Tabloy on the GAIA benchmark
GAIA is a public test for AI assistants, made by researchers from Meta, Hugging Face and AutoGPT. Its questions come from real tasks, and each one has a single short answer. Getting to it means searching the web and reading files, often over many steps. People answer about 92% of the questions correctly.
On 7 October 2026 we ran Tabloy on all 165 questions of the GAIA validation set. Tabloy answered 128 of them correctly, a score of 77.6%.
The result
| Level | Questions | Correct | Score |
|---|---|---|---|
| Level 1 | 53 | 43 | 81.1% |
| Level 2 | 86 | 68 | 79.1% |
| Level 3 | 26 | 17 | 65.4% |
| All | 165 | 128 | 77.6% |
The levels go from shorter questions to long tasks. A Level 3 question can take dozens of steps before the answer is found.
Next to other agents
| Agent | Level 3 | All levels |
|---|---|---|
| Tabloy, October 2026 | 65.4% | 77.6% |
| Genspark Super Agent, April 2025 | 58.8% | 75.4% |
| Manus, March 2025 | 57.7% | 73.4% |
| OpenAI Deep Research, February 2025 | 47.6% | 67.4% |
Tabloy has the highest score of the four, and the gap is widest on Level 3, the hardest questions. On Level 1 Manus published a higher score than ours, 86.5% against 81.1%.
The other figures are the results each company published for the same validation set. They date from early 2025 and may be higher today.
How we ran it
- We used the app at app.tabloy.ai with a regular Pro account and Tabloy's most capable model setting.
- Each question was typed into the message box the way a person would type it. When a question comes with a file, the file was attached.
- Every question ended with the instruction the GAIA authors give their own baselines: finish with "FINAL ANSWER:" and keep the answer short.
- Each question got one attempt. Nothing was retried, and no best run was picked out of several.
- The answers were scored with the scoring code of the official GAIA leaderboard.
Some misses are about format. The instruction asks for answers without articles, while a few expected answers keep "The". The official scorer counts those as wrong, and so do we.
What this result is not
This is our own run, like the figures the other companies published. It is not an entry on the official GAIA leaderboard, which scores a separate set of 300 questions with hidden answers.
The GAIA terms ask that the questions are not reposted, so this page shows the scores only.
Sources
- GAIA: a benchmark for General AI Assistants, the paper by the GAIA authors
- GAIA leaderboard on Hugging Face
- Introducing deep research, OpenAI
- Genspark and Manus GAIA scores, a summary of the published results