Daily snapshot · immutable permalink
AI model usage rankings - September 15, 2026
As published on September 15, 2026, DeepSeek: DeepSeek V4 Flash 0423 led real OpenRouter usage share at 12.3%. Google: Gemma 4 26B A4B (free), Qwen: Qwen3.8 27B (batch), Meta: Muse Spark 1.3 entered the leaderboard for the first time. NVIDIA: Nemotron 3.5 Lightning (free) moved 11 places up since September 14, 2026.
- #1 DeepSeek: DeepSeek V4 Flash 0423 - 12.3% usage share
- #2 Tencent: Hy4 preview - 10.2% usage share
- #3 DeepSeek: DeepSeek V4.1 Flash (batch) - 10.1% usage share
- #4 Z.ai: GLM 5.3 Flash (batch) - 8.9% usage share
- #5 OpenAI: GPT-5.6 Luna (batch) - 7.1% usage share
- #6 Xiaomi: MiMo-V2.5 - 6.5% usage share
- #7 Tencent: Hy3 - 4.4% usage share
- #8 NVIDIA: Nemotron 3 Ultra (free) - 2.5% usage share
- #9 Z.ai: GLM 5.3 (batch) - 2.4% usage share
- #10 Google: Gemini 3.8 Flash (batch) - 2.3% usage share
| # | Model | Usage share | Rank change |
|---|---|---|---|
| 01 | DeepSeek: DeepSeek V4 Flash 0423 | 12.3% | +1 (-0.9pts) |
| 02 | Tencent: Hy4 preview | 10.2% | +1 (+0.7pts) |
| 03 | DeepSeek: DeepSeek V4.1 Flash (batch) | 10.1% | +2 (+3.1pts) |
| 04 | Z.ai: GLM 5.3 Flash (batch) | 8.9% | no change |
| 05 | OpenAI: GPT-5.6 Luna (batch) | 7.1% | -4 (-7pts) |
| 06 | Xiaomi: MiMo-V2.5 | 6.5% | no change |
| 07 | Tencent: Hy3 | 4.4% | no change |
| 08 | NVIDIA: Nemotron 3 Ultra (free) | 2.5% | no change |
| 09 | Z.ai: GLM 5.3 (batch) | 2.4% | +3 (+0.7pts) |
| 10 | Google: Gemini 3.8 Flash (batch) | 2.3% | +1 (+0.7pts) |
| 11 | DeepSeek: DeepSeek V4 Pro 0423 | 2.1% | -2 (+0.1pts) |
| 12 | OpenAI: GPT-5.6 Sol (batch) | 2.1% | -2 (+0.4pts) |
| 13 | Meta: Muse Spark 1.3 Contributor | 1.7% | no change |
| 14 | Anthropic: Claude Sonnet 5 (batch) | 1.4% | no change |
| 15 | Z.ai: GLM 5.2 | 1.3% | no change |
| 16 | Upstage: Solar Pro 4 | 1.3% | no change |
| 17 | MiniMax: MiniMax M3 (batch) | 1.3% | no change |
| 18 | MoonshotAI: Kimi K3 (batch) | 1.2% | no change |
| 19 | Anthropic: Claude Opus 5 (batch) | 1.1% | no change |
| 20 | Poolside: Laguna S 2.1 (free) | 1.0% | no change |
| 21 | inclusionAI: Ling 3.0 Flash Fin | 0.9% | no change |
| 22 | OpenAI: gpt-oss-120b (batch) | 0.7% | +5 (+0.3pts) |
| 23 | NVIDIA: Nemotron 3.5 Lightning (free) | 0.7% | +11 (+0.3pts) |
| 24 | Google: Gemini 3 Flash Preview (batch) | 0.7% | -2 (0pts) |
| 25 | OpenAI: GPT-6 Astra (batch) | 0.6% | -1 (0pts) |
| 26 | Google: Gemini 2.5 Flash Lite | 0.6% | -3 (0pts) |
| 27 | OpenAI: GPT-5.6 Terra (batch) | 0.5% | -1 (0pts) |
| 28 | inclusionAI: Ling 3.0 Flash VL | 0.5% | no change |
| 29 | Google: Gemini 3.7 Flash (batch) | 0.5% | -4 (0pts) |
| 30 | Nex AGI: Nex-N2.5-Pro | 0.4% | +11 (+0.2pts) |
| 31 | Google: Gemini 3.1 Flash Lite (batch) | 0.4% | -2 (0pts) |
| 32 | Google: Gemini 2.5 Flash | 0.4% | -1 (0pts) |
| 33 | Google: Gemma 4 31B (free) | 0.4% | +5 (+0.1pts) |
| 34 | Anthropic: Claude Fable 5.1 (batch) | 0.4% | -2 (0pts) |
| 35 | Qwen: Qwen3.8 Flash | 0.4% | -5 (0pts) |
| 36 | Anthropic: Claude Sonnet 4.6 (batch) | 0.4% | -3 (0pts) |
| 37 | DeepSeek: DeepSeek V3.2 | 0.4% | -2 (0pts) |
| 38 | SpaceXAI: Grok 4.6 (batch) | 0.3% | -2 (0pts) |
| 39 | Anthropic: Claude Opus 4.8 (batch) | 0.3% | -2 (0pts) |
| 40 | Thinking Machines: Inkling | 0.3% | -1 (0pts) |
| 41 | OpenAI: GPT-5.6 Luna Pro (batch) | 0.3% | +4 (0pts) |
| 42 | Qwen: Qwen3.7 Flash | 0.3% | -2 (0pts) |
| 43 | Google: Gemma 4 26B A4B (free) New | 0.2% | - |
| 44 | inclusionAI: Ling 3.0 Flash Sante | 0.2% | no change |
| 45 | Qwen: Qwen3.8 Max | 0.2% | -2 (0pts) |
| 46 | Qwen: Qwen3.8 27B (batch) New | 0.2% | - |
| 47 | Meta: Muse Spark 1.3 New | 0.2% | - |
This page is a permanent, unchanging record of real OpenRouter usage share as it stood on September 15, 2026. Later corrections to the data pipeline appear as new dated snapshots - this page's figures never change.
Source: OpenRouter (openrouter.ai/rankings), as of September 15, 2026.
Benchmark scores: Artificial Analysis (artificialanalysis.ai) via OpenRouter (openrouter.ai/rankings).
Methodology: www.smophy.ai/benchmark/methodology
Token counts originate from each provider's own tokenizer and are not directly comparable across providers.
