ANTHROPIC LEAKED A FILE THEY SPENT 4 YEARS BUILDING WHERE 10 AGENTS MAKE THEIR OWN LLM IN 60 MINUTES WORTH $1.2M A YEAR ten agents cost $0 an hour, and the model they assemble in that hour generates $100K a month corpus → clean → 5,000 pairs → LoRA → 300 tests → endpoint nine agents prepare the data, and only the tenth one rents a GPU for 40 minutes at $1.70 the tuning itself is 40% of the hour, and all the quality comes from the other five stages nobody pays for Haiku finds the corpus and cleans it, Sonnet 5 labels and tunes, Opus 5 writes 300 tests and then tries to break the result training a model from scratch is 100,000 GPU-hours - 150,000× more compute than anything here on a single GPU that's roughly a decade, and in this file it's 40 minutes you don't train a model in an hour, you tune someone else's - and that's exactly why it's possible at all your part is 0.2% of the weights, the other 99.8% someone else trained on their own money the layer is thin, but the data underneath it is the asset nobody can take from you $1.2M a year is 334 customers at $299, and that arithmetic doesn't move the tuning costs $1.70, the data you tuned it on costs everything else save this and paste it into Claude Code - let it build you these 10 roles ↓