In September and October 2026, OpenAI, Anthropic, Google, Meta and Mistral all shipped a new model, each with a table in which their own model wins. Those tables can all be accurate and still say nothing about your work.
AI benchmarks are standardised test sets used to measure language model performance, such as a collection of coding tasks, exam questions or agent tasks with a fixed scoring method. They're useful for ranking models roughly, but rarely for choosing one: gaps at the top often fall within measurement error, the tasks look little like business work, and almost no benchmark measures Dutch. For an investment decision you need a test set of your own.
Why do the vendor's own numbers tell you so little?
A vendor's numbers tell you little because the vendor chooses which tests go in the table, which settings they run with and how often. All three choices favour the vendor's model.
An October 2026 example. At the launch of Mistral Large 4, Mistral wrote that, through the evaluator Vals.ai, the model beats GPT-6 Astra on legal and financial tasks. Vals.ai's own pages put Mistral Large 4 one point ahead on the finance agent test, inside the error margin, and almost eight points behind on the legal research test. On the overall Vals Index it ranks 34th of 49; GPT-6 Astra ranks 7th. What else we know is in our piece on Mistral Large 4.
It works the other way too, honestly. With Claude Opus 5.5, Anthropic published 66.4% on Terminal-Bench 4.0 against 57.9% for GPT-6 Astra, and added itself that the standard error is about 2.6 points and that "benchmark margins have become a less reliable guide to real-world differences", in its own announcement. A vendor writing that down says a lot.
The research institute Epoch AI has documented the gap for a while. A model whose developer reported 65% on the science test GPQA Diamond scored 55% in Epoch's own measurements, averaged over sixteen runs, according to Epoch's methodology page. Developers often don't disclose the prompt or temperature they used.
Where does a benchmark score go wrong?
A benchmark score goes wrong in three places: the test is in the training data, the test is maxed out, or the score depends on wording. All three have been measured.
| Problem | What happens | Measured example |
|---|---|---|
| Contamination | Test questions are in the training data and the model recognises them | On GSM1k, a fresh set of maths questions in the style of the well-known GSM8K, models scored up to 8 percentage points lower |
| Saturation | All top models score near the maximum, differences vanish | Models exceed 90% on MMLU, which is why Humanity's Last Exam was built |
| Prompt sensitivity | Small formatting changes in the prompt change the score | Up to 76 percentage points of difference from changing prompt formatting alone |
The last row matters most for a company. If a model's score swings that much with wording, a benchmark run with the vendor's prompt says little about how the model performs with the prompts your system will send.
Leaderboards built on human preference have their own weak spots. LMArena has visitors compare two anonymous answers. The paper The Leaderboard Illusion found that Meta tested 27 private variants on the platform before the Llama 4 launch, and that Google and OpenAI together received about 40% of all vote data, against 29.7% for 83 open models combined. A vendor allowed to practise more scores higher. Length counts too: LMArena now corrects for answer length and formatting, because longer answers with more bullet points systematically win more often.
Which independent leaderboards are worth using?
Independent leaderboards are worth using when they test every model under the same settings, publish their method and show an error margin. Two meet that bar for a Dutch company.
Artificial Analysis runs every model with the same prompts and temperature on ten evaluations (Intelligence Index v4.3.2), weighted towards agent tasks and coding. Some tests are kept private to prevent contamination. The methodology page also says what the index isn't: "a primarily text-based, English-language evaluation suite". Its separate multilingual index covers sixteen languages, not Dutch. And the index changes often: four versions in September 2026 alone. Never compare scores across versions.
EuroEval, formerly ScandEval, does measure Dutch. It tests models on Dutch tasks: sentiment, recognising names and organisations, reading comprehension, summarising, simplifying into plain language, detecting hallucinations and calling functions. Each model runs ten times on resampled test sets, so the Dutch leaderboard shows a margin per score. In October 2026 GPT-6 Astra and GPT-6.1 Sol topped it, with the open model GLM-5.3-Flash within the margin of third place. Claude 5.5, Gemini 4 and Mistral Large 4 weren't listed yet. A leaderboard always lags the launches.
Use them for shortlisting. A model clearly lower on EuroEval doesn't need your own test for Dutch text. Between models within each other's margin, your own test decides. How far open models lag in general is covered in open-weight models versus the frontier.
How do you build your own test set for an AI model?
You build your own test set for an AI model from real examples of the work it will do, with what counts as a good answer fixed in advance. This is how we do it.
1. Collect 50 to 200 real cases
Take them from your inbox, ticketing system or archive, including the awkward ones: the customer who asks three questions in one email, the invoice with an unusual layout, the contract in English. OpenAI's evaluation guidance warns against test sets that "don't faithfully reproduce production traffic patterns". Fifty is enough to see large differences. For differences of a few percent you need hundreds.
2. Define what's correct for each case
For extraction tasks it's simple: the right amount, the right date, the right category. You score those with code, which Anthropic calls the fastest and most reliable method. For written text, write a rubric: which information must be in it, what must never be.
3. Have another model grade, and check the grader
A language model as grader works, with known biases. The original research on the method found over 80% agreement with human graders, about the level humans reach among themselves, but also a preference for the first answer shown, for longer answers and for its own answers. So use a different model as grader than the one under test, and have a colleague review twenty gradings before you trust the rest.
4. Run each model several times
Run the full set at least three times per model. If the score moves more between runs than between two models, you have no winner. For one agent test, Artificial Analysis reports how often a model answers a question correctly five times in a row for exactly that reason: an answer that's only sometimes right still needs checking.
5. Measure cost and speed too
Record what a run costs and how long it takes per model. A model that scores two percent better but costs four times as much per task is rarely the right choice. How to calculate that is in AI model cost per task.
When should you switch to a new model?
Switch to a new model when it's demonstrably better on your own test set, outside the run-to-run margin, at a cost per task worth the difference. A launch benchmark table is never enough for that.
The test set you build for your first model choice then becomes your main operational tool. You run a new release through it within a day, and you know whether the launch changes anything for you. You use the same set for every change to prompts or integrations, as we describe for AI agents in production. With open models you run yourself, the test set also decides whether a smaller model is enough, the subject of small language models for business.
For the wider question of how to judge AI investments and who helps with that, see our guide to hiring AI consulting. Want a test set built for your own processes, so you know where you stand within a day of every release? We do that as part of AI consulting.