Speed in AI models is measured primarily in tokens per second (TPS) — how quickly the model generates output. This matters for user experience, but it also has an interesting implication for detection: faster, lower-cost models (designed for throughput) tend to use more formulaic patterns than slower, higher-quality models.
Speed Rankings — 2025
| Model | Speed (TPS) | Quality Tier | Detectability |
|---|---|---|---|
| Groq (Llama 3.1 70B) | ~800 TPS | Good | High — formulaic output |
| Gemini 1.5 Flash | ~250 TPS | Good | High — speed model patterns |
| GPT-4o mini | ~160 TPS | Good | High — vocab markers present |
| Claude 3 Haiku | ~150 TPS | Good | Medium-High |
| GPT-4o | ~80 TPS | Excellent | High — vocab patterns strong |
| Claude 3.5 Sonnet | ~70 TPS | Excellent | Medium — hedging style |
| Gemini 1.5 Pro | ~60 TPS | Excellent | High — encyclopedic style |
| DeepSeek V3 | ~55 TPS | Excellent | High — very formal output |
TPS figures are approximate output token throughput via API. Varies significantly by load, region, and prompt type.
Speed vs. Detection: The Relationship
There's a consistent pattern: faster models trained for throughput produce more predictable output. More predictable output = more detectable. This makes intuitive sense — speed optimization often means reducing diversity in token selection.
Slower, higher-quality models like Claude 3.5 Sonnet and GPT-4o are harder to detect because they have more "temperature" in their outputs — more variance in word choice and sentence structure. They're still detectable (70–80%), but less so than flash/mini models (85–92%).
Which Model Produces the Most Detectable Text?
Based on our testing across all 12 detection signals:
- Groq (Llama 3.1) — fastest and most formulaic. Very high detection rate.
- GPT-4o / GPT-4o mini — strong vocabulary marker pattern. Consistent detection.
- Gemini 1.5 Flash/Pro — encyclopedic style is distinctive. Detects reliably.
- DeepSeek V3 — very formal academic register even on casual prompts.
- Claude 3.5 Sonnet — least detectable of the major models. Epistemic hedging style is distinct but scores lower on standard vocab signals.