We timed 346 large language models from Bengaluru on 12 December 2025 and ranked the 309 that answered by median response time. The fastest answered in 344 ms. The median model took 1.7 seconds, and 96 models took longer than 3 seconds.
Free report · 121 pages · GRX10 Research
A caller notices a pause of more than a second. In a cascaded voice agent the language model is only one step, after speech-to-text and before text-to-speech, so a model that takes 1.7 seconds to answer makes the whole call feel slow.
GRX10 Voice AI uses a single audio-to-audio model, and its first-word response is under 700 ms on real phone calls. We ran this benchmark to see how far general-purpose language models are from that target.
| # | Model | Median latency |
|---|---|---|
| 1 | OpenAI: gpt-oss-safeguard-20b | 344 ms |
| 2 | Qwen: Qwen-Plus | 351 ms |
| 3 | Qwen: Qwen-Turbo | 429 ms |
| 4 | Qwen: Qwen VL Plus | 431 ms |
| 5 | Qwen: Qwen Plus 0728 | 453 ms |
| 6 | Mistral: Codestral 2508 | 472 ms |
| 7 | Qwen: Qwen-Max | 507 ms |
| 8 | Mistral Large 2411 | 508 ms |
| 9 | TheDrummer: Skyfall 36B V2 | 510 ms |
| 10 | Mistral: Ministral 3 8B 2512 | 554 ms |
| 11 | Mistral: Ministral 3 3B 2512 | 556 ms |
| 12 | OpenAI: GPT-4 Turbo Preview | 578 ms |
| 13 | Meta: Llama 4 Maverick | 582 ms |
| 14 | Google: Gemma 3 4B | 583 ms |
| 15 | Cogito V2 Preview Llama 109B | 593 ms |
| 16 | Mistral: Pixtral 12B | 598 ms |
| 17 | NVIDIA: Llama 3.1 Nemotron Ultra 253B v1 | 607 ms |
| 18 | Mistral: Ministral 3B | 618 ms |
| 19 | Mistral: Ministral 8B | 618 ms |
| 20 | Qwen: Qwen VL Max | 631 ms |
Each model received the same short prompt twice through OpenRouter from Bengaluru, India, on 12 December 2025. We recorded the full round-trip time including network latency and ranked models by median. Results change with provider load and region, so treat them as a snapshot.
We'll call you back within one working day.
Expect a call back within one working day.