GAIA-2

Next-generation dynamic, stateful multi-turn environment extension of GAIA.27 Oct 2026

One row per model and compatible score measure. Mean = equal-weight arithmetic average of imported reports. Distinct configurations reported for the same model in a source table remain separate. Reports can overlap; the range is not a confidence interval. Unbundle to inspect each observation.
		Evidence	What the result shows
Claude-4-Sonnet →Anthropic	40.1% 2 reports · range 38.1–42.1%	GAIA-2 · Gaia2 full · Adaptability pass@1[1] View 2 reports 42.1% · GAIA-2 · Gaia2 full · Adaptability pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.3 (source units; see paper for uncertainty definition) [1] 38.1% · GAIA-2 · Gaia2 full · Adaptability pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.2 (source units; see paper for uncertainty definition) [1]	Published observations span 4 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. Leads this displayed comparison.
GPT-5 →OpenAI	35.3% 2 reports · range 30.2–40.4%	GAIA-2 · Gaia2 full · Adaptability pass@1[1] View 2 reports 40.4% · GAIA-2 · Gaia2 full · Adaptability pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.2 (source units; see paper for uncertainty definition) [1] 30.2% · GAIA-2 · Gaia2 full · Adaptability pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.1 (source units; see paper for uncertainty definition) [1]	Published observations span 10.2 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #2 of 11 in this displayed comparison; 4.8 pp below the leader.
Kimi-K2 →Kimi	24% 1 report	GAIA-2 · Gaia2 full · Adaptability pass@1[1]Gaia2 · baseline ReAct · three runs	#3 of 11 in this displayed comparison; 16.1 pp below the leader. Source-reported ± 1.9 (source units; see paper for uncertainty definition) Source notes Source cell: 24.0 ±1.9
GPT-5 (minimal) →OpenAI	19.2% 1 report	GAIA-2 · Gaia2 full · Adaptability pass@1[1]Gaia2 · baseline ReAct · three runs	#4 of 11 in this displayed comparison; 20.9 pp below the leader. Source-reported ± 1.8 (source units; see paper for uncertainty definition) Source notes Source cell: 19.2 ±1.8
Gemini-2.5-Pro →Google	17.5% 1 report	GAIA-2 · Gaia2 full · Adaptability pass@1[1]Gaia2 · baseline ReAct · three runs	#5 of 11 in this displayed comparison; 22.6 pp below the leader. Source-reported ± 1.7 (source units; see paper for uncertainty definition) Source notes Source cell: 17.5 ±1.7
Qwen3-235B →Qwen	12.15% 2 reports · range 8.1–16.2%	GAIA-2 · Gaia2 full · Adaptability pass@1[1] View 2 reports 16.2% · GAIA-2 · Gaia2 full · Adaptability pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.9 (source units; see paper for uncertainty definition) [1] 8.1% · GAIA-2 · Gaia2 full · Adaptability pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 1.2 (source units; see paper for uncertainty definition) [1]	Published observations span 8.1 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #6 of 11 in this displayed comparison; 27.95 pp below the leader.
GPT-OSS 120B →OpenAI	10.6% 1 report	GAIA-2 · Gaia2 full · Adaptability pass@1[1]Gaia2 · baseline ReAct · three runs	#7 of 11 in this displayed comparison; 29.5 pp below the leader. Source-reported ± 1.4 (source units; see paper for uncertainty definition) Source notes Source cell: 10.6 ±1.4
GPT-4o →OpenAI	6.2% 1 report	GAIA-2 · Gaia2 full · Adaptability pass@1[1]Gaia2 · baseline ReAct · three runs	#8 of 11 in this displayed comparison; 33.9 pp below the leader. Source-reported ± 1.1 (source units; see paper for uncertainty definition) Source notes Source cell: 6.2 ±1.1
Llama 4 Maverick →Meta	5% 1 report	GAIA-2 · Gaia2 full · Adaptability pass@1[1]Gaia2 · baseline ReAct · three runs	#9 of 11 in this displayed comparison; 35.1 pp below the leader. Source-reported ± 1.0 (source units; see paper for uncertainty definition) Source notes Source cell: 5.0 ±1.0
Grok-4 →xAI	4.4% 1 report	GAIA-2 · Gaia2 full · Adaptability pass@1[1]Gaia2 · baseline ReAct · three runs	#10 of 11 in this displayed comparison; 35.7 pp below the leader. Source-reported ± 1.6 (source units; see paper for uncertainty definition) Source notes Source cell: 4.4 ±1.6
Llama 3.3 70B Instruct →Meta	1.9% 1 report	GAIA-2 · Gaia2 full · Adaptability pass@1[1]Gaia2 · baseline ReAct · three runs	#11 of 11 in this displayed comparison; 38.2 pp below the leader. Source-reported ± 0.6 (source units; see paper for uncertainty definition) Source notes Source cell: 1.9 ±0.6
Claude-4-Sonnet →Anthropic	30.2% 2 reports · range 27.9–32.5%	GAIA-2 · Gaia2 full · Agent2Agent pass@1[1] View 2 reports 32.5% · GAIA-2 · Gaia2 full · Agent2Agent pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.1 (source units; see paper for uncertainty definition) [1] 27.9% · GAIA-2 · Gaia2 full · Agent2Agent pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.0 (source units; see paper for uncertainty definition) [1]	Published observations span 4.6 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. Leads this displayed comparison.
