Tested models
How AI models do on six real author requests in Elabora, how long they take, and what they cost.
We try AI models on a test bench: six real author requests, on a copy of the same course (tighten a sentence, reorder a screen, add a question the plan calls for, translate a screen, add sounds, find a flaw that only shows on screen). Each result is checked automatically, by checks written in advance.
Latest run: 4 November 2026, with the current version of Elabora (after 0.5.1) and OpenCode 2.2.5.
The chat’s free models
These are the models Elabora’s chat offers, with OpenCode, with no account or key. Two runs each, under the chat’s rules.
| Model | Success | Typical time | Cost |
|---|---|---|---|
| DeepSeek V4.1 Flash | 12 of 12 | 1 min 30 | free |
| GLM-5.3 Flash | 11 of 12 | 2 min 20 | free |
- DeepSeek V4.1 Flash: the most reliable. It reads Elabora’s documentation, checks on screen and points out what it notices on the way. Rarely slips in a sentence in English.
- GLM-5.3 Flash: economical and effective; once edited the wrong sentence. Ran on 1 November, before the chat asked for shorter final answers: its line does not quite compare with DeepSeek’s.
For comparison
The same requests, on 1 November 2026, one run each.
| Model | Success | Typical time | Cost per request |
|---|---|---|---|
| Claude Haiku 5.5 | 5 of 6 | 30 s | about 1 US cent |
| Claude Sonnet 5.5 | 6 of 6 | 45 s | about 15 US cents |
| GPT 6 Luna | 5 of 6 | 50 s | included in the ChatGPT subscription |
| GPT 6.1 Terra | 6 of 6 | 40 s | included in the ChatGPT subscription |
- Claude Haiku 5.5: fast and very cheap; did not find how to change a component.
- Claude Sonnet 5.5: the steadiest, with the clearest explanations.
- GPT 6 Luna: fast; edited the wrong sentence.
- GPT 6.1 Terra: fast and reliable.
Claude ran under the rules of Elabora’s chat, through OpenCode, with a pay-as-you-go API key. GPT ran in Codex, with its own terminal: not quite the same test, since Codex can change the files itself, where the chat requires Elabora’s tools.
The limits of these figures
- Few runs. Two per free model, one for the others: a difference of one request proves nothing.
- Free models change without notice. Their provider can update, replace or withdraw them: a result holds at its date.
- An automatic check says whether the request was done, not whether the text is well written. We read the answers, but that judgement is not in the table.
- Times depend on the load of the servers and the computer: they give an order of magnitude.
To choose your AI, see Working with your AI.