Closed-weight AI usage on OpenRouter rose 13-fold while its share of tokens fell from 74% to 28%. Open-weight usage increased 95-fold. The total expanded 34-fold. A percentage chart captures the change in relative scale, but hides much of the growth behind it.[1]
These figures measure traffic routed through OpenRouter. They exclude requests sent directly to OpenAI and Anthropic APIs, and usage in their own applications such as ChatGPT and Claude. They also omit open-weight models run directly through other hosts or on private infrastructure. The charts therefore cannot establish total provider usage growth or market-wide open/closed shares. Growth on OpenRouter could reflect new demand, traffic moving between routes, or both.[1]
The more recent pattern is just as revealing. Closed-weight models recovered some share as their growth accelerated. Successive releases from OpenAI, Anthropic and DeepSeek attracted substantial usage, challenging a simple account of models becoming interchangeable. The strategic question is whether providers can keep delivering improvements that customers value, at a cost that allows them to earn acceptable returns.
- Falling share can coexist with rapid growth. Closed-weight weekly tokens increased from 3,60 trillion to 46,70 trillion.
- Recent competition has been more balanced. Closed-weight share of volume recovered from approximately 26% in July to 31% in September.
- New releases can attract usage quickly. DeepSeek V4.1 Flash, GPT-6 Luna and Claude Opus 5.5 became substantial sources of OpenRouter traffic within weeks.
- Capability and economics need to be assessed together. Higher performance, lower prices and fewer tokens per completed task can all change demand.
- Usage growth does not settle the moat question. Providers still need to convert that activity into returns after serving customers and financing the next generation.
OpenRouter measures one route to the market
The missing direct traffic is material to interpretation. A change in OpenRouter share could coexist with a different pattern in direct subscriptions and API usage. The OpenRouter series cannot tell us how quickly direct OpenAI or Anthropic usage grew, whether it grew faster than routed usage, or whether either provider gained or lost share across all channels. Ramp provides a complementary connected-provider sample, discussed below, rather than a complete measure of all direct activity. The 13-fold figure belongs to closed-weight tokens on OpenRouter, rather than to the whole closed-weight market.
This also narrows what the chart can say about competitive moats. A shared routing platform makes it possible to compare models without capturing all the advantages a provider might build through its own applications, workflow integration, enterprise relationships or distribution. Those potential advantages require separate evidence. OpenRouter remains a useful lens on competition between models, but an incomplete lens on competition between the businesses behind them.
A growing denominator changes the interpretation
Between the complete weeks beginning 20 October 2025 and 5 October 2026, open-weight volume increased from 1,27 trillion to 120,95 trillion tokens per week. Closed-weight volume increased from 3,60 trillion to 46,70 trillion. Together, they grew from 4,87 trillion to 167,65 trillion tokens per week. Open-weight models accounted for approximately 74% of the increase.[1]
The open-weight share gain is real. It does not follow that closed-weight demand collapsed. Both categories expanded substantially, and one expanded faster. Nor does their simultaneous growth prove there was no substitution: these aggregate figures cannot reveal how much closed-weight usage might otherwise have occurred.
The three panels answer different questions. Absolute volumes establish the scale of activity. Growth rates show how quickly that scale changes. Percentage shares show how the activity is divided. Reading the share chart alone leaves out two of those questions.
Closed-weight share has recently stabilised as growth accelerated
The latest months are inconsistent with a smooth, uninterrupted decline in closed-weight share. In July, average weekly closed-weight volume was 14,94 trillion tokens. It reached 25,86 trillion in August and 39,36 trillion in September. Its share of the corresponding combined volumes rose from 26,0% to 29,5% and then 30,6%. These shares are calculated from monthly average weekly volumes, rather than by averaging weekly percentages.[1]
Closed-weight monthly growth was approximately 73% in August and 52% in September, compared with 45% and 44% for open-weight models. By September, the trailing three-month average of monthly growth rates had converged to approximately 45% for closed-weight models and 47% for open-weight models.[1]
There is still considerable weekly volatility. Since July, closed-weight share has moved within a broad range of approximately 25% to 34%, ending at 28% in the final complete week. Stabilisation describes a recent competitive pattern, rather than an established long-term equilibrium.
