A model posts a 92% on a reasoning benchmark that sat at 40% eighteen months ago. The chart is blue and soaring. Everyone shares it. Then a grad student tweets that the same model fails on the same problems when you swap the variable names from a and b to Alice and Bob. That gap between the headline and the reality is where most AGI benchmark results live, and learning to read it is a genuinely useful skill right now.
This isn’t a knock on benchmarks. They’re the only shared ruler we have. But a ruler tells you nothing until you know what it’s made of and who’s holding it. Here’s how to look at the numbers like someone who builds them rather than someone who retweets them.
What AGI benchmarks are actually measuring (and what they aren’t)
No benchmark measures general intelligence, because nobody agrees on a definition precise enough to test. What benchmarks measure is performance on a fixed set of tasks with a scoring rule. That’s it. ARC-AGI measures pattern completion on abstract visual puzzles. GPQA measures graduate-level science questions that a domain expert can answer but a search engine struggles with. SWE-bench measures whether a model can resolve real GitHub issues in a repository. Humanity’s Last Exam measures breadth across expert-written questions designed to resist memorisation.
Each of those is a proxy. A good proxy correlates with something you care about. It never becomes the thing itself. When you see a model claim a record on one of these, the first question isn’t “how high?” It’s “high on what version, with what scaffolding, and scored by whom?”
The scaffolding problem
Many benchmark results are produced by a system, not a bare model. The model generates candidates, a verifier ranks them, a tool executes the code, a retry loop runs until something compiles. That’s a legitimate engineering approach — real products work this way — but it means a “single model” score can hide a small pipeline behind it. If a lab reports a jump from 34% to 71%, ask whether the baseline was also run with the same harness. If not, part of that jump is scaffolding, not capability.
The saturation problem
Benchmarks die. MMLU was a serious test in 2021 and is now a formality; top models cluster near the ceiling and the differences between them are noise. When every frontier model lands within two points of each other, the benchmark has stopped discriminating. This is why new ones keep appearing, and why a sudden burst of fresh benchmark names is usually a sign the old ones ran out of headroom rather than a sign of sudden breakthroughs. If you want to track which thresholds genuinely moved and which just got retitled, the milestone tracking on this site is a decent running record.
Four numbers to check before you believe a score
Most reporting skips the details that determine whether a result means anything. These four matter more than the headline figure.
- Sample size and variance. A 3-point gain on a 200-question eval is often inside the error bars. Ask for confidence intervals. Labs rarely publish them, which is itself information.
- Contamination. If the eval questions leaked into training data, the score measures recall, not reasoning. Serious benchmark authors now hold out private sets for exactly this reason.
- Pass@k versus pass@1. A model that solves 60% of tasks given ten attempts is a different product from one that solves 60% on the first try. Both get written up as “60%.”
- Cost per correct answer. If the new model is 5 points better but 40× more expensive to run, that’s a research result, not a capability you can use.
A useful habit: read the evaluation appendix before the abstract. Papers that hide their setup in a footnote are usually hiding something else too. There’s a method to skimming these efficiently, and this guide to reading AGI research papers quickly saves a lot of wasted afternoons.
Where benchmark gains quietly stop translating
The strangest thing about the last two years of results is the mismatch between benchmark curves and felt capability. Models got dramatically better at coding benchmarks, and moderately better at actually shipping a feature in a messy codebase. Models aced exam-style chemistry, and still confidently invented a plausible-sounding reagent that doesn’t exist.
The reason is that benchmarks are clean. They have a right answer, a clear prompt, and no ambiguity about when the task is finished. Real work has none of those. It has a stakeholder who wants something slightly different from what they said, a data pipeline that silently drops rows, and a definition of “done” that changes on Thursday.
So a 15-point jump on a clean benchmark can translate into almost nothing in practice, while a 5-point jump on a benchmark built from real messy tasks can change how a whole team works. Direction of travel is not the same as distance travelled. That distinction is the whole argument in this review of what actually progressed and what stalled, and it’s worth internalising before you quote any single figure.
Why labs have an incentive to publish flattering results
Benchmark results are marketing. That’s not cynicism, it’s just what happens when a number drives fundraising, recruiting, and press coverage. The structural pressure produces a few predictable patterns.
Selective reporting is the most common: a lab runs the new model on a dozen evals and publishes the four where it wins. Chained comparisons are the next: new model compared to an old baseline, never to the current competitor, using a harness tuned for the new model. Then there’s the “frontier claim,” where a result on one narrow benchmark is presented as evidence of general advancement.
None of this requires bad faith. It requires a marketing team and a deadline. The practical filter is to look for third-party reproduction. Independent evaluation organisations and academic groups re-running the same evals are far more informative than the original announcement. If nobody outside the lab has reproduced a result after a few months, treat it as provisional. Telling a real announcement from a press beat is a skill you can build with a checklist, and it pays off every launch cycle.
What a genuinely informative result looks like
The results that deserve attention tend to share a shape. They show a capability appearing where the model previously scored near zero — not a smooth improvement from 70% to 78%, but a jump from 4% to 55% on something that was considered out of reach. They come with ablations showing which component caused the gain. They include failure cases in the main text, not buried in an appendix. And they survive a change of surface form: the same model that solves the problem with one set of names solves it with another.
That last test is the cheapest one you can run yourself, and it’s brutal. Rephrase the question. Shuffle the answer order. Add an irrelevant sentence. If the score collapses, the model learned the format, not the reasoning. A benchmark result that survives paraphrase is worth ten that don’t.
Watch for the evaluations that build in adversarial variation from the start, and for labs that publish their own negative results. When a team reports that their method didn’t help on three of five tasks, that’s a signal the other two numbers are trustworthy. Scarcity of that behaviour is the single best reason to be sceptical of the whole genre.
So read the numbers. Read them closely, and read the appendix. But hold them loosely — the benchmark that looks decisive this quarter will look quaint in eighteen months, and the capabilities that matter will still be the ones nobody figured out how to score.