{
  "video": {
    "id": "prGP0huECPo",
    "title": "Without Caching, Your AI Agent Costs 10x More - George Cameron, Artificial Analysis",
    "duration": 1183,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "prGP0huECPo",
    "title": "Without Caching, Your AI Agent Costs 10x More - George Cameron, Artificial Analysis",
    "one_liner": "For long-running agents, cost is driven by input tokens, cache hit percentage and turns per task rather than output tokens: Opus 5.5 at $21 per task on AA Briefcase would cost over $200 with no cache hits.",
    "summary": "George Cameron of Artificial Analysis argues that the cost of intelligence in mid-late 2026 differs from even 6-9 months earlier, because individual agent tasks now cost many dollars and models differ in cost per task by multiple orders of magnitude. He breaks down cost per task into answer tokens, reasoning tokens, input tokens, cache hits and number of steps, showing that input tokens, caching and turns matter most. He says the hard part is not knowing which model is cheaper but making the trade-off decision, using the intelligence-versus-cost Pareto and sorting tasks into 'unlimited intelligence' and 'limited intelligence'. For production rollouts, he recommends creating your own benchmark from a shortlist of models.",
    "key_points": [
      "On the AA Briefcase benchmark (agents working among thousands of files on tasks such as reports, PowerPoints and marketing videos), Opus averages $21 per task: about 13 cents per language model call across 165 steps.",
      "Cached input tokens get roughly a 10x discount on the input token price; with a 0% cache hit rate, for example a self-hosted vLLM setup without prompt caching enabled, the $21 task costs over $200.",
      "A year ago the cost discussion centred on output tokens and reasoning verbosity; Cameron says the key agent cost metrics today are input tokens, cache hit percentage and number of turns per task.",
      "Reasoning effort affects more than reasoning tokens: it changes the number of turns, and reasoning tokens feed through message history into more input tokens. Opus 5.5 can cost $1 per task at low effort and $21 at max.",
      "On one chart, Opus and Fable cost above $5 per task while some models cost 5 cents or less; Cameron puts the gap between leading proprietary models and much cheaper but still capable models at two orders of magnitude, citing GPT-6 Luna at 7 cents per task.",
      "The cost of reaching a given intelligence tier falls about 10x in around a year: a score of 30 on the Intelligence Index cost a bit over $1 when first reached, about 10 cents a year later, and now 1-10 cents. By that logic, $1 million of spend might become $100k in a year and potentially around $10k in two.",
      "The Artificial Analysis Intelligence Index synthesises 10 evaluations covering areas such as knowledge work, scientific reasoning, coding and terminal use, and hallucination; Opus 5.5 currently leads it, but models do not always follow the general ranking, with the Gemini series very good at navigating complex spreadsheets in the AA Analyst Agent benchmark.",
      "Open weights models continue to stay within 3 to 9 months of the leading proprietary models, and GLM 5.3 Flash is cited as a much cheaper, open weights and very capable option."
    ],
    "takeaways": [
      "Track input tokens, cache hit percentage and turns per task as your core agent cost metrics, and check that prompt caching actually works, especially on self-hosted setups and when choosing inference providers.",
      "Treat reasoning effort as a cost decision even when keeping the same model, because it also changes turns and input tokens across the agent's trajectory.",
      "Sort tasks into two buckets: for limited intelligence tasks (e.g. 'How much did I spend on subscriptions last month?'), find the intelligence ceiling and the cheapest way to reach it; for unlimited intelligence tasks (e.g. analysing your business's cost drivers), ask how much you will pay for more intelligence.",
      "Use the intelligence-versus-cost Pareto to shortlist models, then test the shortlist in your own context and agents for smaller prototype workloads.",
      "When rolling out in production, create your own benchmark on a shortlist of up to five models to get trade-off curves for your own task, given the magnitudes of cost now involved."
    ],
    "topics": [
      "agents",
      "inference-cost",
      "prompt-caching",
      "benchmarks",
      "model-selection",
      "reasoning-effort",
      "open-weights"
    ],
    "tools": [
      "Artificial Analysis",
      "Artificial Analysis Intelligence Index",
      "AA Analyst Agent",
      "AA Briefcase",
      "Optima",
      "Anthropic",
      "Opus 5.5",
      "Fable",
      "OpenAI",
      "GPT-6 Luna",
      "Astra",
      "GPT-4",
      "GPT-3.5 Turbo",
      "Google",
      "Gemini",
      "Nvidia",
      "GLM 5.3 Flash",
      "vLLM",
      "MCP"
    ],
    "quotes": [
      {
        "text": "$21 per task, this costs over $200.",
        "at": "13:17",
        "url": "https://www.youtube.com/watch?v=prGP0huECPo&t=797s"
      },
      {
        "text": "there's no points for knowing that GLM 5.3 Flash is cheaper than Opus 5.5.",
        "at": "15:33",
        "url": "https://www.youtube.com/watch?v=prGP0huECPo&t=933s"
      },
      {
        "text": "So the challenge for us as builders is making the trade-off decision.",
        "at": "15:43",
        "url": "https://www.youtube.com/watch?v=prGP0huECPo&t=943s"
      },
      {
        "text": "I'm not getting any benefit from more intelligence.",
        "at": "17:46",
        "url": "https://www.youtube.com/watch?v=prGP0huECPo&t=1066s"
      }
    ],
    "words": 3617
  },
  "summary_url": "/#prGP0huECPo",
  "transcript": {
    "html": "/transcripts/prGP0huECPo.html",
    "txt": "/transcripts/prGP0huECPo.txt",
    "vtt": "/transcripts/prGP0huECPo.vtt"
  }
}