The internet answers the question "which model is best at law" with rankings. Mostly in English and on American law. We were interested in something narrower: which model writes the best complaint letter under Czech consumer law. We measured that in September and October 2026.
We expected the most expensive model to write the best letter.
We took 11 real Czech consumer complaints from the development part of a set of 200 publicly published complaints filed after 6 January 2023, when the amendment to consumer law (Act No. 374/2022 Coll.) took effect. Each model wrote one letter per case. The fields differed: e-shops and subscriptions, energy, telecommunications, insurance, transport. We removed personal data from the texts in advance and unified the facts.
Each model wrote the letter through the full Legamik, that is, with the same database of laws, the same source materials and the same instructions. Only the model that writes changed. The models were GPT-5.6 Sol, GPT-6 Sol, GPT-6 Astra (OpenAI), Claude Opus 5, Claude Opus 5.5 and Claude Fable 5.1 (Anthropic). Later we also compared GPT-6 Sol with the newer GPT-6.1 Sol.
Three different models, Grok, Claude Fable and GPT-6 Astra, evaluated the letters blind, without being told who wrote them. No model evaluated itself. Each evaluator received the full text of the cited provisions and scored the handling of the law, the facts, the argumentation, the demands and readability on a scale of 1 to 5. Separately, each counted critical legal errors on the basis of which a company could justifiably reject the letter, and invented facts, that is, anything the consumer did not write.
The most expensive model did not win
In the first round, the ranking according to the only evaluator who saw all four models at the time was: GPT-6 Astra 3.64, GPT-5.6 Sol 3.09, Claude Fable 5.1 3.00, Claude Opus 5 2.91.
Claude Fable is the most expensive and the slowest of all the models tested. A letter took it 165 seconds on average. It wrote deep, long arguments, but it more often invented details in them. With the evaluators Grok and Fable, the Claude models had invented facts in 6 to 8 letters out of 11, the GPT models in 0 to 2. The strictest evaluator, Astra, found invented facts in Fable’s letters in 10 out of 11.
Reading pairs of letters side by side did not look that way. In all ten pairs, the longer letter from Claude Fable won. It seemed more convincing. But that reading was done by Claude Fable 5.1 itself, in a separate session, not by a person. It was therefore also evaluating its own letters.
The blind evaluators, who knew the exact wording of the law, saw it differently. On the shorter GPT Sol letters they agreed with each other much more (a difference of 0.18 points between their scores) than on the long Claude Opus letters (1.18 points). We did not separate the effect of length itself from the model, but a longer text leaves more room for a drawn-out citation or an invented circumstance. A company does not read an essay. It looks for whether the claim has support in the facts and in the law, and every statement the consumer did not write is a weak spot.
After tuning, a cheaper model was enough
GPT-6 Sol is the cheaper and faster sibling of the Astra model. After two rounds of tuning the instructions, it reached scores of 3.73, 3.64 and 3.64 with the three evaluators. With both evaluators who could assess both models, it came out better than Astra, and in 33 evaluations it had a single critical error. Yet it cost about a fifth as much, and it wrote a letter in about half the time.
We did not find an invented section
In the main rounds of the measurement, the evaluators did not find a single citation of a section that does not exist. That matches the fact that the model writes from the texts of the law it is given, not from memory. The errors that remain are of a different kind: an existing section applied to a situation it does not apply to. That is a harder problem.
Today, GPT-6.1 Sol writes the letters in Legamik. Compared with GPT-6 Sol, it writes letters richer in content, and what it adds is in the consumer’s favor. We added to its instructions an explicit ban on inventing attachments and circumstances. We dropped the Claude models: they were slower, more expensive, and did not convince on quality. That is why we now build on GPT models. With a new version of any model this can change, and that is why we measure again. Since September we have changed the model that writes the letters three times, each time based on the measurement.
When you write a complaint with a general-purpose chatbot
We cover what to watch for in the article A complaint via ChatGPT: why free may not be enough.
- Do not judge a letter by length and confidence. A letter that sticks only to your facts is stronger than a longer letter with invented details.
- Check whether the cited section really applies to your situation. That a section exists is not enough.
Limits of the measurement
Eleven cases is a small sample. We treat differences below 0.1 points as noise.
Scores and critical errors were determined by language models. I read the letters too, Jaroslav Škrabálek, founder of Legamik. I have a law degree, but I am not a practicing attorney and I do not have that day-to-day practice.
The results hold for specific versions of the models and for Czech consumer law. Models change every few months, and that is why we repeat the measurement.
What to do next
Legamik will prepare a draft complaint letter from your description, using a database of laws and terms and conditions, with the model we select based on measurement. You read the letter, edit it and send it yourself.
Legamik is not a law firm and does not provide legal advice. In complex disputes we recommend consulting an attorney.