New models are changing the competitive balance
The release sequence helps explain why the recent period deserves closer attention. Provider records distinguish several events that a broad reference to “the launch of GPT-6” or “DeepSeek Flash” would combine:
| Date in 2026 | Release or pricing event |
|---|---|
| 9 July | OpenAI released the GPT-5.6 family, including Sol, Terra and Luna.[3] |
| 30 July | OpenAI reduced GPT-5.6 Luna pricing by 80% and Terra pricing by 20%, and introduced Fast mode for Sol.[3] |
| 31 July | DeepSeek released the V4 Flash update, reporting stronger agent capabilities.[5] |
| 3 September | OpenAI released GPT-6 Astra.[3] |
| 10 September | DeepSeek released V4.1 Flash and reduced API prices.[5] |
| 22 September | Anthropic released Claude Opus 5.5. OpenAI released GPT-6 Sol and Luna.[3], [4] |
| 29 September | OpenAI released GPT-6.1 Sol.[3] |
OpenRouter’s underlying model chart provides corroborating usage evidence. GPT-5.6 Luna processed approximately 1,95 trillion tokens in the week beginning 27 July, which straddled the price reduction. By the week beginning 7 September, it processed 18,18 trillion. That is a substantial increase following a major improvement in affordability, and a reason not to attribute all of the GPT-5.6 family’s effect to the flagship Sol model.[2], [3]
DeepSeek V4.1 Flash processed 15,77 trillion tokens in its first full week, beginning 14 September, and 40,97 trillion in the week beginning 5 October. In that final week, GPT-6 Luna processed 7,08 trillion and Claude Opus 5.5 processed 5,02 trillion. These were already substantial volumes shortly after their September releases.[2]
These observations support the conclusion that the market responds rapidly to new model offerings. They do not isolate the effect of capability from pricing, availability, distribution or changes in customer workloads. GPT-6.1 Sol arrived especially late in the observation window: only one complete week in this graphic falls entirely after its release. Its individual contribution, and that of GPT-6 Astra and GPT-5.6 Sol, cannot be established from the aggregate open/closed chart alone. Ramp’s model-level data provides additional corroboration below. OpenRouter’s public model chart names the nine leading models in each week and groups the remainder as “Others”.[2]
Ramp provides a complementary view of provider activity
Ramp helps address the coverage gap. Its Token Spend Management product imports model usage and billing data from connected providers. The token-volume and token-spending charts therefore provide evidence of provider activity in a business sample beyond the OpenRouter routing platform. They are distinct from Ramp’s broader adoption series.[9]
For the complete week of 5–11 October 2026, Ramp reports the following shares in that connected-provider sample:
| Model maker | Share of token volume | Share of token spending |
|---|---|---|
| Anthropic | 58,9% | 60,1% |
| OpenAI | 38,2% | 36,4% |
Anthropic and OpenAI together accounted for approximately 97% of both measures. The contrast with OpenRouter’s open-weight majority demonstrates why platform token shares should not be treated as shares of the entire market. Ramp’s sample is also selective: these figures are not estimates of the providers’ worldwide shares.[9]
The model detail is particularly revealing. By 5–11 October, Claude Opus 5.5 accounted for 35,1% of Ramp’s observed token volume. GPT-6 Luna accounted for 9,1%, GPT-6.1 Sol for 6,5% and Claude Sonnet 5.5 for 6,4%. All four had been released in the preceding few weeks. At the same time, GPT-5.6 Sol’s share was 2,7% and Opus 5’s was 4,2%, illustrating the movement of activity towards newer offerings.[9]
There is evidence of expansion as well as changing mix. Between the weeks beginning 7 September and 5 October, Anthropic token volume in Ramp’s underlying records increased by approximately 56%, and OpenAI volume by 50%. Within that period, GPT-6 Astra volume increased by approximately 61% from its first full week, beginning 7 September, to the following week. Opus 5.5 volume increased by approximately 46% between its first two full weeks, beginning 28 September and 5 October. These growth calculations use like-duration weeks and avoid treating a partial launch week as a full baseline.[9]
Ramp displays volumes on a common relative index, where 100 represents the peak of any single series in the selected chart. Its connected-provider cohort is not held constant here, and changes in connected customers or workloads may contribute to measured growth. The data corroborates rapid uptake of new models alongside provider-level expansion, without isolating capability as the sole cause or establishing worldwide growth.
Commoditisation and capability-led growth can coexist
There are good reasons to take the commoditisation argument seriously. More models compete for workloads, and customers can choose increasingly capable alternatives at lower prices. But lower prices do not establish that all models are equivalent or that improvements in capability have ceased to matter.
Research published in the Journal of Economic Perspectives, using OpenRouter data, documents both falling prices and substantial differentiation between models. It finds frequent turnover among leading suppliers and differences in demand across applications, with no single model dominating every use case. That combination is consistent with strong competition alongside continuing demand for differentiated capability.[7]
A model that can complete a difficult coding task reliably may become useful for work that a cheaper but less capable model cannot finish. Greater reliability can also reduce retries and human checking. Improvements can therefore expand the work customers attempt, as well as redirect existing work between models. This is a plausible interpretation of the rapid uptake around new releases, rather than a causal finding established by these charts.