GPT-5 →OpenAI	21.25% 2 reports · range 17.9–24.6%	GAIA-2 · Gaia2 full · Agent2Agent pass@1[1] View 2 reports 24.6% · GAIA-2 · Gaia2 full · Agent2Agent pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.0 (source units; see paper for uncertainty definition) [1] 17.9% · GAIA-2 · Gaia2 full · Agent2Agent pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 1.8 (source units; see paper for uncertainty definition) [1]	Published observations span 6.7 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #2 of 11 in this displayed comparison; 8.95 pp below the leader.
Gemini-2.5-Pro →Google	20.4% 1 report	GAIA-2 · Gaia2 full · Agent2Agent pass@1[1]Gaia2 · baseline ReAct · three runs	#3 of 11 in this displayed comparison; 9.8 pp below the leader. Source-reported ± 1.8 (source units; see paper for uncertainty definition) Source notes Source cell: 20.4 ±1.8
Kimi-K2 →Kimi	18.3% 1 report	GAIA-2 · Gaia2 full · Agent2Agent pass@1[1]Gaia2 · baseline ReAct · three runs	#4 of 11 in this displayed comparison; 11.9 pp below the leader. Source-reported ± 1.8 (source units; see paper for uncertainty definition) Source notes Source cell: 18.3 ±1.8
Grok-4 →xAI	14.4% 1 report	GAIA-2 · Gaia2 full · Agent2Agent pass@1[1]Gaia2 · baseline ReAct · three runs	#5 of 11 in this displayed comparison; 15.8 pp below the leader. Source-reported ± 2.8 (source units; see paper for uncertainty definition) Source notes Source cell: 14.4 ±2.8
GPT-5 (minimal) →OpenAI	11.5% 1 report	GAIA-2 · Gaia2 full · Agent2Agent pass@1[1]Gaia2 · baseline ReAct · three runs	#6 of 11 in this displayed comparison; 18.7 pp below the leader. Source-reported ± 1.5 (source units; see paper for uncertainty definition) Source notes Source cell: 11.5 ±1.5
Qwen3-235B →Qwen	10.95% 2 reports · range 9.4–12.5%	GAIA-2 · Gaia2 full · Agent2Agent pass@1[1] View 2 reports 12.5% · GAIA-2 · Gaia2 full · Agent2Agent pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.6 (source units; see paper for uncertainty definition) [1] 9.4% · GAIA-2 · Gaia2 full · Agent2Agent pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 1.3 (source units; see paper for uncertainty definition) [1]	Published observations span 3.1 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #7 of 11 in this displayed comparison; 19.25 pp below the leader.
