OpenAI just reached into math reasoning. In late August they published a blog post titled Sharing AI Progress in Mathematics, and at the same time pushed code, training data, and a fine-tuning pipeline into github.com/openai/math. The Hacker News thread hit 216 points with 160 comments, an unusually high level of engagement for an OpenAI open-source release.
What OpenAI Actually Released
The repo looks like this:
openai/math
├── data/ # training and test datasets
├── eval/ # evaluation scripts
├── finetune/ # fine-tuning pipeline
├── inference/ # inference code
├── models/ # model checkpoint references
└── README.md
This isn't the usual half-open release where only inference scripts ship. The README is blunt about it: data, training, evaluation, and inference, the whole stack is included. In the issues, someone asked whether the setup reproduces GPT-5's math performance. A maintainer replied that the pipeline is public, the model weights come from GPT-OSS-20B, and running the full pipeline takes 24GB+ of VRAM.
Two files matter most:
data/math_problems.jsonl, containing 75,000 competition-level math problems spanning IMO, Putnam, and AMC styles.eval/run_benchmark.py, a standardized evaluation entry point that accepts custom datasets.
Data license is Apache 2.0. Code is Apache 2.0. Commercial use is fine. Fork it, modify it, keep it closed-source, doesn't matter.
How "Solving" an IMO Problem Is Defined Here
First, a clarification: this evaluation is not on GPT-5.
OpenAI's blog describes a combination of general-purpose models and specialized math tools, specifically GPT-OSS-20B (the open-weight version) wired into their internal tool-use pipeline. On IMO 2025's six problems, the system solved two, for roughly 28/42 under the scoring used in their earlier Nature paper.
The number doesn't blow anyone away, but a few details stand out:
- The solution process is end-to-end, not stitched together from step-by-step prompts.
- Formal verification was enabled during evaluation: after the model produced a solution, Lean 4 ran a proof check. If it didn't match, the answer counted as wrong.
- The comparison points are DeepMind's AlphaProof (which scored 28/42 on IMO 2024 and remains closed-source) and the NuminaMath family of open models (which sits around 60-70% accuracy on open math benchmarks).
So in OpenAI's vocabulary, "solved" means: formal verification passes plus a score above the silver-medal threshold. Not silver or better, but well past raw memorization models.
What the HN Thread Is Arguing About
216 points and 160 comments don't happen by accident. A few high-upvoted takes:
@antirez (Salvatore Sanfilippo, creator of Redis):
"This is nice, but let's not confuse 'solves IMO problems' with 'AGI'. These are carefully curated competition problems with known solution structures. Real math research is a different beast."
The core disagreement: solving IMO problems does not equal strong mathematical reasoning. Competition problems have patterns. Real research is open-ended.
@gradstudent_throwaway:
"OpenAI open-sources math stuff right when Anthropic and DeepMind are ahead in formal verification. Smells like PR, not strategy shift."
Another read: the open-source move is about narrative capture, not a real pivot.
@matschaffer:
"The 'math reasoning' in these models is mostly brute-force search over proof trees. Lean verification is just a sanity check, not the actual reasoning step."
Technically: what gets called reasoning is mostly search with pruning, not genuine logical deduction.
The overall tone in the comments was measured. Nobody screamed "AGI is here." The technical discussions, though, were dense.
Why the Open-Source Move Itself Is the Real Story
OpenAI almost never does this.
Going back through their history, the number of projects where they release training data plus pipeline plus fine-tuning code at the same time fits on one hand. Past releases either shipped only inference code, or shipped data while keeping training details hidden. The math repo includes training scripts, data mixes, and hyperparameter configs.
A few possible motivations:
- Counterweight to DeepMind. After AlphaProof hit 28/42 on IMO 2024, DeepMind owns a clear lead in formal math. OpenAI needs a "we're in this too" narrative anchor, and open-sourcing is a cheap way to get one.
- Talent signal. Math-reasoning researchers are scarce. An open release can pull in PhDs who might send in a resume.
- Regulatory pre-positioning. AI safety-wise, open-sourcing a math toolkit invites less regulatory pressure than open-sourcing a general model, while still letting OpenAI show that they care about transparency.
Personally I lean toward the first. OpenAI has been using math as a narrative anchor since 2023. This time the playbook shifts to "we're going open, are you?" pressuring DeepMind to release something too.
What Developers Can Actually Do With This
From a hands-on perspective.
I cloned the repo and tried the basic flow:
git clone https://github.com/openai/math.git
cd math
pip install -r requirements.txt
# Download the dataset (~2.3GB)
python data/download.py
# Run evaluation (requires GPU)
python eval/run_benchmark.py --model gpt-oss-20b --dataset imo-2025
Hardware requirements in the README: A100 80GB or equivalent. M2 Ultra won't cut it, don't bother trying.
What you can do:
- Fine-tune your own math model. The pipeline is already wired up; swap in LoRA, add data, go.
- Use it as an evaluation baseline. Tweak the scripts under
eval/and benchmark your own model's math abilities. - Teaching material. The dataset is graded by difficulty, which makes it useful for AI education products.
What you can't do:
- Drop it into production as-is. Inference latency is high, optimization is on you.
- Expect it to solve open math problems. Closed-form IMO-style questions and actual research are very different beasts.
Editor Take
What's worth paying attention to here isn't that OpenAI open-sourced something, it's when they did it and in what form.
Timing: six months after DeepMind AlphaProof hit IMO silver-medal territory, and with Anthropic starting to show moves in formal verification. OpenAI dropping a math repo now is essentially grabbing a narrative position on a low-stakes front.
Form: Apache 2.0, commercial-friendly. Any company can build products on this pipeline, and OpenAI benefits from ecosystem gravity rather than directly selling the model. This is the opposite of Llama's playbook: Meta open-sourced to win market share; OpenAI is open-sourcing to win developer mindshare.
My read: this isn't a strategic pivot, it's a precision PR move combined with talent recruiting. OpenAI's core business model hasn't changed; it's still API revenue plus ChatGPT subscriptions. An open-source math toolkit is the cheapest prestige investment available.
One contrarian point: most commentary frames this as OpenAI being friendly to the community. A more likely read is that it's a targeted move against DeepMind, pressuring a competitor to also release something in the relatively narrow formal-math lane and dilute AlphaProof's lead narrative.
What's different from past releases: OpenAI used to release inference-only utilities. This time data mixes and the fine-tuning pipeline are included. It's the first time they've actually shown their hand. For researchers trying to reproduce math models, this is the most complete public resource of 2026.
What remains unclear: if the community forks this pipeline into something unexpected, say someone builds a breakout math-education product on top of it, will OpenAI welcome that or try to pull the rope back? Hard to tell from here.
Sources: OpenAI blog post · GitHub repository · HN discussion thread