Terence Tao calls out OpenAI's math benchmarks as scaffolded theater

What happened

On September 11, Terence Tao — Fields medalist, and probably the mathematician whose opinion the field weighs most heavily — published a blog post titled *A severe misalignment of AI in mathematics*. The post argues that the gap between what frontier labs claim their models can do in math and what those models actually do, unaided, has become large enough to constitute a research-integrity problem. The Economist picked it up the same day under the headline *Top mathematicians are outraged by OpenAI's methods*, and by Friday morning the HN thread was sitting at 777 points.

Tao's core complaint isn't that the models are bad at math — it's that the reported numbers describe a system, not a model. When OpenAI announces that its latest model solved N problems from the IMO or the Putnam, that number typically reflects a pipeline: a human-written prompt scaffold, external tool calls (Python, SymPy, Lean), best-of-K sampling with a verifier picking the winner, and in some cases retrieval over a corpus of solved problems. Strip those away and the raw model performance is dramatically lower. The reported number is real; it's just not a number about the model.

The specific incident that lit the fuse, per the Economist, was an OpenAI announcement claiming near-frontier performance on research-level problems, without disclosing that the runs used a substantial scaffolding harness and multiple independent samples per problem. Several named mathematicians — including at least two IMO problem-setters — went on record saying they had not been told their problems would be evaluated this way, and that if the same rules were applied to human contestants, the results would be disqualified.

Why it matters

This is the AI equivalent of a benchmark laundering scandal, and it matters beyond mathematics because the same pattern shows up everywhere frontier labs report capabilities. SWE-bench numbers depend heavily on the harness. GPQA numbers depend on whether the model gets to think for 30 seconds or 30 minutes. ARC-AGI numbers, famously, depend on whether you're allowed 1,024 samples with a verifier. The industry has quietly normalized reporting the ceiling of a scaffolded system as if it were the floor of a model's raw capability, and Tao is the first person with enough standing in a technical field to make the accusation stick.

The defense — which OpenAI and others make explicitly — is that scaffolding is part of the product. Nobody uses raw completions in production; everyone uses agents with tools. If the scaffolded system solves the problem, the system solved the problem. That's a defensible position for a product announcement. It is not a defensible position for a scientific claim about reasoning ability, which is what the math-community objection is really about. Tao's phrasing is careful here: he distinguishes between engineering achievement (real, impressive) and epistemic claim (misleading, and in some cases actively damaging to the mathematics community, which now has to spend time explaining to funders and administrators why an LLM did not, in fact, prove a novel theorem last week).

The deeper problem is a measurement one. Best-of-K with a verifier is a fundamentally different regime from pass@1, and the gap between them grows superlinearly for problems where verification is cheaper than generation — which describes most of formal math. A model that solves a problem 1 in 500 attempts, with an oracle picking the winning attempt, is not the same artifact as a model that solves it on the first try. Reporting the two under the same label is, to use Tao's word, misalignment. Every serious eval methodology paper of the last two years has said this out loud; every frontier-lab press release has ignored it.

Community reaction on HN split predictably. The 'this is obvious' camp pointed out that anyone reading benchmark papers carefully already knew scaffolding was doing most of the work. The 'this is a big deal' camp countered that 'anyone reading carefully' is roughly 200 people worldwide, and the other several billion — including the ones writing checks and setting policy — read the press releases. Both are correct. The Tao post matters not because the information is new but because the messenger changes what happens next.

What this means for your stack

If you're picking an LLM for a reasoning-heavy pipeline — code review, incident triage, contract analysis, anything where the model needs to actually chain steps rather than pattern-match — treat every published benchmark as a question, not an answer. Ask three things before you trust a number: what was the sampling budget, what tools did the model have, and what does raw pass@1 look like on the same task. If a vendor won't tell you, that's the answer.

For internal evals, this is a good week to audit your own methodology. If you're running best-of-N with a verifier in your test harness — and a lot of teams are, often without realizing it, because the eval framework does it by default — you're measuring the system, not the model. That's fine if the system is what ships. It's a disaster if you're using the number to decide which model to fine-tune, because you'll pick the model whose scaffolding is best-tuned to your verifier, not the one with the best underlying reasoning. The two diverge quickly.

The practical fix is boring and known: report pass@1, pass@k, and scaffolded-system numbers as three separate columns, and never let anyone in a slide deck collapse them. Internally, this takes a week. Externally, an entire industry has spent two years avoiding it.

Looking ahead

Tao's post will not, by itself, change how OpenAI or Anthropic or Google DeepMind reports benchmarks. What it may do is give reviewers at NeurIPS and ICML enough political cover to start rejecting papers whose 'model X achieves Y%' claims don't include a scaffolding ablation. That's a slow lever, but it's the right one — and if it works, expect the next round of frontier-lab announcements to quietly start including the pass@1 numbers they've been leaving out. Watch for whether the next GPT or Claude release includes a raw-capability column. If it does, Tao won. If it doesn't, the misalignment gets worse, and the next scandal will be uglier.