Numbers you can check and cite.
Three sets of figures from thinkhard.si, each with the date it was measured and a line to cite it by. Anyone writing or teaching about AI at work is welcome to use them.
19models from Anthropic, OpenAI, Google, Mistral, DeepSeek, xAI and Groq
AI model API prices, per million tokens
The list prices for input, output and cached input tokens and each model's context length, checked against each provider's own pricing page. They are checked again on the first day of every month, and an announced price change takes over on its date.
Cite asAI model API list prices, thinkhard.si, checked November 5, 2026. https://thinkhard.si/prices/
1.2out of 10: the average score of 319 work tasks written as the one line most people type
What a one-line prompt scores
We took the 319 work tasks the repository held at the time, wrote each one as the single line most people type, and scored it on the Prompt Grader's ten checks. 261 scored 1 out of 10, and none scored above 3. Two short additions lifted the average from 1.2 to 6.8: one saying who the answer is for, how long and what shape, and one saying what to do when a fact is missing.
Cite asWhat a one-line prompt scores, thinkhard.si, October 24, 2026. https://thinkhard.si/guides/what-a-one-line-prompt-scores/
96%of 100 hand-labeled work prompts scored within one point by the Prompt Grader's free checks
How accurate the site's own tools are
A tool that checks your work should say how often it is right. Each figure below comes from a labeled test set kept off the site, so the rules are never tuned to the test, and each carries the date it was measured.
- Prompt Grader. On 100 work prompts, each labeled by hand from the checks' definitions alone, the score was exactly right for 67% and within one point for 96%. The mean F1 across the checks that read wording was 0.93. Measured November 11, 2026.
- Prompt search. On 172 labeled searches across all 1,024 prompts, the right prompt came in the top three for 92% of the searches the rules were never tuned on, and 97% of the ones they were. It came first 78% of the time. Measured November 11, 2026.
- The prompts themselves. Every prompt in the repository scores 9 or 10 on the same ten checks before it is published. Today 1,017 score 10, and the 7 that score 9 are prompts for image and video models. How every prompt is made and tested.
- Safe Paste. Not measured yet. We are building a labeled set of real-looking documents and will publish how many names, numbers and addresses it finds and misses, by kind.
Cite asAccuracy of the thinkhard.si Prompt Grader and prompt search, thinkhard.si, November 11, 2026. https://thinkhard.si/research/#accuracy
For writers and journalists
thinkhard.si is a set of free tools for getting better answers from AI at work. It went live on October 20, 2026. It has 21 free tools and a repository of 1,024 tested prompt templates, and most of the tools run in the reader's own browser, so what people type stays on their device (the trust center says tool by tool what is sent where). It is independent, with no advertising and no trackers from other companies.
If you are writing about AI at work and want a figure we have not published, or want to check one before you quote it, write to contact@thinkhard.si. The mark is at thinkhard.si/assets/img/icon-512.png, and the about page says what it stands for; please use it unchanged.
