DeepSeek’s very cheap multimodal model (763B parameters), tuned for agent and coding harnesses at a low cost per task. It has a 1M tokens context window, and its weights are openly available. Through DeepSeek’s API it costs $0.15 off-peak · $0.30 peak per million input tokens and $0.60 off-peak · $1.20 peak per million output tokens.
Among the tools we track, it’s available in Cursor and Amp. An October 17 update (1017) shipped at the same price. Peak pricing applies on weekdays from 01:00–04:00 and 06:00–10:00 UTC.
What DeepSeek reports
- DeepSeek reports 77.0 on DeepSWE v1.1
- DeepSeek reports 42.3 on Terminal-Bench 4.0 (Claude Code harness)
Vendor-reported figures, not independently verified by Code×Radar.
Related changes
- Cline 4.1.26 changes default models and adds a model setting for cloud tasks · October 26, 2026
- Augment retires DeepSeek V4 Flash, GLM 5.2 and Kimi K2.7 Code · September 15, 2026
Sources
- api-docs.deepseek.com: pricing
- api-docs.deepseek.com: updates
- huggingface.co: DeepSeek V4.1 Flash 1017
Facts last checked November 15, 2026.