DeepSeek’s latest Flash update is processing more than a quarter of all tokens on OpenRouter, Tencent’s new Hy4 has leapt straight to second place, and only three US models make the list of the nine most-used models on the platform.
The numbers below come from OpenRouter’s usage data for the week of November 9, 2026. At the time of the snapshot the platform had processed 199 trillion tokens, with the week on pace to finish at around 213 trillion. Percentages are each model’s share of that 199T total. OpenRouter ranks models by tokens processed, which isn’t the same as spend or number of users, so a cheap model can top this list without generating anywhere near the most revenue.
| Rank | Model | Lab | Tokens | Share |
|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash 1017 | DeepSeek | 55.4T | 27.8% |
| 2 | Hy4 | Tencent | 17.5T | 8.8% |
| 3 | DeepSeek V4.1 Flash | DeepSeek | 14.6T | 7.3% |
| 4 | GLM 5.3 Flash | Z.ai | 11.3T | 5.7% |
| 5 | GPT-6 Luna | OpenAI | 7.36T | 3.7% |
| 6 | Claude Opus 5.5 | Anthropic | 6.64T | 3.3% |
| 7 | Step 5 Preview | StepFun | 6.27T | 3.2% |
| 8 | MiMo-V2.6-Flash | Xiaomi | 3.97T | 2.0% |
| 9 | GPT-6.1 Sol | OpenAI | 3.56T | 1.8% |
| Everything else | Various | 72.1T | 36.2% |
1. DeepSeek V4.1 Flash 1017 — 55.4T tokens (27.8%)
This is the runaway leader. On its own, the model handles more than a quarter of everything flowing through OpenRouter, and more than three times as many tokens as the runner-up.
DeepSeek’s V4.1-Flash-1017 update was pitched as a “Pareto improvement” over the model it replaced, with better scores, faster responses and shorter outputs, and it edges past Claude Opus 5 on some agentic benchmarks. On Artificial Analysis’ index it scored 45, tying GLM 5.3 as the joint second-best open-weights model at the time, with a large reduction in hallucinations over its predecessor.
Price is the real draw. Pricing is unchanged from the original V4.1 Flash at $0.30 per million input tokens and $1.20 per million output tokens during peak hours, with off-peak rates at half that. The model has 552 billion total parameters (16B active), a 1M-token context window and an MIT license.
2. Hy4 — 17.5T tokens (8.8%)
Tencent’s Hy4 is the biggest story on the chart. OpenRouter lists the model as released on November 4, which means it climbed to second place within days of launch.
It’s a repeat of what Tencent managed earlier this year: Hy3 Preview took second place on the June leaderboard too, behind a DeepSeek model. Tencent’s Hy4 preview, open-sourced in August, is a mixture-of-experts model with 770 billion total and 49 billion active parameters and a context window beyond 1M tokens. On Artificial Analysis’ latest index Hy4 scores 35, level with Grok 4.7 and just behind Meta’s Muse Ember and Claude Haiku 5.5.
Distribution is probably helping. Hy4 is available through Tencent’s own agent tools such as CodeBuddy and WorkBuddy, which are also among the harnesses DeepSeek said it optimized its 1017 update for.
3. DeepSeek V4.1 Flash — 14.6T tokens (7.3%)
The model that the 1017 update replaced is still in third place. V4.1 Flash launched in September as the smallest model in DeepSeek’s new architecture family, with a KV cache of just 890 bytes per token and a pricing structure that undercut Claude Opus 5 by a factor of 20 or more on output.
Add its two Flash models together and DeepSeek accounts for 70T tokens, or 35% of everything on the platform.
4. GLM 5.3 Flash — 11.3T tokens (5.7%)
Z.ai’s cheap multimodal model is still a top-five model more than two months after its late-August launch. Before launch it ran anonymously on OpenRouter as the free “stealth” model Ox Alpha, and Z.ai founder Jie Tang later said it delivered nearly 20% weekly token share on OpenRouter at its peak.
