Open-Weight AI Models Grab Half The Traffic, Just 13% Of The Cash
Open-weight AI models handled roughly 50% of tokens on the Requesty AI gateway in the week of Aug. 3, up from under 5% in January, while taking about 13% of spend.
Key Takeaways
- Open-weight models rose from under 5% of Requesty’s tokens in January to about 50% by the week of Aug. 3
- Those models took about 13% of dollar spend, and a closed frontier token cost about 6.4 times an open one after caching
- Anthropic, OpenAI and Google together took 86% of spend on the gateway
- Blended cost per token fell 57% from December to August, even as closed-model list prices rose
What The Gateway Data Shows
Requesty routes API calls from applications to AI model providers for more than 1,000 companies. Chief executive Thibault Jaigu published the figures on Aug. 29, drawn from the gateway’s paid production traffic.
Open-weight models, whose weights anyone can download and run, rose from under 5% of tokens in January to roughly half by the week of Aug. 3. Tokens are the small chunks of text that AI models read and write, and providers bill by the token.
Spend tells a different story. Open-weight models took about 13% of the dollars on the gateway. Anthropic took 51%, OpenAI 21% and Google 15%.
The gap comes down to price. After caching discounts, a closed frontier token cost about 6.4 times as much as an open-weight one.
Cheaper Models And Caching Cut The Bill
Blended cost per token on the gateway fell 57% between December and August, even as closed-model list prices rose. Requesty attributes about half of that drop to the shift toward cheaper models and nearly all the rest to caching.
Cache re-reads, meaning text a model has already seen, made up 78% of tokens processed over 30 days. Fresh input was 19%, and visible output was under 2%.
By step 51 of an agent session, requests cost 23% of list price, compared with 92% at step one. Average context per request grew from 9.6k tokens in January to 28k in August.
Models Turn Over Fast
In the week of Aug. 3, 88% of tokens ran on models first served less than six months earlier. Half ran on models younger than 90 days.
Among 672 organizations sending more than 5 million tokens a week, the median one sent 62% of its traffic to models under 90 days old. After Kimi K3 launched, its own lab lost the majority of its token traffic to rival hosts within 11 days.
Hardware Benchmarks Show The Price Gap
Benchmark runs from SemiAnalysis‘s InferenceX project show how cheaply open models can run. OpenAI’s open-weight gpt-oss-120b sustained 4,782 tokens per second per GPU on AMD’s MI325X, at about $0.064 per million tokens.
Alibaba’s Qwen3.8-Flash-Next reached 4,974 tokens per second per GPU on Nvidia’s H100 at about $0.065 per million tokens. The earlier Qwen3.5 managed 1,584 tokens per second on the same chip at $0.21, roughly three times the cost.
| Model | Hardware | Throughput (tokens/s/GPU) | Cost per 1M tokens | Test window |
|---|---|---|---|---|
| gpt-oss-120b | AMD MI325X | 4,782 | $0.064 | Mar. 9 to May 30, 2026 |
| Qwen3.8-Flash-Next | Nvidia H100 | 4,974 | $0.065 | Through Aug. 27, 2026 |
| Kimi K2.6 | Nvidia GB200 NVL72 | 4,594 | $0.11 | June 20 to 21, 2026 |
| Qwen3.5 | Nvidia H100 | 1,584 | $0.21 | July 5, 2026 |
The runs cover widely available chips such as the H100 and MI325X, not only Nvidia’s newest rack systems. They do not include the full cost of running a model in production, such as engineering, storage and reliability work.
Also Read: Rival Model Beaten by Claude Opus 5.5 at Quarter Cost on Release
What It Means For Data Center Spending
J.P. Morgan estimates hyperscaler capex will reach $697 billion in 2026, and Futurum Group puts it near $690 billion, with Amazon budgeting about $200 billion. Goldman Sachs projects global AI-related investment above $1 trillion this year.
Those budgets rely on inference revenue per token. If routine workloads keep moving to open models at a fraction of closed-model prices, more of that revenue has to come from the harder tasks that stay on closed models.
The Case Against Reading Too Much Into It
Token count is not the same as value delivered. Much of the open-weight volume is likely bulk work, such as summarization and classification, that would never have run on expensive closed models, so closed labs may be giving up little revenue.
The data also comes from a single gateway. No second gateway operator has published matching figures, and Requesty does not disclose its customer list or total token volume.
Read Next: OpenAI Latest GPT-6 Cache Update Adds Cost, Latency Controls
How The Data Was Collected
The Requesty figures cover paid production traffic in 2026, mostly from a 30-day window ending in late August, and are reported as percentages and ratios rather than raw volumes. They exclude bring-your-own-key traffic, internal accounts and, in some cases, the largest accounts, and the EU-hosted gateway may lean toward European customers. The hardware figures come from SemiAnalysis InferenceX runs between March and August 2026, and the capex figures are forecasts from bank and advisory research rather than company disclosures.
