Video models
Fastest at the median: MiniMax: H3 Max, 14 s a clip; slowest: Kling: Video v3.0 Standard, 2 min 10 s.
| Model | Median time | 9 in 10 within | Median charge | Generations | Accounts |
|---|---|---|---|---|---|
| MiniMax: H3 Max | 14 s | 20 s | $0.50 | 53 | 52 |
| HeyGen: HeyGen Video | 23 s | 1 min 6 s | $0.10 | 11 | 5 |
| SpaceXAI: Grok Imagine Video 1.5 Lite | 35 s | 46 s | $0.16 | 11 | 6 |
| Google: Veo 3.1 Lite | 50 s | 1 min 45 s | $0.30 | 17 | 10 |
| ByteDance: Seedance 2.0 Mini | 1 min 46 s | 2 min 5 s | $0.318 | 12 | 10 |
| Kling: Video v3.0 Standard | 2 min 10 s | 3 min 14 s | $0.63 | 30 | 27 |
Image models
Fastest at the median: Google: Nano Banana 2 (Gemini 3.1 Flash Image), 11 s a image; slowest: OpenAI: GPT Image 2.5 Sunburst, 18 s.
| Model | Median time | 9 in 10 within | Median charge | Generations | Accounts |
|---|---|---|---|---|---|
| Google: Nano Banana 2 (Gemini 3.1 Flash Image) | 11 s | 13 s | $0.067 | 88 | 21 |
| Google: Nano Banana 2.1 | 14 s | 22 s | $0.048 | 481 | 19 |
| OpenAI: GPT Image 2.5 Flare | 14 s | 21 s | $0.006 | 20 | 10 |
| OpenAI: GPT Image 2.5 Sunburst | 18 s | 26 s | $0.006 | 96 | 31 |
Speech and music models
Fastest at the median: OpenAI: GPT Audio Mini, 3.5 s a generation.
| Model | Median time | 9 in 10 within | Median charge | Generations | Accounts |
|---|---|---|---|---|---|
| OpenAI: GPT Audio Mini | 3.5 s | 7.4 s | <$0.01 | 44 | 7 |
Text models
Fastest at the median: OpenAI: GPT Astra Latest, 2.2 s a answer; slowest: Google: Gemini 4 Argon, 1 min 24 s. A text answer's time grows with its length, so output speed is the fairer comparison.
| Model | Median time | 9 in 10 within | Output speed | Median charge | Generations | Accounts |
|---|---|---|---|---|---|---|
| OpenAI: GPT Astra Latest | 2.2 s | 9.2 s | 29 tokens/s | $0.007 | 107 | 11 |
| Google: Gemini 3.5 Flash | 3.7 s | 18 s | 101 tokens/s | <$0.01 | 17 | 8 |
| Anthropic: Claude Sonnet 5 | 13 s | 25 s | 70 tokens/s | $0.011 | 16 | 6 |
| Google: Gemini 4 Argon | 1 min 24 s | 1 min 54 s | 25 tokens/s | $0.096 | 13 | 3 |
How this is measured
Every generation on GenMagic, in the studio and through the API, records its model, when it was asked for, when the finished result came back and what it was charged. Time runs from the request to the finished result as GenMagic saw it; for video that includes the wait in the model's queue until the clip was collected, which is the wait you would have. Only completed generations count, over the last 90 days (since Aug 6, 2026). The charge is the median of what those generations were billed at the time, so it moves with the settings people chose (a longer clip, a larger image); each model's current price is on its page and on the video, image and LLM price lists.
A model appears only after at least 10 completed generations from at least 3 different accounts, so no one person's use can be read from it, and it is published in aggregate only: never a prompt, an output or an account. More models join as people use them. How every other figure on GenMagic is sourced is on our sources page.