SPARKITMeasured June 2026
GAIA
SPARKIT answered 73.2% of the 127 evaluated GAIA questions correctly, compared with 58.2% for Exa and 57.0% for Brave.
127 questions · Accuracy · Updated June 24, 2026
Results
| System | Evaluation mode | Accuracy |
|---|---|---|
| SPARKITSPARKIT research-agent run | SPARKIT research-agent run | 73.2% |
| ExaSearch API result | Search API result | 58.2% |
| BraveSearch API result | Search API result | 57.0% |
Methodology summary
- 01The evaluation uses 127 questions from the GAIA validation benchmark.
- 02Each system is scored against the benchmark answer key, and the reported value is the percentage of evaluated questions answered correctly.
- 03The comparison tests an end-to-end research-agent workflow against search API results on questions where retrieval and reasoning both matter.
Sources and supporting artifacts
GAIA benchmark paper ↗The primary publication describing the public benchmark.GAIA retrieval case-study artifact (JSON) →A representative Level 2 question with system outputs, timing, and grading. This is not the complete benchmark dataset.Hallucinated vs. fetched case study →A readable walkthrough of the retrieval failure exposed by that question.
Limitations
- 01GAIA spans broad real-world tasks and does not isolate scientific literature performance.
- 02Search indexes and provider behavior change over time, so these results are a dated snapshot rather than a permanent ranking.
- 03The downloadable case study documents one representative question, not the complete 127-question run-level dataset.
- 04Benchmark accuracy does not guarantee that every citation, calculation, or conclusion in a production report is correct.