Testing Math Conjectures with AI Burns Over $350 in Tokens Per Session
doodlestein · x · 2026-07-30
A developer attempting to use an AI agent (Fable) to prove or find a counter-example to Kaplanski's Zero Divisor Conjecture documented an astonishing consumption of compute.
Due to the complexity of the task, the agent consumed millions of tokens in just a couple of days. At API rates, this single session cost over $350 to run (not counting another couple million tokens from a previous session that was rate-limited).
Related event: AI Agents Face High Costs and Bottlenecks in Math Research(3 posts)→
More from coding & agent
- Context Compaction Unlocks ARC-AGI-3 SOTA: The Untapped Potential of Harness Engineering — eldonredwards · 2026-07-30
- Meta & CMU Paper: Agentic Context Management Boosts Long-Horizon Task Performance by 27% — rohanpaul_ai · 2026-07-30
- From Micromanaging Processes to Pure Intent: The Evolution of Agent Interaction — intellectronica · 2026-07-30
- Ghostwriter: An Open-Source Plugin for AI-Generated Terminal Tab Titles — smtabatabaie · 2026-07-30
- Managing 10 Agents vs. 100 Humans: The AI Era Management Debate — HanchungLee · 2026-07-30
- OpenAI Codex Security complained about: blocks content after 20 mins of running — VanillaMammoth1937 · 2026-07-30