AI benchmarking is like competing who has bigger Schwartz. It keeps the competition going, but outside of a small group of people, the numbers have almost no meaning.
To illustrate my point, let’s go to BenchLLM. We can see that Gemini 4 Carbon has a score of 88.79. The Agentic Score is not known right now, though the Coding score is at 81. Coming in at a close second is Claude Opus 5.5 with a score of 86.4, an Agentic score of 90 and a Coding score of 83. Third is GPT Astra with 86.04, 76 and 75, respectively.
Of course, that’s what BenchLLM people think. People at Artificial Analysis rank Gemini 4 Carbon (high) at 60, Claude Opus 5.5 (max) at 59 and Gemini 4 Carbon (medium) at 58. (These scores are based on the Intelligence Index.)
To get more numbers, you can visit LiveBench, Arena Leaderboard or Scale AI.
Everyone has their own methods and is getting their own numbers. But they are all Schwartz competition numbers — like ‘DistressBench’ which gives you the quality of crisis support that the particular model delivers.
This is the problem with benchmarks. If I take you to the year 1989, we were comparing 286, 386 and 486. While 286 was between 6 MHz to 25 MHz, the 386 was 12 MHz — 40 MHz, and the 486 was 16 MHz — 160 MHz.
This probably looks like a series of numbers with no meaning. But I’m sure some of you recognize that I’m talking about Intel processors and the respective clock speed they were able to achieve. Back then, these were mind-boggling speeds. Imagine the marketing people at that time, trying to entice you to buy a computer which can clock 160 MHz. And in case you want to know how far we got since then, the Intel Core i9–14900KS, when overclocked and cooled with liquid nitrogen, can get past 9.2 GHz. That’s almost 58 times faster. Note: MHz and GHz stands for Mega/Giga Hertz.
But how many people know what processor they have in their computers, phones, tablets or watches? Most people find this kind of thing so boring!
And this is where we are with AI, when it comes to understanding what it can do and how we can measure anything useful coming out of it.
Not only we don’t know how these models work inside, we have a hard time designing meaningful metrics.
Here is a conclusion from Prime Intellect, a company which trains and deploys custom AI models: ‘Designing evaluations for frontier models is hard, because capabilities are advancing quickly. Evaluations with ambiguous prompts, broken graders or setups that allow for reward hacking risk giving a false impression of current capabilities and shortcomings.’
Here’s another one from Interconnects AI: ‘The problem with trying to benchmark real-world usage today is that it is extremely jagged, just like the models.’ And: ‘The benchmarking industry comes down to trust. I have a solid amount of trust for Artificial Analysis, but not the most I do for anyone in the ecosystem. Artificial Analysis’s business model is to make money from the labs to verify their results. They make a lot of money doing this, which is fine. Other benchmarking companies have a similar problem, such as Arena, which does pre-release testing for labs to pay the bills.’
Or here’s a great article from Quanta Magazine, titled ‘He Wants AI Benchmarking To Be a Science, Not a Dark Art.’ In the article, Moritz Hardt stated: ‘The benchmarking ecosystem is broken.’
And this post, titled ‘Teaching an old monkey new tricks,’ shows that AI vendors do everything to score high, including cheating. Money talks.
Here is an idea about how to improve things.
Perhaps one day, the vendors of these AI models will realize that if they want businesses to adopt this new technology, they will have to spell out what exactly it can be used for and how to measure its accuracy. This story, ‘The cognitive dissonance of the Co-World,’ shows that, based on the current architecture, AI models can’t achieve two things at the same time: certainty and scope. The higher the certainty, the narrower the scope and vice versa.
Further, the vendors have to create a level of trust, where the customer will be provided a performance guarantee.
We can use an analogy of FDA approvals before the drugs are ready for the market. The vendor identifies a symptom or disease to be cured, runs lab tests, moves things to human trials (based on the results) and eventually gets approved. Everything is documented, tests are traceable and the scope is well defined. Now let’s use this analogy for AI models deployment. The vendor identifies a use-case, trains the model, states the verifiable outcomes and puts the trained model on the market.
Then, to create a level of trust, an independent certification body creates a public data set for the model training and certifies the results. Yes, the argument for model overfitting is valid. However, it is up to the vendor to take the data set, split it into training and validation sets and then tune it for the best result.
That’s the vendor side. The certification body then provides a way for the customer to upload their own subset of data and compare which model performs better within their own context.
Suddenly, the number of parameters, tokens or an Intelligence Index are completely irrelevant. The score for various use cases is the simple number which will help the customer decide what to buy.
What would be examples of the use cases?
It should be something which fits this criteria:
the correctness can be scored,
ground truth is easily established,
drift can be measured
The use cases could be then:
document data extraction — invoices, bank statements, forms
KYC/AML alert triage — banks or any financial institution requires this by law
clinical documentation and coding
contract clause extraction
I can hear the argument that AI benchmarked at the PhD level can do much more. While this is true, it rings hollow because with wide scope, it loses its accuracy. And to the business, accuracy and reliability trumps everything.
The recurrent pattern? Every time a new technology is introduced, the euphoria and excitement about all the new possibilities fills the air. Eventually the business reality sets in and with that comes standards and regulations. If the AI vendors are serious about business, they have to follow that path.