{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreifa4wbru5qk246ydfxs2neiehnhfvmxfkn7btzcwww4zztteih5dq",
    "uri": "at://did:plc:3hmagzdut76xycuvulaezm3j/app.bsky.feed.post/3mkxl72dy3qc2"
  },
  "path": "/2026/05/03/burning-tokens.html",
  "publishedAt": "2026-05-03T16:20:30.000Z",
  "site": "https://www.manton.org",
  "tags": [
    "Michael Tsai quoted this",
    "a trailing narrative"
  ],
  "textContent": "With the higher GPT-5.5 pricing, I wasn’t sure if in practice it would matter, so I preemptively disabled `/fast` in Codex. Seems like a non-issue. I can’t get anywhere close to using half of my tokens. Whenever I check it’s at 80% or higher remaining.\n\nI’m now back to running “high” and `/fast` for everything. If I think a problem is difficult, I’ll bump to “xhigh” and won’t think twice about it.\n\nMichael Tsai quoted this from X:\n\n> I have a friend in apple.\n>\n> He has over 200 dollars credit on claude everyday to spend.\n\nI find this difficult to believe. Even though Claude Code is more stingy with tokens than Codex, $200/day seems outrageous. If this is even close to true, it’s no wonder Anthropic is making so much money. But the popular opinion about Claude might be a trailing narrative, before everyone notices how good GPT-5.5 is.",
  "title": "Burning tokens"
}