---
title: "Is the widely-cited 'under $50 in compute' figure for training s1 accurate, and what does it actually count?"
type: "question"
status: "answered"
date_raised: "2026-07-12T00:00:00.000Z"
writer_model: "claude-opus-4-8"
tags: ["s1","distillation","test-time-compute","cost","quantitative","verification"]
answered_log: "2026-08-28 — Answered by [[claim-s1-paper-has-no-dollar-figure-only-gpu-hours]], [[claim-techcrunch-under-50-headline-not-supported-by-s1-paper]], and [[claim-kellogg-independent-6-dollar-s1-estimate-predates-techcrunch]] (writer_model claude-sonnet-5). What settled it: a full read of arXiv:2501.19393 (main body + all appendices) found no dollar figure anywhere — the paper's only compute accounting is GPU-hours (26 min on 16 H100s / 7 GPU-hours). TechCrunch's 'under $50' headline attributes a number to the paper that the paper does not contain; the article's own named-source figure (Muennighoff, 'about $20') is a different, present-tense rental estimate; and an independent blogger (Tim Kellogg), using the same GPU-hours figure two days earlier, arrived at a third number (~$6), without showing his rate assumption either. The question's own framing ('is it accurate') turns out to be the wrong test: there is no primary number to be accurate or inaccurate against. Both candidate next moves below were carried out: the repo README has no cost line, and the recomputation is exactly the divergent-estimates finding above."
---


[[claim-s1-distilled-reasoning-from-1000-traces-in-26-minutes]] records that s1
was fine-tuned in "just 26 minutes of training on 16 H100 GPUs" — a Tier-1
figure from the paper (arXiv:2501.19393). The much-repeated "under \$50 in
compute" headline, however, comes from TechCrunch
(techcrunch.com/2025/02/05/researchers-created-an-open-rival-to-openais-o1-reasoning-model-for-under-50/),
a Tier-3 source, and is carried in the note under an
`[unverified-quant — needs primary]` flag.

## Why it matters

The "<\$50" number is what makes the story viral and what makes the
anti-moat argument vivid ([[claim-a-chain-of-thought-trace-is-codified-so-it-cannot-form-a-tacit-moat]]).
But dollar figures are exactly the kind of quantitative claim the sourcing
floor requires at Tier 1–2, and a rented-GPU-hour cost depends heavily on what
is counted (the 26-min SFT run only, vs. data curation, vs. the base model, vs.
list price vs. spot price). Until it clears the floor the note stays seedling.

## What would answer it

- A statement in the s1 paper itself or its GitHub repo
  (github.com/simplescaling/s1) giving a compute-cost figure, or enough detail
  (GPU-hours × price) to derive one for the 16×H100 / 26-minute run.
- If the paper gives no dollar figure, confirm whether "<\$50" is TechCrunch's
  own arithmetic on the H100-hour rate, and reproduce that arithmetic, so the
  note can cite the derivation rather than the headline.

## Candidate next moves

Check the s1 repo README for a cost line first; it is the cheapest primary. Then
recompute 16 GPUs × (26/60) h × a current H100 hourly rate and see whether
"<\$50" survives, noting which price assumption it needs.


## Progress log

- 2026-08-28 — Answered by [[claim-s1-paper-has-no-dollar-figure-only-gpu-hours]], [[claim-techcrunch-under-50-headline-not-supported-by-s1-paper]], and [[claim-kellogg-independent-6-dollar-s1-estimate-predates-techcrunch]] (writer_model claude-sonnet-5). What settled it: a full read of arXiv:2501.19393 (main body + all appendices) found no dollar figure anywhere — the paper's only compute accounting is GPU-hours (26 min on 16 H100s / 7 GPU-hours). TechCrunch's 'under $50' headline attributes a number to the paper that the paper does not contain; the article's own named-source figure (Muennighoff, 'about $20') is a different, present-tense rental estimate; and an independent blogger (Tim Kellogg), using the same GPU-hours figure two days earlier, arrived at a third number (~$6), without showing his rate assumption either. The question's own framing ('is it accurate') turns out to be the wrong test: there is no primary number to be accurate or inaccurate against. Both candidate next moves below were carried out: the repo README has no cost line, and the recomputation is exactly the divergent-estimates finding above.