<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>Cost per Task on AI Pulse</title>
    <link>https://cuigh.com/tags/cost-per-task/</link>
    <description>Recent content in Cost per Task on AI Pulse</description>
    <generator>Hugo</generator>
    <language>en-US</language>
    <lastBuildDate>Sat, 25 Jul 2026 12:44:04 +0800</lastBuildDate>
    <atom:link href="https://cuigh.com/tags/cost-per-task/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Claude Opus 5: Token price unchanged, but cost per task just hit a new floor</title>
      <link>https://cuigh.com/posts/claude-opus-5-cost-per-task-2026/</link>
      <pubDate>Sat, 25 Jul 2026 12:44:04 +0800</pubDate>
      <guid>https://cuigh.com/posts/claude-opus-5-cost-per-task-2026/</guid>
      <description>Anthropic released Claude Opus 5 on July 24. Token prices did not move. What moved is cost per task: half on CursorBench 3.2 at max effort, one-third on OSWorld 2.0, three times the next-best score on ARC-AGI 3. The yardstick for picking models is shifting from per-token to per-task, and unit intelligence cost is the new moat.</description>
    </item>
  </channel>
</rss>
