Which model should you pick?
Filter by legal jurisdiction, budget, context and vision. The ranking uses published prices, each model official positioning, context size and the speed our probes actually measure.
Ranking
| # | Version | Provider | Data jurisdiction | Price $/M (in / out) | Measured speed (first token) | Context | Our probe | Score |
|---|---|---|---|---|---|---|---|---|
| 1 | Claude Haiku 5.5 | Anthropic | US | 0,1 / 0,5 USD | — | 1 M | Not probed (no key) | 77,6 |
| 2 | GPT 6 Luna | OpenAI | US | 0,1 / 0,5 USD | — | 1,05 M | Not probed (no key) | 76,7 |
| 3 | Jamba Jamba Mini | AI21 Labs | AUTRE | 0,2 / 0,4 USD | — | — | Not probed (no key) | 74,5 |
| 4 | Mistral Small 4 | Mistral AI | EU | 0,15 / 0,6 USD | 339 ms | 262 k | Operational | 68,3 |
| 5 | GLM GLM-5.3-Flash | Z.AI | CN | 0,15 / 0,5 USD | — | — | Not probed (no key) | 67,8 |
| 6 | DeepSeek V4.1 Flash 1017 | DeepSeek | CN | 0,3 / 1,2 USD | 342 ms | 1,05 M | Operational | 55,7 |
| 7 | Gemini 3.1 Flash-Lite | US | 0,25 / 1,5 USD | 163 ms | 1,05 M | Operational | 49,7 | |
| 8 | Gemini 3.8 Flash | US | 0,75 / 3,75 USD | 158 ms | 1,05 M | Operational | 48,6 | |
| 9 | Mistral Large 4 | Mistral AI | EU | 0,7 / 2,1 USD | 337 ms | 1 M | Operational | 47,7 |
| 10 | Gemini 3.5 Flash-Lite | US | 0,3 / 2,5 USD | 160 ms | 1,05 M | Operational | 45,0 | |
| 11 | DeepSeek V4 Pro 0813 | DeepSeek | CN | 1,32 / 3,96 USD | 343 ms | 1,05 M | Operational | 43,1 |
| 12 | Grok 4.3 | xAI | US | 1,25 / 2,5 USD | — | 1 M | Not probed (no key) | 42,2 |
Why this ranking
1. Claude Haiku 5.5 — Anthropic (US) · Volume, automatisations, prototypage
Weighting : Cost 89 · Speed 45 · Capability 55 · Context 99 · Our probe 60
2. GPT 6 Luna — OpenAI (US) · Fort volume, coût minimal
Weighting : Cost 89 · Speed 45 · Capability 50 · Context 100 · Our probe 60
3. Jamba Jamba Mini — AI21 Labs (AUTRE) · Version légère du même archipel
Weighting : Cost 100 · Speed 45 · Capability 50 · Context 20 · Our probe 60
Score 0-100: cost (input price 30%, output 70%), capability (official positioning in the range), context (log scale), speed (measured first token; neutral 45/100 when not probed), state (100 operational, 55 degraded, 0 down, 60 not measured).