GPT-OSS 120B →OpenAI	10.6% 1 report	GAIA-2 · Gaia2 full · Agent2Agent pass@1[1]Gaia2 · baseline ReAct · three runs	#8 of 11 in this displayed comparison; 19.6 pp below the leader. Source-reported ± 1.4 (source units; see paper for uncertainty definition) Source notes Source cell: 10.6 ±1.4
Llama 4 Maverick →Meta	9.2% 1 report	GAIA-2 · Gaia2 full · Agent2Agent pass@1[1]Gaia2 · baseline ReAct · three runs	#9 of 11 in this displayed comparison; 21 pp below the leader. Source-reported ± 1.3 (source units; see paper for uncertainty definition) Source notes Source cell: 9.2 ±1.3
GPT-4o →OpenAI	5.2% 1 report	GAIA-2 · Gaia2 full · Agent2Agent pass@1[1]Gaia2 · baseline ReAct · three runs	#10 of 11 in this displayed comparison; 25 pp below the leader. Source-reported ± 1.0 (source units; see paper for uncertainty definition) Source notes Source cell: 5.2 ±1.0
Llama 3.3 70B Instruct →Meta	4.6% 1 report	GAIA-2 · Gaia2 full · Agent2Agent pass@1[1]Gaia2 · baseline ReAct · three runs	#11 of 11 in this displayed comparison; 25.6 pp below the leader. Source-reported ± 1.0 (source units; see paper for uncertainty definition) Source notes Source cell: 4.6 ±1.0
GPT-5 →OpenAI	45.75% 2 reports · range 39.6–51.9%	GAIA-2 · Gaia2 full · Ambiguity pass@1[1] View 2 reports 51.9% · GAIA-2 · Gaia2 full · Ambiguity pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.3 (source units; see paper for uncertainty definition) [1] 39.6% · GAIA-2 · Gaia2 full · Ambiguity pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.2 (source units; see paper for uncertainty definition) [1]	Published observations span 12.3 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. Leads this displayed comparison.
Claude-4-Sonnet →Anthropic	25.75% 2 reports · range 24.2–27.3%	GAIA-2 · Gaia2 full · Ambiguity pass@1[1] View 2 reports 27.3% · GAIA-2 · Gaia2 full · Ambiguity pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.0 (source units; see paper for uncertainty definition) [1] 24.2% · GAIA-2 · Gaia2 full · Ambiguity pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.0 (source units; see paper for uncertainty definition) [1]	Published observations span 3.1 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #2 of 11 in this displayed comparison; 20 pp below the leader.
GPT-5 (minimal) →OpenAI	20.6% 1 report	GAIA-2 · Gaia2 full · Ambiguity pass@1[1]Gaia2 · baseline ReAct · three runs	#3 of 11 in this displayed comparison; 25.15 pp below the leader. Source-reported ± 1.8 (source units; see paper for uncertainty definition) Source notes Source cell: 20.6 ±1.8
Gemini-2.5-Pro →Google	18.1% 1 report	GAIA-2 · Gaia2 full · Ambiguity pass@1[1]Gaia2 · baseline ReAct · three runs	#4 of 11 in this displayed comparison; 27.65 pp below the leader. Source-reported ± 1.8 (source units; see paper for uncertainty definition) Source notes Source cell: 18.1 ±1.8
Grok-4 →xAI	9.4% 1 report	GAIA-2 · Gaia2 full · Ambiguity pass@1[1]Gaia2 · baseline ReAct · three runs	#5 of 11 in this displayed comparison; 36.35 pp below the leader. Source-reported ± 2.3 (source units; see paper for uncertainty definition) Source notes Source cell: 9.4 ±2.3
GPT-OSS 120B →OpenAI	8.3% 1 report	GAIA-2 · Gaia2 full · Ambiguity pass@1[1]Gaia2 · baseline ReAct · three runs	#6 of 11 in this displayed comparison; 37.45 pp below the leader. Source-reported ± 1.3 (source units; see paper for uncertainty definition) Source notes Source cell: 8.3 ±1.3
Kimi-K2 →Kimi	8.3% 1 report	GAIA-2 · Gaia2 full · Ambiguity pass@1[1]Gaia2 · baseline ReAct · three runs	#6 of 11 in this displayed comparison; 37.45 pp below the leader. Source-reported ± 1.3 (source units; see paper for uncertainty definition) Source notes Source cell: 8.3 ±1.3
Qwen3-235B →Qwen	8.25% 2 reports · range 6.5–10%	GAIA-2 · Gaia2 full · Ambiguity pass@1[1] View 2 reports 10% · GAIA-2 · Gaia2 full · Ambiguity pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.4 (source units; see paper for uncertainty definition) [1] 6.5% · GAIA-2 · Gaia2 full · Ambiguity pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 1.1 (source units; see paper for uncertainty definition) [1]	Published observations span 3.5 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #8 of 11 in this displayed comparison; 37.5 pp below the leader.
