Why AI Benchmark Scores Are So Often Misleading

On 30 September 2026, Google launched Gemini 4 Argon with claims that sound strong: 77.9 percent on DeepSWE v1.1, ahead of GPT-6 Astra, Claude Fable 5.1 and Opus 5.5. First place on knowledge work according to Vals AI. Then an analyst at Jefferies added one sentence: benchmark wins still need validation in the real world.

Status note: unverified. The 77.9 percent DeepSWE figure and the Vals AI ranking come from third-party measurement, but the results have not been published as a complete report that can be re-checked. The comparison against other models may differ when configuration, cost, or allowed task types are not matched. The Jefferies statement is analysis, not a measurement.

Those four sentences are a good enough example of why a leaderboard cannot be taken at face value. The claims are real, the method is explained, and none of it answers the question that matters: is this model better at the work people will actually do.

Four reasons the numbers mislead

First, the test is chosen, not natural. A benchmark does not emerge on its own. A group of people picks what gets measured, and that choice determines the outcome. When a product team picks the benchmark it wins on, that is technically legitimate and entirely self-serving.

Second, the testing is often the vendor's own. The numbers above come from internal testing. In internal testing the cases are known in advance and can be adjusted. That does not make the numbers false. It means they measure an easier condition than users meet.

Third, training data contamination. If benchmark questions already sit in a model's training data, a score on that benchmark is no longer a measure of ability. Older benchmarks are the easiest thing for a model to memorise.

Fourth, what is being measured is a small delta. As the base improves, a few points on a benchmark can come from a small change in decoding rather than from capability. On LiveBench, DeepSeek V4.1 Flash released on 10 September 2026 scored 81.1 against Anthropic's 83.4. That gap of about 3 percent is clean: it was about 15 percent earlier in 2026 and 9 percent in May.

What actually moves: the gap, not the ranking

Model standings change every month. What is more interesting is the distance between them, and that distance is shrinking on a visible pattern.

Fifteen percent early in 2026 became about 9 percent in May, then about 3 percent in September. That is not one big leap, it is three consecutive ones. If the pattern holds, in a few years the distance between the best model in the world and a good open-weight model may shrink to nothing.

What shrinks the gap is not a single discovery but three things happening together. Cleaner and larger datasets. Leaner architectures. And compute directed at reinforcement learning instead of spent entirely on pretraining.

A short history: why benchmarks exist

Before 2020, the way to judge a model was to look at sample outputs and decide. That failed for two reasons. First, there was no number comparable across papers, so two research groups could never measure the same thing. Second, and more quietly, the samples a paper chose were almost always its best ones.

Then public benchmarks arrived. GLUE in 2018, SuperGLUE in 2019, and SQuAD alongside them. Each gave one comparable number, and that changed the industry within two years.

But it had a side effect that was not visible at the time: a model trained specifically on the test set scores high on it without gaining ability anywhere else. That is benchmark contamination, and nearly every older benchmark has it now.

Which is why evaluation moved to harder ground. LiveBench is refreshed periodically so its questions are new. METR measures something else entirely: not how good an answer is, but how long a task takes a model to complete when work is measured in hours rather than seconds.

What happens when a benchmark is trained on

The obvious case is training on the test set directly. A model that has seen benchmark examples scores well on them and stays exactly where it was on anything new.

Harder to detect is indirect contamination: data that is similar but not identical. A textbook that contains practice problems of the same shape, API documentation that copied examples from a test suite, a forum where people discuss the tests. All of it give the model a look at the shape of the problem without ever seeing the answer.

That is why a benchmark created deliberately after a model shipped is worth more than an old one. In theory they measure the same thing, but one has been contaminated and the other has not.

How to read a number you did not produce

Four things to check before drawing a conclusion from a ranking table.

• Who ran it. Third-party testing is worth more than vendor testing. METR, the organisation that examined OpenAI's agent incidents, was named in the FTC probe opened on 30 September 2026; it is known for testing claims rather than selling models.

Table: four signs a number needs a second look

• Whether the score is their own. A vendor testing its own model is not automatically lying, but it is interested. A figure that really holds up has to be reproduced by somebody else.
• Whether the test predates the training data. For a model trained on the open web, almost every old benchmark is already inside it. The relatively clean ones were made or revised after the model's training window.
• How close the gap is. A difference under two points on any test should be read as a tie, not a win.

Numbers that need their status attached

Several figures in circulation deserve different labels.

The 99 steps down to 8 in the distilled Rho-1 come from Reka's own testing. The 6.70 seconds for a 14.3 second video in Video DeltaNet comes from its authors. Both are technically plausible and genuinely interesting, and neither is third-party verified.

The "kernel often picked 14 times" figure on Reka Stream is a vendor claim with an obvious leak: the model picks its own kernel, and there is no external control over what gets chosen. A number like that is worth reading as an internal report rather than as evidence.

The most commonly misread numbers are the ones with no comparison at all. "Pixel perfect" on one model means nothing if the reference image came from a different source, or if the prompt was a different length.

Analysis: why the leaderboard will not disappear

It will stay, and not because of prestige. The reason is structural. A ranking is the cheapest way to communicate something that is genuinely hard to communicate in prose.

Two models with very different architectures, data and decoding can be two points apart. The ranking loses all of that information. But without a ranking, comparison is hard, and the market will find another way to compare.

What changes instead is how a ranking is built. A leaderboard that includes a confidence interval, the sample size, and a date is more useful than one that just prints a winner in bold.

Deeper dive: the four numbers behind the score

The useful table is not the leaderboard. It is four questions that are easy to skip. This is the part not present in any original source.

The fourth question is the most revealing. Nearly every vendor page shows five best results and quietly drops the rest. One complete table showing the actual spread is more informative than ten repeated highlights of the same run.

The third question has a practical consequence. If two models differ by 1.8 points on any test, picking one on that number is close to picking a horoscope. What should decide is cost, latency and failure rate in your own environment, not a position in somebody else's table.

What to watch

Three things. First, whether contamination-resistant tests become more common. A community-built test made deliberately after a model was released has a different value from an old test the model has already memorised.

Second, whether vendors start auditing themselves and publishing the failures. When the best reports include what did not work, the numbers become more useful precisely because they do not look good.

Third, whether the gap between open-weight and frontier actually closes. If it does, the question of whether to use a closed model stops applying, and that is a bigger change than any ranking that comes out this month.