🔮 The curious economics of a $6 AI agent #597

Good morning from London. The AI spend of your nightmares has hit $1.8 million for Amazon's Claude project, which ran for five months. A senior employee admitted that it's hard to determine the exact cost of AI-related expenses.

One of the authors experienced a similar situation with their OpenClaw agent, R Mini Arnold. The costs for this bot increased as it grew more complex, and it took time to switch to cheaper models. The bot once blew through $500 a day, leading to a tedious audit that revealed many processes were running older, more expensive models like Opus 4.5.

The price war between Anthropic and OpenAI, in response to Chinese advancements, has helped bring costs down. RMA, the bot, now uses the author's token allowance from their $200-a-month subscription and falls back to cheaper models such as DeepSeek v4 Flash or Pro if OpenAI is unavailable. For harder tasks, RMA can use OpenAI's top model, 5.6 Sol, or Anthropic's Fable, both of which are paid for by the token.

RMA is cheaper than before but more capable, as it plugs into the Manus API for certain tasks, uses Claude Code and Codex for coding, and has access to Prism (the internal research graph) and Elicit for academic papers. Despite the drop in costs, RMA still costs the author $2,000 per year.

The author wonders if the $494 a day spent on AI has disappeared from revenue, but acknowledges it was just an anomaly. In July, top companies spent a median of $7,400 per employee on AI, with the top 10% spending $650. The median firm spent $11.95 per employee.