V4.1 Flash 跑分与价格汇总：编程超 Opus 5，HLE 仍落后
DeepSWE v1.1 得 77.0 高于 Opus 5 的 74.0，Terminal-Bench 4.0 从 31.2 升到 42.3，API 闲时输入 0.15 美元/百万 token，MIT 权重已上 Hugging Face，但 HLE 42.6 仍低于 Opus 5 的 56.3。
⚡ DeepSeek drops V4.1-Flash-1017 — beats Claude Opus 5 on coding at 1/10th the cost DeepSeek's latest Flash update scores higher on coding benchmarks than Claude Opus 5 — at ~$0.20 per task vs $2–$10 for rivals. Weights are free on Hugging Face under MIT. KEY POINTS • DeepSWE v1.1: 77.0 vs 74.0 for Claude Opus 5; Terminal-Bench 4.0 jumps 31.2 → 42.3 • API price: $0.15/M input tokens off-peak — the per-task cost math checks out • MIT-licensed weights on Hugging Face — run it anywhere, verify every claim • DeepSeek also opened Search & Fetch APIs ($0.75/1,000 calls), free via agent tools till Nov 1 WHY IT MATTERS Agentic coding keeps getting dramatically cheaper every few months — and open weights mean anyone can reproduce the scores instead of trusting a vendor scorecard. WHAT TO WATCH All benchmarks are DeepSeek's own — independent reproduction is still pending. Raw reasoning still trails Opus 5 (HLE 42.6 vs 56.3). Watch Artificial Analysis for the real verdict. #DeepSeek #OpenWeights