---
title: "The s1 paper's own compute accounting for training gives GPU-hours, never a dollar figure, and covers only the fine-tuning run itself"
type: "claim"
status: "seedling"
sources: [{"source_url":"https://arxiv.org/abs/2501.19393","source_author":"Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, Tatsunori Hashimoto","source_title":"s1: Simple test-time scaling","source_date":"2025-01-31 (v1); read as v3, 2025-03-01","source_venue":"arXiv:2501.19393 [cs.CL]","source_quote":"The training takes just 26 minutes on 16 NVIDIA H100 GPUs.","source_tier":1,"source_sha":"598932c9f96849cd52495d8b3e12ba4d224e41d5588d2278a78f76c744d8a3bc"},{"source_url":"https://github.com/simplescaling/s1","source_author":"simplescaling (Niklas Muennighoff et al.)","source_title":"s1: Simple test-time scaling","source_date":"content dated through 2025-03; accessed 2026-08-28","source_venue":"GitHub, simplescaling/s1 (README)","source_tier":1,"source_sha":"a476a2e6a9a8d19d5d3ceb9dc49ed73fc76ceafa7e5d50703aa609cc3afa5f86"}]
audit_status: "capture-verified — the capture read the full text of arXiv:2501.19393 (main body plus all appendices, including Appendix D 'Training details') via extract_pdf at capture time searching specifically for any dollar figure, and separately read the GitHub README; neither contains one. Promotion (headless, no network) did not independently re-fetch. | AUDIT 2026-08-29 (claude-fable-5, cross-model; writer claude-sonnet-5): independently re-fetched arXiv:2501.19393 via extract_pdf (sha256 byte-identical to source_sha, 46pp) and swept the full text — no '$', 'dollar', or 'USD' anywhere; 'The training takes just 26 minutes on 16 NVIDIA H100 GPUs.' verbatim (§2 and Appendix D 'Training details'); 'To finetune on 59K samples, we use 394 H100 GPU hours while s1-32B only required 7 H100 GPU hours.' verbatim; 59,029-question pool confirmed. GitHub README re-fetched: 'we recommend 16 H100 GPUs i.e. 2 nodes with 8 each', no dollar figure. CONFIRMED, no changes."
provenance: "Promotion from 10-inbox/raw/2026-08-28-is-the-widely-cited-under-50-in-compute.md, 2026-08-28 (headless)"
origin: "batch"
derived_from: "10-inbox/raw/2026-08-28-is-the-widely-cited-under-50-in-compute.md"
writer_model: "claude-sonnet-5"
date_created: "2026-08-28T00:00:00.000Z"
tags: ["s1","distillation","test-time-compute","cost","quantitative","compute-accounting","reasoning-models"]
audits: ["2026-08-29 claude-fable-5"]
seek_code_commit: "7d6d9ed"
---


A full read of arXiv:2501.19393 — main body and every appendix, including
Appendix D, "Training details" — turns up no dollar ($) figure anywhere in
the document. The paper's own accounting of what training cost is entirely
in GPU-hours: "The training takes just 26 minutes on 16 NVIDIA H100 GPUs,"
and, comparing the final run (fine-tuned on 1,000 curated examples) against
an ablation trained on the full 59,029-question pool it was drawn from,
"s1-32B only required 7 H100 GPU hours" against 394 for the full-pool
version. The paper's own hosted repository (github.com/simplescaling/s1),
which carries the training scripts, model weights, and data, gives the same
"16 H100 GPUs" hardware recommendation and likewise states no cost in
dollars.

Both figures cover only the supervised fine-tuning (SFT) run described in
[[claim-s1-distilled-reasoning-from-1000-traces-in-26-minutes|the vault's
existing note on that run]]. Neither source prices two upstream steps the
run depended on: generating the 1,000 distilled reasoning traces via the
Gemini 2.0 Flash Thinking Experimental API the training data was drawn from,
or pretraining the base
[[claim-hoffmann-2022-compute-optimal-scaling-splits-equally-between-parameters-and-data|compute-optimally
trained]] Qwen2.5-32B-Instruct model the SFT run started from. "26 minutes
on 16 H100s" is real and Tier-1, but it is an accounting of one visible step
in a longer, partly unpriced chain — the fact that resolves what a widely
circulated "under \$50" figure could and could not have been checked
against; see
[[claim-techcrunch-under-50-headline-not-supported-by-s1-paper]].

> [!note] Seek's commentary:
> "26 minutes, 16 H100s" is a genuinely cheap number, and it's real — Tier 1,
> stated twice, in the paper and in the repo. What it isn't is the whole
> bill. The trace-generation API calls and the base model's own pretraining
> are the two steps nobody in this chain ever priced, and they're exactly
> the steps that don't fit on a slide: harder to meter, less flattering to
> the "reasoning is now this cheap" story. A number this precise, this
> quotable, and this incomplete is how a partial accounting gets read as a
> total one. — Seek