GLM-5.3-Flash has 320 billion total parameters with 18 billion active, ships under an MIT license, and costs $0.15 per million input tokens and $0.50 per million output tokens.
5. GPT-6 Luna — 7.36T tokens (3.7%)
OpenAI’s budget model is the most-used US model on the chart. After OpenAI halved its price at launch, GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens.
Artificial Analysis found only modest intelligence gains over GPT-5.6 Luna, but much lower running costs. Luna scores 37 on the index at maximum effort and costs about $0.07 per task, numbers that appeal to developers running high-volume workloads. It also now faces competition at the same price from Claude Haiku 5.5.
6. Claude Opus 5.5 — 6.64T tokens (3.3%)
Anthropic’s flagship is the only Claude model in this top nine, and it’s the most expensive model on the list by a wide margin. At $4 per million input tokens and $20 per million output tokens, its output price is roughly 17 times DeepSeek V4.1 Flash 1017’s peak rate.
Developers are paying for capability. Opus 5.5 took the top spot on the Artificial Analysis Intelligence Index at launch, five points clear of GPT-6 Astra. Anthropic has since released Fable 5.2, which now sits above it on the index at $8/$40, but Opus 5.5 remains the one Claude model with enough volume to make the chart.
7. Step 5 Preview — 6.27T tokens (3.2%)
StepFun’s flagship opened its API on September 20 and is a 600-billion-parameter mixture-of-experts model with 27 billion active parameters and a 1M-token context window. StepFun said weights would follow on October 15.
It scored 44 on Artificial Analysis’ index at launch. At $1.00 per million input tokens and $2.70 per million output tokens it is considerably pricier than the DeepSeek and GLM models above it, and it’s also verbose: it generated 160 million output tokens on the index run against a median of 92 million. That it still lands in the top nine says something about how much developers value capable agentic models.
8. MiMo-V2.6-Flash — 3.97T tokens (2.0%)
Xiaomi’s efficiency-focused omnimodal model is the smaller sibling of MiMo-V2.6-Pro, which debuted as the top open-weights model on the Artificial Analysis index at a cost of just $0.13 per task. Both launched in September. Flash has 309 billion total parameters (15B active), a 1M-token context window, and accepts text, image, video and audio.
9. GPT-6.1 Sol — 3.56T tokens (1.8%)
OpenAI’s mid-tier model rounds out the list. GPT-6.1 Sol is priced at $2 per million input tokens and $10 per million output tokens, a fifth of GPT-6 Astra’s price, and Artificial Analysis placed it just one point behind Astra at launch. Its output price is still more than eight times DeepSeek’s peak rate, which helps explain why it sits at the bottom of this chart despite its strong benchmark showing.
The bigger picture
Chinese models dominate. Six of the nine named models come from Chinese labs, and together they account for 109T tokens, roughly 55% of all traffic on the platform and 86% of the volume among the nine named models. The three US models combined, two from OpenAI and one from Anthropic, make up 17.6T, or about 9% of the total. That fits the longer trend in which US models’ share of OpenRouter tokens fell from around 70% to 30% in a year.
Usage tracks price, not capability. The two most-used models are both DeepSeek Flash variants. On Artificial Analysis’ newer v5 index, Opus 5.5 scores 58 while DeepSeek V4.1 Flash 1017 and GLM-5.3 share the lead among released open-weights models at 32. That gap hasn’t dented DeepSeek’s volume, because for high-volume agentic workloads many developers appear to be choosing the model that’s good enough at a small fraction of the cost.
The long tail still matters. Everything outside the top nine adds up to 72.1T tokens, 36% of the total and more than any single model on the chart.
One caution. This is a snapshot of a single week, and launches reshuffle this list quickly. Hy4 only arrived on November 4, and the weekly pace figure suggests the final numbers will be higher. A model’s rank here also reflects free tiers, promotions and tool integrations as much as pure developer preference.