V4.1 Flash 对比 Opus 5.5/Sol：DeepSWE 77% 且便宜近 3 倍
帖子称 V4.1 Flash 在 DeepSWE 编程任务约 77% 成功率、每任务约 $0.20，优于 GPT-6.1 Sol 的约 75%/$0.65；开源权重约 763B，输入 $0.15/输出 $0.30。
Just dropped a model that actually shifts the balance again. DeepSeek-V4.1-Flash-1017. Chinese lab put out something that in some coding tasks matches or edges out Opus 5.5 and Sol, while being dramatically cheaper and much smaller. Here's what actually matters: Performance: It doesn't beat Opus 5.5 across the board, but it's already in the top tier. The standout is DeepSWE (coding tasks): - Flash: ~77% success at ~$0.20 per task - Best GPT-6.1 Sol setup: ~75% at ~$0.65 Almost 3x cheaper and slightly better. For heavy refactoring and iteration loops the gap is real. Price: Basically free next to the frontier models. Input ~$0.15 / Output ~$0.30. On a 100k in + 10k out example: 86-93% cheaper than Sol, 93-97% cheaper than Opus. With their Harness and high cache hit it gets even lower. In practice it can feel 100x+ more usable than burning through Astra or Opus on everything. Size: Open weights. Around 763B total (weights ~510GB). For context, a lot of US frontier models are estimated in the 10T range, Kimi K3 around 2.8T. This is already in the range where you can run it locally on a few DGX Sparks or a maxed Mac Studio. You don't need a massive datacenter to get close to frontier performance anymore. How they did it: They took an existing base and improved it with RL + self-feedback. They raised success rate on hard tasks while also cutting runtime and output length. Stronger and cheaper at the same time is rare. Right now DeepSeek looks like one of the strongest labs on quality / price / accessibility. Balance shifted again.
这条是引用着另一条写的：x.com/deepseek_ai/status/2111375640717033777