LLM Ranking 2026

The report is based on tests of language models conducted via official chatbots (with the exception of EuroLLM). The aim was to assess how individual models – available through these interfaces – handle open-ended tasks. Both the testing approach and the selection of tasks stem from a simple assumption: this is how a typical user engages with AI. For this reason, the evaluations were carried out by people – 11 in total, including myself, representing a range of professions. When calculating the averages, a mechanism known from ski jumping was used, whereby one extreme score on either end is discarded. To make the process easier, participants were given assessment criteria and model answers for strictly factual questions. The models themselves vary in capability.

A few of the most important conclusions for me:

1. The biggest surprise was the performance of the Llama 4 and Qwen 3.5 Plus models. This suggests that the gap between open models and their commercial counterparts (such as ChatGPT) in everyday use may not be that large. Credit should also be given to Google for the development of Gemini, which ultimately took first place (in the initial pilot in 2024 it did not reach the podium).
2. Media attention focused on the low ranking of Polish models. If we look at the results without any assumptions or caveats, it should be noted that European models generally performed worse than their American and Chinese counterparts. At the same time, in the report I explain my perspective on the results, which bears no resemblance to headlines such as “Polish bots are idiots”, and I strongly oppose such use of my report. A model should be chosen to suit the task it is meant to perform and I emphasised this point strongly in the report.
3. The one criterion on which all models failed was Polish humour. I put forward the idea that this may be the ultimate test of whether AI truly matches human ability.
4. Overall, tasks related to professional work – which are the most important to me given my role as a trainer and consultant – were handled better by the models than tasks in the broadly defined socio-cultural category. The tests included, for example, writing emails and generating presentation content.
5. What I hope for is a discussion on the direction of AI development in Poland and Europe, necessarily from a practical and business perspective. In the report and in my media statements, I stress the need to create a regulatory framework that ensures efficient funding, openness to risk and operational flexibility for startups. As regards LLMs, if we want to achieve a position comparable to top-tier models (so-called SOTA), we need to make swift and well-considered decisions. Moving in this direction, we should be more open to the needs of the typical user and the quality of the content we request on a daily basis. Ultimately, these criteria determine who we are willing to pay our hypothetical $20 per month to. To my mind, this direction is reflected in headlines about sovereign AI – without a European SOTA model, there will be no sovereign AI.

For the record, I should add that on 8 February I asked representatives of Bielik for an interview and received no response; on 24 February, the responses for evaluation were generated and the report was published on 9 March.

I encourage you to read the full report and the accompanying articles carefully – links can be found below.

Summary publication date: 19 March 2026.

PS As I already wrote in the report, I am strongly rooting for European models, including, naturally, Polish ones above all.

Language models should be selected according to the tasks we need to carry out and the regulatory conditions we operate under.

The report responds to questions I am often asked during training sessions, including those specifically concerning language models developed in Poland and Europe and the extent to which they are universal.

The fact that something is not universal does not mean it is bad.