for others to use or adapt themselves.
The low spending relative to the high volume is an important part of the development. According to Vercel, open models processed about a quarter of the tokens and accounted for less than 4 per cent of costs. Closed frontier models remained dominant in applications where errors are more costly or risky.
The figures therefore show two markets side by side. For summarising, simple assistance and other large-scale work, inexpensive open models can be attractive. For complex programming tasks, agents and other higher-risk applications, companies continue to pay more often for closed models. According to Vercel, the four largest American frontier labs together received 95 per cent of spending on the platform.
Adoption among companies is also increasing, according to Vercel. Approximately one in eight business customers ran an open-weight model in production in June. This figure is compiled differently from the token share: it counts models regardless of which provider delivers them through the platform. The two percentages therefore cannot be added together directly.
The development expands companies’ choice, but also their responsibility. Open weights give organisations more control over infrastructure and adaptations, while requiring them to arrange more themselves for security, maintenance, evaluation and compliance with regulations. Vercel’s data are also aggregated and may be revised in later reports.
One story, several perspectives
What is established
- Vercel saw open-weight models process 29 per cent of gateway tokens in June 2026.
- Their share of spending was below 4 per cent.
- The data come from a single aggregated platform and are not a complete market measurement.
Left
Arguments Open models can make power and knowledge more widely available and prevent a small number of companies from determining the infrastructure and standards. Public institutions should encourage open development, with strong requirements for safety and transparency.
Values Open access, digital autonomy and democratic control.
Consequences Greater openness can accelerate innovation and reduce dependence, but requires public investment in oversight and safety.
Centre
Arguments Organisations should weigh risk, cost, privacy and reliability for each application. Open and closed models can coexist as long as provider and user can demonstrate compliance with safety and transparency rules.
Values Proportionality, reliability and practicability.
Consequences A risk-based approach prevents policy from favouring one technical model too early, but requires complex assessment.
Right
Arguments Companies should retain the freedom to choose the best models without unnecessary regulations. Open models can bring competition and lower costs, but liability and security must not be shifted onto society.
Values Competition, innovation and individual responsibility.
Consequences Fewer rules in advance can accelerate development, while clear liability remains necessary after the event when harm occurs.
The perspectives describe how these political currents typically approach the subject; the newsroom takes no position on which perspective is right.
Fact-check Approved · Nour Haddad — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The main percentages, measurement period and limitations were checked against Vercel. The text presents the business consequences as context, not as a directly proven market effect.
- confirmed Open-weight models processed 29 per cent of the tokens passing through Vercel’s AI Gateway in June 2026. — Vercel reports 29 per cent of gateway tokens in June. source
- confirmed Less than 4 per cent of spending went to open-weight models. — Vercel cites less than 4 per cent of spend. source
- confirmed DeepSeek accounted for 22.6 per cent of token volume. — This appears in the summary of the Vercel report. source
- confirmed Approximately one in eight business customers ran an open-weight model in production. — Vercel reports this adoption estimate separately from the token share. source
Editor's note
The figures come from aggregated data from Vercel’s AI Gateway covering June 2026 and are not market-wide statistics. The independent second source confirms the broad thrust, but uses the same Vercel data.Sources
More on this in Dutch media
- Tweakers — „kunstmatige intelligentie”
- AD — „kunstmatige intelligentie”
- RTL Nieuws — „kunstmatige intelligentie”