GPT-4o →OpenAI	4.4% 1 report	GAIA-2 · Gaia2 full · Ambiguity pass@1[1]Gaia2 · baseline ReAct · three runs	#9 of 11 in this displayed comparison; 41.35 pp below the leader. Source-reported ± 0.9 (source units; see paper for uncertainty definition) Source notes Source cell: 4.4 ±0.9
Llama 4 Maverick →Meta	2.1% 1 report	GAIA-2 · Gaia2 full · Ambiguity pass@1[1]Gaia2 · baseline ReAct · three runs	#10 of 11 in this displayed comparison; 43.65 pp below the leader. Source-reported ± 0.7 (source units; see paper for uncertainty definition) Source notes Source cell: 2.1 ±0.7
Llama 3.3 70B Instruct →Meta	1.7% 1 report	GAIA-2 · Gaia2 full · Ambiguity pass@1[1]Gaia2 · baseline ReAct · three runs	#11 of 11 in this displayed comparison; 44.05 pp below the leader. Source-reported ± 0.6 (source units; see paper for uncertainty definition) Source notes Source cell: 1.7 ±0.6
GPT-5 →OpenAI	60.95% 2 reports · range 52.7–69.2%	GAIA-2 · Gaia2 full · Execution pass@1[1] View 2 reports 69.2% · GAIA-2 · Gaia2 full · Execution pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.1 (source units; see paper for uncertainty definition) [1] 52.7% · GAIA-2 · Gaia2 full · Execution pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.3 (source units; see paper for uncertainty definition) [1]	Published observations span 16.5 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. Leads this displayed comparison.
Claude-4-Sonnet →Anthropic	60% 2 reports · range 57.9–62.1%	GAIA-2 · Gaia2 full · Execution pass@1[1] View 2 reports 62.1% · GAIA-2 · Gaia2 full · Execution pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.2 (source units; see paper for uncertainty definition) [1] 57.9% · GAIA-2 · Gaia2 full · Execution pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.3 (source units; see paper for uncertainty definition) [1]	Published observations span 4.2 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #2 of 11 in this displayed comparison; 0.95 pp below the leader.
Gemini-2.5-Pro →Google	39.2% 1 report	GAIA-2 · Gaia2 full · Execution pass@1[1]Gaia2 · baseline ReAct · three runs	#3 of 11 in this displayed comparison; 21.75 pp below the leader. Source-reported ± 2.2 (source units; see paper for uncertainty definition) Source notes Source cell: 39.2 ±2.2
Kimi-K2 →Kimi	34.2% 1 report	GAIA-2 · Gaia2 full · Execution pass@1[1]Gaia2 · baseline ReAct · three runs	#4 of 11 in this displayed comparison; 26.75 pp below the leader. Source-reported ± 2.2 (source units; see paper for uncertainty definition) Source notes Source cell: 34.2 ±2.2
GPT-5 (minimal) →OpenAI	31.9% 1 report	GAIA-2 · Gaia2 full · Execution pass@1[1]Gaia2 · baseline ReAct · three runs	#5 of 11 in this displayed comparison; 29.05 pp below the leader. Source-reported ± 2.1 (source units; see paper for uncertainty definition) Source notes Source cell: 31.9 ±2.1
Qwen3-235B →Qwen	25.4% 2 reports · range 22.7–28.1%	GAIA-2 · Gaia2 full · Execution pass@1[1] View 2 reports 28.1% · GAIA-2 · Gaia2 full · Execution pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.1 (source units; see paper for uncertainty definition) [1] 22.7% · GAIA-2 · Gaia2 full · Execution pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 1.9 (source units; see paper for uncertainty definition) [1]	Published observations span 5.4 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #6 of 11 in this displayed comparison; 35.55 pp below the leader.
GPT-OSS 120B →OpenAI	17.9% 1 report	GAIA-2 · Gaia2 full · Execution pass@1[1]Gaia2 · baseline ReAct · three runs	#7 of 11 in this displayed comparison; 43.05 pp below the leader. Source-reported ± 1.8 (source units; see paper for uncertainty definition) Source notes Source cell: 17.9 ±1.8
Llama 4 Maverick →Meta	13.8% 1 report	GAIA-2 · Gaia2 full · Execution pass@1[1]Gaia2 · baseline ReAct · three runs	#8 of 11 in this displayed comparison; 47.15 pp below the leader. Source-reported ± 1.6 (source units; see paper for uncertainty definition) Source notes Source cell: 13.8 ±1.6
Grok-4 →xAI	8.8% 1 report	GAIA-2 · Gaia2 full · Execution pass@1[1]Gaia2 · baseline ReAct · three runs	#9 of 11 in this displayed comparison; 52.15 pp below the leader. Source-reported ± 2.2 (source units; see paper for uncertainty definition) Source notes Source cell: 8.8 ±2.2
GPT-4o →OpenAI	8.3% 1 report	GAIA-2 · Gaia2 full · Execution pass@1[1]Gaia2 · baseline ReAct · three runs	#10 of 11 in this displayed comparison; 52.65 pp below the leader. Source-reported ± 1.3 (source units; see paper for uncertainty definition) Source notes Source cell: 8.3 ±1.3
Llama 3.3 70B Instruct →Meta	7.1% 1 report	GAIA-2 · Gaia2 full · Execution pass@1[1]Gaia2 · baseline ReAct · three runs	#11 of 11 in this displayed comparison; 53.85 pp below the leader. Source-reported ± 1.2 (source units; see paper for uncertainty definition) Source notes Source cell: 7.1 ±1.2
GPT-5 →OpenAI	31.85% 2 reports · range 28.3–35.4%	GAIA-2 · Gaia2 full · Noise pass@1[1] View 2 reports 35.4% · GAIA-2 · Gaia2 full · Noise pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.2 (source units; see paper for uncertainty definition) [1] 28.3% · GAIA-2 · Gaia2 full · Noise pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.1 (source units; see paper for uncertainty definition) [1]	Published observations span 7.1 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. Leads this displayed comparison.