The relevant comparison is often the cost of an acceptable completed task. A lower price per token helps, but so can fewer tokens, fewer failed attempts or less human intervention. Anthropic’s Opus 5.5 announcement illustrates this combination: it reduced input and output token prices by 20% relative to Opus 5 and reported an approximately 40% reduction in typical task costs through both pricing and efficiency. These are provider-reported results, rather than an independent measurement of every customer workload.[4]
The moat depends on the ability to keep improving
The data provides little support for treating ownership of closed weights, by itself, as proof of a durable competitive moat. Open-weight competitors can attract large volumes, and the DeepSeek results show that rapid uptake of new capability is available to both categories. The distinction between open and closed weights does not determine which model will deliver the best result for a particular task.[2], [8]
A stronger potential advantage is the ability to repeatedly deliver improvements customers value before competitors match them. That could sustain demand even as older capabilities become widely available. However, an advantage that must be renewed through each generation creates a different economic challenge from one that persists with little further investment.
Can the additional gross profit earned from better models cover the continuing cost of reaching and operating at the frontier? Usage growth is one input to that question. The answer also depends on realised revenue per task, inference costs, research and training expenditure, infrastructure commitments and how long an improvement remains commercially valuable.
Faster growth can coexist with lower prices and higher investment requirements. A capable new model can win traffic without generating sufficient returns on its development. Conversely, a provider can reduce cost per completed task while earning more if efficiency, adoption and customer value improve sufficiently. The charts cannot determine which outcome applies.
This is the strategic distinction behind the moat debate. Successive model launches show that competitive positions can change quickly. They leave providers with the task of turning repeated technical progress into durable economic returns.
Business adoption provides a separate signal of expansion
Separately from its Token Spend Management measures, Ramp’s AI Index tracks observed payments across a broader business population, including provider subscriptions and API spending. The share of its businesses making an observed AI payment increased from 35,56% in January 2025 to 55,86% in September 2026. In September, Anthropic’s adoption rate was 43,77% and OpenAI’s was 40,63%. Businesses can pay both providers, so these vendor percentages overlap and cannot be added as market shares.[6]
For business leaders, the implication is to assess models against the outcomes and total costs of real workflows, and to reassess those choices as capabilities change. For providers, strong usage is evidence of demand to serve. Demonstrating a moat also requires evidence that the capability generating that demand can be renewed profitably.
Do you like this Fast Fact? Subscribe to our email newsletters to receive updates on topics you select, at a frequency you control.
Sources and methodology
All usage calculations refer to the data retrieved on 17 October 2026. The headline comparison uses 51 complete weeks beginning 20 October 2025 to 5 October 2026. The incomplete week beginning 12 October is excluded. Monthly growth compares mean weekly token volumes grouped by the week-start month, using November 2025 to September 2026. These are not exact calendar-month totals. The moving average is the trailing arithmetic mean of three monthly growth rates. CAGR annualises the 350-day interval and is not a forecast.
OpenRouter covers traffic through its platform, rather than the entire AI market or ChatGPT consumer usage. Its open/closed chart excludes stealth models. The model-level top-models chart has a different presentation, including an “Others” group; its denominator has not been substituted for the open/closed denominator. Token counts are affected by workload, model behaviour and tokenization and do not directly measure useful work, revenue or profit. Ramp’s adoption measure covers observed payments in its broader business sample. Its token volume, spending and price measures cover the smaller Token Spend Management cohort using connected providers. No Ramp token volumes or spending have been added to OpenRouter totals, and the two datasets are not assumed to have identical coverage or to be deduplicated.
- OpenRouter, Open / Closed rankings and underlying weekly chart data. Global Advisors calculations. Data retrieved 17 October 2026. Licensed under CC BY 4.0.
- OpenRouter, Top Models and underlying model chart data, retrieved 17 October 2026. Model definitions corroborated against the OpenRouter pages for GPT-5.6 Sol, GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5 and DeepSeek V4.1 Flash.
- OpenAI API changelog, entries dated 9 and 30 July, and 3, 22 and 29 September 2026. See also the GPT-6.1 Sol model documentation. Release dates refer to the API, which is relevant to routed usage.
- Anthropic, Introducing Claude Opus 5.5, 22 September 2026. Performance and task-cost claims are attributed to Anthropic.
- DeepSeek API change log, entries dated 24 April, 31 July and 10 September 2026. The separate update dated 17 October is outside this graphic’s observation window.
- Ramp AI Index, monthly adoption data and methodology. Adoption means an observed positive AI payment in the month; free usage and personal accounts may be missed.
- Demirer, Mert, Andrey Fradkin and Nadav Tadelis. The Emerging Market for Intelligence: How Firms Buy and Sell AI. Journal of Economic Perspectives, 40(3), Summer 2026, pp. 23–46.
- DeepSeek’s official V4.1 Flash model card and weight repository.
- Ramp AI Index, Token volume by model maker; Token volume by model; and the Token spend view on the same page. Published leaderboard shares and underlying weekly model records, retrieved 17 October 2026. Token Spend Management imports usage and billing from connected providers and is not representative of all Ramp AI spending. Indices share a common scale within a chart; growth ratios are calculated from the underlying weekly records, without a fixed-cohort adjustment.
