Harvard and Crypto AI Platform Chutes Release 6.1 Billion-Request Dataset, Find 99% of Repeat Queries Land Within 15 Minutes

A decentralized AI company and a set of university researchers just handed the public one of the biggest windows yet into how people and machines actually use large language models.

Chutes, an inference platform running on Bittensor's Subnet 64, partnered with researchers from Harvard and the University of Chicago to release a dataset covering 6,122,413,756 LLM requests, according to Crypto Briefing. The data spans 9,174 different models and 314,970 anonymized users, collected over a full year of production traffic from April 11, 2025 to April 12, 2026.

The dataset is free and public, posted on GitHub and a Harvard-hosted Amazon S3 bucket. It does not contain any actual prompts or AI responses. Instead it captures metadata: request timing, token counts, latency, and time-to-first-token measurements, according to Crypto Briefing. Over the year, Chutes says it processed roughly 35.8 trillion input tokens and 2.52 trillion output tokens.

The 15-Minute Finding

99% of repeat requests to the same model occur within a 15-minute window, per the dataset's documentation. For engineers building AI infrastructure, that means servers can cache recently-used data and route similar requests to the same machine without wasting much compute on requests that never repeat.

The research team found that "prefix-aware routing strategies" (sending similar requests to servers that already have relevant data loaded) can hit near-optimal cache performance with only minor load imbalance across servers, according to Crypto Briefing. Companies running AI at scale can cut costs significantly if they route traffic strategically instead of randomly.

The dataset also shows a clear trend line: output lengths dropped over the year, from hundreds of tokens per response down to fewer than 100. The researchers interpret that as a shift away from humans asking for long explanations and toward automated, "agentic" systems firing off short, targeted requests, according to Crypto Briefing. LLM Stats reports that current frontier models like OpenAI's GPT-6 Astra, Anthropic's Fable 5.1, and Grok 4.7 are increasingly benchmarked on tasks like coding (CursorBench, DeepSWE) and terminal automation, not open-ended chat—work that produces short, function-style outputs rather than essays.

Privacy and Incentives Worth Scrutinizing

Chutes ran an opt-in period from March to July 2026 where users got a 25% discount on their usage bill for letting their traffic be included in the dataset, according to Crypto Briefing. User identifiers rotate every three months as an added anonymization layer.

A fair skeptic would ask whether paying users for their data biases the sample. Heavy users chasing a discount might behave differently than the platform's average customer, and a dataset built around a financial incentive isn't automatically representative of AI usage broadly. None of the available reporting establishes that this happened, but it's a real methodological question anyone using this data for academic or commercial modeling should account for before drawing sweeping conclusions.

On privacy, rotating anonymized IDs every three months limits how easily a single user's full history can be reconstructed, but metadata alone, timing patterns, token volumes, and model choice can sometimes be enough to re-identify heavy or unusual users even without seeing actual prompts. No source here claims this has happened to anyone in the dataset, and Chutes' methodology documentation was built with that risk in mind. It remains an open question rather than a proven flaw.

Why a Crypto Platform Is Doing This

Chutes runs on Bittensor's Subnet 64, a decentralized network where payments flow through Bittensor's native token, TAO. Publishing a documented dataset co-authored with Harvard is also good marketing: it gives a blockchain-based AI infrastructure project academic credibility and a public case that decentralized inference can match centralized Big Tech data centers on real production traffic. That's a legitimate business incentive, not evidence of dishonesty, but readers should know the underlying commercial motive sits alongside the pure research value.

The dataset is now sitting on GitHub and Harvard's S3 servers for any researcher, company, or grad student to download and dig through. The open question going forward is whether independent researchers, working with no financial tie to Chutes or Bittensor, replicate the same 99%-within-15-minutes caching pattern on other platforms' traffic. If they do, it becomes an industry standard. If they don't, it stays a Chutes-specific finding.