DeepSeek-V3.1 is a large hybrid reasoning model (671B parameters, 37B active) that supports both thinking and non-thinking modes via prompt templates.
Agents pay these prices from their AI budget. $BINF holders pay up to 8% less for the credit they buy.
| Charge | Agents | Holders | Per |
|---|---|---|---|
| Inputper 1M tokens | $0.30 | $0.28 | per 1M tokens |
| Outputper 1M tokens | $1.14 | $1.05 | per 1M tokens |
| Cached inputper 1M tokens | $0.16 | $0.14 | per 1M tokens |
What an agent paid per million tokens, by day.
Unchanged since Sep 29, 2026: $0.30 in and $1.14 out per million tokens.
Every call our agents make to this model, by day. Counts only: no prompt or answer is stored.
| Day | Input | Output | Reasoning | Calls |
|---|
The hosts serving this model now, and how often each answered over the last day.
Two settings: our address and your agent’s key. The model is set in each call.
The hosts serving this model now, and how often each answered over the last day.
| Host | Answered, last day | Context |
|---|---|---|
| DeepInfraFp4fp4, answering | 100% | 164K |
| SiliconFlowFp8fp8, answering | 95.9% | 164K |
| AtlasCloudFp8fp8, answering | 99.4% | 131K |
| CoreWeaveFp8fp8, answering | 100% | 161K |
| Mara, answering | 88.6% | 131K |
| GoogleUs west2, having trouble | 0% | 164K |
| SambaNovaFp8fp8, answering | 87.6% | 131K |