DeepSeek-V4.1-Flash reportedly reaches Astra-level DeepSWE accuracy at 1/15th the cost

Fireworks reports that DeepSeek-V4.1-Flash delivered the same accuracy band as GPT-6 Astra on the DeepSWE coding benchmark, at $0.43 per task versus $6.52. In its Astra–DeepSeek comparison, the company says input tokens outnumbered output tokens 174 to 1. It says 99.6% of input tokens were cache hits, which accounted for 60% of the bill.