Which model?
Tested in October 2026. Models change often: this page gives the results of that test, and the settings’ Recommended, from our tests follows the latest one.
In October 2026, we graded twelve models out of ten, like at school: eleven tests, seven of them on real photos of a year-8 pupil’s notebook (pages taken askew, phone shadows, crossed-out words), the others without a photo (formulas, geometry figures, spelling, helps on the back). Two AI judges compared the cards with a hand-checked reading of each page; every model ran twice. Costs are in US cents per lesson.
| Model | Grade | On photos | Per lesson | Time | In short |
|---|---|---|---|---|---|
| Gemini Flash | 8.5 | 8.7 | 1.3 ¢ | 18 s | Our advice: nearly as good as the best for a fraction of the price, and the most consistent from one run to the next. The default with Gemini and OpenRouter. |
| Claude Sonnet | 8.4 | 8.8 | 3.2 ¢ | 14 s | Among the best on photos, and the fastest. The default with Claude. |
| GPT Sol | 8.4 | 8.5 | 2.1 ¢ | 27 s | A good deal on the OpenAI side. The default with GPT. |
| Claude Opus | 8.5 | 8.8 | 6.4 ¢ | 18 s | As good as Sonnet on photos, twice the price. |
| GPT Astra | 8.7 | 8.8 | 10.3 ¢ | 28 s | The best average, but eight times the price of Gemini Flash. |
| Gemini Pro | 8.0 | 8.3 | 5.3 ¢ | 27 s | Behind Gemini Flash, for four times the price. |
| GPT Luna | 7.8 | 7.7 | 0.1 ¢ | 23 s | Almost free, but misreads handwriting on a page taken askew. |
To avoid:
- Kimi, DeepSeek Flash and Grok: slow (Grok takes 2 to 3 minutes a lesson, Kimi ran out of time three times), or poor on pages taken askew;
- small models such as Claude Haiku, GPT Mini, Mistral Medium or GLM Flash: on a hard page, they make up a lesson instead of reading it.
Whichever the model, check the cards in the review before sending them: even the best make a few mistakes, especially on hard-to-read handwriting. The helps on the back (explanations and ways to remember) were the weakest test for every model.
These are the models under Recommended, from our tests in the settings (see
The model), in their exact
tested versions (for instance gemini-3.8-flash, claude-sonnet-5-5, gpt-6.1-sol; on
OpenRouter google/gemini-3.8-flash…). Notosaurus doesn’t recommend the ~…-latest aliases:
a new version can read differently or cost more, so it is tested before it replaces one.