DeepSeek V4.1-Flash matches Claude Opus 5 on coding at a fraction of the cost

DeepSeek released V4.1-Flash on September 10, 2026, with MIT-licensed weights on Hugging Face. It scores 74.2 on DeepSWE v1.1 versus Claude Opus 5's 74.0, but trails on hard reasoning (36.8 vs 56.3 on Humanity's Last Exam). From September 14, all V4-Pro API requests are rerouted to V4.1-Flash at Flash pricing.

• 552B total parameters in MoE layout, 8B active on prefill and 16B on decode
• Off-peak pricing: $0.15 per 1M input tokens, $0.003 cached, $0.60 output
• KV cache is 890 bytes per token with at least 72 hours of persistence
• 1M-token context, pre-trained on 45 trillion tokens