DeepSeek V4.1-Flash reportedly reaches Astra-level DeepSWE accuracy at 1/15th the cost
Fireworks says input tokens outnumbered output tokens 174 to 1 in its DeepSWE comparison, with cache hits accounting for 60% of the bill.
TLDR
Fireworks reports that DeepSeek V4.1-Flash on its platform reached the same accuracy band as GPT-6 Astra on the DeepSWE coding benchmark, costing $0.43 per task versus $6.52. The company says input tokens outnumbered output tokens 174 to 1, and 99.6% of input tokens were cache hits. Those cache hits still accounted for 60% of the bill.
Combined views
16
1 Source, first seen 15d ago