Claude Fable 5 closes 82% of a human research record in nanoGPT speedrun
Two releases landed tonight that say a lot about where AI is heading: a benchmark that finally measures whether models can do research rather than talk about it, and DeepSeek making its API cheaper exactly when nobody is watching.
Prime Intellect's NanoGPT Speedrun Frontier is the first public, at-scale test of autonomous AI research — and Claude Fable 5 lapped the field. The team ran 153 autonomous runs across 18 frontier models, each agent getting the same repo, a rulebook, one prompt, and eight H200 GPUs for up to eight days, tasked with beating the nanoGPT optimizer speedrun's baseline of 3,290 steps toward a validated loss of 3.28. Fable 5 closed 81.7% of the gap to the human record of 2,600 steps; Opus 5 managed 53.6% and Kimi K3 52.2% in the serial era, while GPT-5.6 Sol needed six days and roughly 2.9 billion tokens to reach 35.9%. The striking part isn't any single number but what the winners did: no run invented a new method — they all won on disciplined optimizer work like preconditioning, learning-rate schedules, and late weight averaging. Execution quality, not idea generation, is where frontier models now separate.
The timing matters too. Anthropic's internal automated-R&D eval runs its optimization on a single CPU node, and OpenAI's GPT-5.6 Sol system card cites less than a day on one H100 — this leaderboard is the first time those private yardsticks get a public, comparable arena, with traces, scratchpads, and per-run ledgers open for anyone to audit. We covered Prime Intellect's earlier finding in July that open models nearly match Opus 5 at AI research; the sobering update is that this time the open-weights pack trails badly, and the gap between first and tenth place looks like a capability chasm, not noise. If autonomous research skill becomes a headline metric, "which model" starts mattering as much as "how big."
DeepSeek is making its API cheaper every weekend, starting August 23 — off-peak pricing all day Saturday and Sunday, Beijing time. Per its official API docs, off-peak rates are half of peak rates (peak hours are 01:00–04:00 and 06:00–10:00 UTC), so from Sunday developers running batch jobs or scheduled pipelines on V4 Flash or V4 Pro get the same work done at half price by shifting it into the weekend. It's a clever use of idle weekend GPU capacity, and it doubles down on the peak-hour billing system we covered when it launched — DeepSeek hikes API prices up to 4.7x with new peak-hour billing — after which Chinese labs largely quit undercutting each other. Weekend discounts don't reverse that repricing; they monetize the quiet hours instead. Watch whether Zhipu, Moonshot, or Alibaba copy the move within a month — if they do, "when you run your job" becomes a real line item in every Chinese startup's inference bill.
What to watch: whether anyone replicates Fable 5's 2,726-step result on an independent stack — the traces are public, so the reproduction attempts should start within days.
Which would you rather bet on — models that execute research brilliantly, or labs that price compute smarter? Tell us in the comments.
Sources: Prime Intellect — NanoGPT Speedrun Frontier · Prime Intellect blog — Measuring Autonomous Research · DeepSeek API docs — Models & Pricing · Cailian Press
