Stanford and University of Washington AI researchers claim they trained AI reasoning model s1, distilled from a Gemini 2.0 model, for under $50 in cloud compute

AI researchers at Stanford and the University of Washington were able to train an AI “reasoning” model for under $50 in cloud compute credits …

TechCrunch 2025-02-06 Maxwell Zeff

Context & Ripple Effects

The result extends a cost-efficiency narrative already visible in DeepSeek’s claim that a competitive open-source model used fewer training chips. Here, the notable mechanism is distillation: researchers say a model derived from Gemini 2.0 can reproduce reasoning-oriented behavior without training a frontier system from scratch.

The claim also sharpens a key constraint on low-cost reasoning work: independent benchmarking of reasoning models can itself be expensive. Later coverage of researchers pursuing “frugal AI” on open-weight systems suggests that access to capital and top-end chips is becoming less decisive for some model-development paths, even if it remains decisive for frontier pretraining.

First-order effects

Stanford and University of Washington researchers gain a low-cost proof point for using a stronger model’s outputs to train a smaller reasoning model; the reported spend is cloud compute for that training step, not the cost of creating Gemini 2.0.
Gemini 2.0 becomes the upstream source model in this experiment, while s1 offers researchers a comparatively inexpensive artifact for testing reasoning-model training and evaluation methods.

Second-order effects

Labs and startups can put more emphasis on distillation, curated training data, and evaluation design rather than treating large-scale pretraining as the only route to reasoning capabilities.
Comparisons among low-cost reasoning models will put more weight on reproducible benchmarks and evaluation budgets, because a cheap training claim alone does not establish broad capability or reliability.

Third-order effects

If such results generalize, the market may separate further between a small group that funds frontier base models and a wider group that adapts, distills, and specializes their capabilities.
Reasoning economics may increasingly turn on access to high-quality teacher models and credible measurement, not merely on raw training-compute expenditure; the extent of that shift depends on whether results hold beyond narrow tasks.

The trend: Reasoning-model development is moving toward a layered supply chain in which expensive frontier models seed cheaper downstream adaptation and specialization.

Discussion

@mims Christopher Mims on bluesky
The latest and most jaw-dropping entry in the “there is no moat for frontier model companies” files: — For $50, researchers created an open rival to OpenAI's o1 ‘reasoning’ model — The model is a distillation of a Google “thinking” model
@yungherzog Herzog on bluesky
Incredible day to be a big tech hater. [embedded post]
@edzitron.com Ed Zitron on bluesky
This is so funny god damn. After all the attempts to tear down DeepSeek's training costs it might turn out that anyone can do it on the cheap. Amazing! [embedded post]
@edzitron.com Ed Zitron on bluesky
Hahahahahaah oh hell yeah baby. Yeah! This is the GOOD STUFF!!!!!' — finance.yahoo.com/news/alphabe... [embedded post]
@khunger Vineet Khunger on bluesky
That does sound really impressive, till you hear that I trained my AI theater acting + quantum cryptography model, the Vineet A1 which is out next week, on a toaster and a pair of Bata shoelaces [embedded post]
@gcampax.com Giovanni Campagna on bluesky
It's a good result, but the description is pure click-bait. We've known you can distill for cheap for a long time, ever since Alpaca distilled the original GPT4. Distill a strong model (Gemini Flash with reasoning), you get a strong-ish model. [embedded post]
@arankomatsuzaki Aran Komatsuzaki on x
Stanford presents: s1: Simple test-time scaling - Seeks the simplest approach to achieve test-time scaling and strong reasoning performance - Exceeds o1-preview on competition math questions by up to 27% (MATH and AIME24) - Model, data, and code are open-source [image]
@sirajraval Siraj Raval on x
😮 OpenAI's o1 was replicated for $6 in 3 steps: 1. Pick 1000 perfect training examples 2. When AI tries to finish thinking, inject “Wait” 3. Force it to double-check (just like telling a student “show your work") 26min training. Runs on laptop. Code: https://github.com/...
@omarsar0 Elvis on x
Nice summary of the s1 paper. “In s1, when the LLM tries to stop thinking with ”</think>", they force it to keep going by replacing it with “Wait”. It'll then begin to second guess and double check it's answer. They do this to trim or extend thinking time (trimming is just [image…
@simonw Simon Willison on x
Published some notes on S1, a new reasoning model that was fine-tuned from Qwen 2.5 32B on just 1,000 examples and $6 of compute Includes notes on running it with Ollama and exploring those 1,000 examples with Datasette Lite https://simonwillison.net/...
@muennighoff Niklas Muennighoff on x
DeepSeek r1 is exciting but misses OpenAI's test-time scaling plot and needs lots of data. We introduce s1 reproducing o1-preview scaling & performance with just 1K samples & a simple test-time intervention. 📜 https://arxiv.org/... [image]

Chronicles