Benchmark · agentic

GAIA

13 results 12 models

GAIA (General AI Assistant) tests whether AI assistants can answer real-world, multi-step questions that require using tools, browsing the web, and reasoning across several sources. The score is the percentage of questions answered correctly.

Read more
Example
A typical item is an everyday-sounding question whose answer is buried across several steps — e.g. "look up a fact in a linked document, then combine it with data from a website to reach one short answer" — requiring browsing and tool use rather than a single lookup.
Scoring
The metric is accuracy: the share of questions whose final answer matches the expected answer. Each question has one unambiguous reference answer, so a response is either right or wrong.
Verification
Answers are checked automatically by exact (quasi-exact) string match against the reference answer, and test-set results are submitted to a public leaderboard where the correct answers are kept private.
Why it matters
It measures practical assistant ability — combining reasoning, web browsing, and tools on tasks that are easy for people but hard for AI — making it a common yardstick for autonomous agents.
Worked example
Task
Using the current English Wikipedia, how many studio albums did the band Radiohead release from 1993 through 2007, inclusive? Give a single integer as your final answer.
Solution
Albums in range: Pablo Honey (1993), The Bends (1995), OK Computer (1997), Kid A (2000), Amnesiac (2001), Hail to the Thief (2003), In Rainbows (2007) → 7. Final answer: 7
Walkthrough
GAIA items have one unambiguous ground-truth answer reached by web browsing plus light aggregation — here, counting discography entries within a date window (compilations and live albums excluded). It is graded by quasi-exact match: the normalized final-answer string must exactly equal the reference '7'.
0 23 46 69 92 2024-09-27 2025-08-13 2026-06-30 Trase · 35.5 · 2024-09-27 Trase · 67 · 2025-02-21 Enterprise h2oGPTe Agent · 65 · 2024-12-23 our agent · 55.1 · 2025-02-04 Deep Research · 67.4 · 2025-02-04 OpenAI Deep Research · 72.6 · 2025-02-21 OWL · 69.1 · 2025-03-07 Manus AI · 86.5 · 2025-03-10 h2oGPTe Agent · 75 · 2025-03-17 Alita · 75.2 · 2025-05-26 ALITA-G · 83.0 · 2025-10-27 MiroThinker v1.0 (72B variant) · 81.9 · 2025-11-14 Agents-A1 · 46.4 · 2026-06-30
Trase Enterprise h2oGPTe Agent our agent Deep Research OpenAI Deep Research OWL Manus AI h2oGPTe Agent Alita ALITA-G MiroThinker v1.0 (72B variant) Agents-A1
Timeline
Date Model Score Source
2026-06-30 Agents-A1 46.4pts Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent
2025-11-14 MiroThinker v1.0 (72B variant) 81.9% MiroThinker v1.0 introduces interactive scaling for open-source research agents
2025-10-27 ALITA-G 83.03% ALITA-G self-evolves agents by generating and curating MCP tools
2025-05-26 Alita 75.15% Alita enables scalable agentic reasoning with minimal predefinition and maximal self-evolution
2025-03-17 h2oGPTe Agent 75.0% H2O.ai's h2oGPTe Agent tops GAIA benchmark with 75% accuracy
2025-03-10 Manus AI 86.5% Manus AI launches as autonomous agent with GAIA benchmark results
2025-03-07 OWL 69.09% OWL achieves #1 on GAIA benchmark with 69.09% score
2025-02-21 OpenAI Deep Research 72.57% Trase tops GAIA leaderboard with 67% test score at 1/100th cost
2025-02-21 Trase 67.0% Trase tops GAIA leaderboard with 67% test score at 1/100th cost
2025-02-04 our agent 55.15% Open-source DeepResearch reproduction achieves 55.15% on GAIA using smolagents
2025-02-04 Deep Research 67.36% Open-source DeepResearch reproduction achieves 55.15% on GAIA using smolagents
2024-12-23 Enterprise h2oGPTe Agent 65.0% H2O.ai's h2oGPTe Agent tops GAIA benchmark with 65% score
2024-09-27 Trase 35.55% Trase tops Hugging Face GAIA leaderboard with 35.55% success rate