V4.1 Flash 1017 成本实测：DeepSWE 每任务约 $0.20，对手 $2–10；附 Search AP
帖子给出 DeepSWE 每任务约 $0.20、输入 $0.30/输出 $1.20 每百万 token、1M 上下文，并介绍用新搜索 API 做新闻打分 agent 的流程，同时承认 Terminal-Bench 与 HLE 落后 Opus 5。
THIS IS A F*CKING GIFT FOR ANYONE RUNNING A TRADING AGENT DeepSeek dropped V4.1 Flash 1017 today: > ~$0.20 a task on DeepSWE. the big rivals incl. Opus 5 cost $2 to $10+ > 77.0 vs Opus 5's 74.0 on DeepSWE (DeepSeek's own numbers) > $0.30 in / $1.20 out per 1M tokens, half price off-peak > 1M context but the part nobody is talking about: they opened a web search API with it. site:, after:, before: operators. and it's free through Claude Code until Nov 1 that's a free news feed for your bot the split: → DeepSeek Search pulls fresh headlines every 15 min → deepseek-flash scores each one, cheap enough to run all day → Opus only sees the few that score high → you only see what survives both where it still loses: Opus 5 wins Terminal-Bench (51.8 vs 42.3) and HLE (56.3 vs 42.6). so the cheap model grinds and the expensive one decides paste this into Claude Code ↓ "Build news_agent.py: 1. Read DEEPSEEK_API_KEY from .env. 2. Every 15 minutes, call DeepSeek's /v1/tools/search for BTC, ETH and NVDA with after:<today>, 20 results each. 3. Send every new headline to deepseek-flash and get back JSON: ticker, bullish or bearish, confidence 0-1, one line why. 4. Log everything to signals.csv with the price at the time of the headline. 5. Only headlines with confidence > 0.8 go to Opus for a second opinion. 6. Paper trading only. No real orders. Show me the code first. Don't run anything until I say go." i post one of these every day. save this one
这条是引用着另一条写的：x.com/0xtrackmind/status/2111389639739125812