{
  "video": {
    "id": "U7ZYBCcdJPg",
    "title": "Tokens Are the Cheap Part — Steve Jamison, Jellyfish",
    "duration": 1052,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "U7ZYBCcdJPg",
    "title": "Tokens Are the Cheap Part — Steve Jamison, Jellyfish",
    "one_liner": "Measure AI by cost per merged PR, counting developer salary plus AI spend: on Jellyfish's data, AI spend rose 60x in a year, yet AI is only about 2.5% of what a merged PR costs, and cost per merged PR fell 26%.",
    "summary": "Steve Jameson, AI research and engineering leader at Jellyfish, uses Jellyfish's engineering workflow data joined with AI telemetry to settle the tension between a 60x rise in AI spend and a 43% rise in merged PRs per engineer. He argues that looking at either number alone hides the trade-off. Instead, add developer salary and AI spend into one 'intelligence cost' per PR. Measured that way, AI is about 2.5% of the cost of a merged PR, cost per merged PR fell 26% in a year, and the companies that spent the most on tokens gained the most throughput and saw the biggest drop in cost per PR. Returns diminish, he says, but most companies are not close to the point where tokens stop being the cheap part.",
    "key_points": [
      "Over one year, AI costs rose 60x and merged PRs per engineer rose 43%. AI spend climbed in early 2026, around when throughput began its steady rise.",
      "Counting salary plus AI as one 'intelligence cost' per PR, AI is about 2.5% of the cost of a merged PR and salary is 97.5% (data from August 2026).",
      "The dollar cost of a merged PR fell 26% over the year despite the 60x rise in AI costs, and 83% of the companies studied saw cost per PR drop. Throughput gains spread salary across more PRs faster than AI costs grew.",
      "Grouped into fifths by token use: in the lightest fifth, AI is about 0.3% of PR cost (~60M tokens per dev per month). In the heaviest fifth it is 8.5% (~2.5B tokens per dev per month), about two orders of magnitude more.",
      "The heaviest-spending fifth raised PR throughput 67% and cut cost per PR 32%. The other four fifths gained about 40% throughput and cut cost per PR about 25%.",
      "Plotting cost per PR against tokens used shows diminishing returns, not zero returns. The top 10% of company-months used 8x the median's tokens and had a 25% lower cost per PR.",
      "In a developer-level study, comparing each developer's heavy-token weeks with their own lighter weeks: about 2x token spend, 27% more merged PRs, and about 18% lower cost per PR.",
      "Quality and PR-size checks: revert rates fell slightly for every group, though within noise, and there were no signs of serious quality problems. The extra throughput does not come from smaller PRs; lines of code per PR went up."
    ],
    "takeaways": [
      "Track total intelligence cost per merged PR (salary plus AI spend) every month, alongside the AI bill and PR throughput.",
      "Don't make AI spend per PR your driving metric. Still track spend and watch for runaway jobs, but don't cap tokens in a way that could block further efficiency gains.",
      "Before investing in cost controls such as model routers, post-training your own model or token limits, ask whether more AI enablement that raises throughput would pay off more.",
      "Watch reverts and other quality metrics alongside cost and throughput as a counterweight to AI slop.",
      "Heading into budget season, bring cost per PR to the table, not just the token bill or the throughput bump."
    ],
    "topics": [
      "ai cost",
      "developer productivity",
      "engineering metrics",
      "token spend",
      "cost per pr",
      "ai budgeting",
      "code quality"
    ],
    "tools": [
      "Jellyfish",
      "LinkedIn",
      "Substack"
    ],
    "quotes": [
      {
        "text": "AI is about 2 and 1/2% of the cost to produce another unit of merged code.",
        "at": "03:38",
        "url": "https://www.youtube.com/watch?v=U7ZYBCcdJPg&t=218s"
      },
      {
        "text": "diminishing returns is not the same as no returns.",
        "at": "08:10",
        "url": "https://www.youtube.com/watch?v=U7ZYBCcdJPg&t=490s"
      },
      {
        "text": "We don't need AI slop mucking up the gears.",
        "at": "16:12",
        "url": "https://www.youtube.com/watch?v=U7ZYBCcdJPg&t=972s"
      },
      {
        "text": "The bill tells you what you spent. Cost per PR tells you whether it worked.",
        "at": "16:25",
        "url": "https://www.youtube.com/watch?v=U7ZYBCcdJPg&t=985s"
      }
    ],
    "words": 2650
  },
  "summary_url": "/#U7ZYBCcdJPg",
  "transcript": {
    "html": "/transcripts/U7ZYBCcdJPg.html",
    "txt": "/transcripts/U7ZYBCcdJPg.txt",
    "vtt": "/transcripts/U7ZYBCcdJPg.vtt"
  }
}