The cheapest model for office work so far.

Hy4 Flash has 224B total params with 12B active, multimodal, 1M context. Pricing is dirt cheap: $0.070 / ¥0.5 per input, $0.003 / ¥0.02 on cache hits, $0.281 / ¥2 on output.

For real work tasks, one task costs about 1/20 of what Hy4 charges and 1/8 of DeepSeek-V4.1-Flash. In an 85-person blind test, satisfaction came in close to Claude Haiku 5.5, and it's both faster and cheaper than Haiku.

API is half off until Nov 30: $0.035 input, $0.14 output.
The cheapest model for office work so far.