Claude-4-Sonnet →Anthropic	29.45% 2 reports · range 27.7–31.2%	GAIA-2 · Gaia2 full · Noise pass@1[1] View 2 reports 31.2% · GAIA-2 · Gaia2 full · Noise pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.1 (source units; see paper for uncertainty definition) [1] 27.7% · GAIA-2 · Gaia2 full · Noise pass@1 · baseline ReAct · three runs · Gaia2 · baseline ReAct · three runs · Source-reported ± 2.0 (source units; see paper for uncertainty definition) [1]	Published observations span 3.5 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #2 of 11 in this displayed comparison; 2.4 pp below the leader.
Gemini-2.5-Pro →Google	20.4% 1 report	GAIA-2 · Gaia2 full · Noise pass@1[1]Gaia2 · baseline ReAct · three runs	#3 of 11 in this displayed comparison; 11.45 pp below the leader. Source-reported ± 1.8 (source units; see paper for uncertainty definition) Source notes Source cell: 20.4 ±1.8
Kimi-K2 →Kimi	18.8% 1 report	GAIA-2 · Gaia2 full · Noise pass@1[1]Gaia2 · baseline ReAct · three runs	#4 of 11 in this displayed comparison; 13.05 pp below the leader. Source-reported ± 1.8 (source units; see paper for uncertainty definition) Source notes Source cell: 18.8 ±1.8
Grok-4 →xAI	15.6% 1 report	GAIA-2 · Gaia2 full · Noise pass@1[1]Gaia2 · baseline ReAct · three runs	#5 of 11 in this displayed comparison; 16.25 pp below the leader. Source-reported ± 2.9 (source units; see paper for uncertainty definition) Source notes Source cell: 15.6 ±2.9
GPT-OSS 120B →OpenAI	14.6% 1 report	GAIA-2 · Gaia2 full · Noise pass@1[1]Gaia2 · baseline ReAct · three runs	#6 of 11 in this displayed comparison; 17.25 pp below the leader. Source-reported ± 1.6 (source units; see paper for uncertainty definition) Source notes Source cell: 14.6 ±1.6

Sources

Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments — published result table ↗
Froger, Romain, Andrews, Pierre, Bettini, Matteo, Budhiraja, Amar, Cabral, Ricardo Silveira, Do, Virginie et al. · 2026
All 24 authors
Froger, Romain, Andrews, Pierre, Bettini, Matteo, Budhiraja, Amar, Cabral, Ricardo Silveira, Do, Virginie, Garreau, Emilien, Gaya, Jean-Baptiste, Laurençon, Hugo, Lecanu, Maxime, Malkan, Kunal, Mekala, Dheeraj, Ménard, Pierre, Bertran, Gerard Moreno-Torres, Piterbarg, Ulyana, Plekhanov, Mikhail, Rita, Mathieu, Rusakov, Andrey, Vorotilov, Vladislav, Wang, Mengjue, Yu, Ian, Benhalloum, Amine, Mialon, Grégoire, Scialom, Thomas
Dates & evaluation details
GAIA-2 · Gaia2 full · Execution pass@1 · baseline ReAct · three runs · 2026-10-27
Published model comparison, 2026. Gaia2 · baseline ReAct · three runs. Pass@1 on the full Gaia2 scenario set using common ReAct scaffolding. Overall is the paper’s equal-weight average of seven capability splits; Gaia2-mini and alternative agent scaffolds are distinct evaluations. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.
GAIA-2 · Gaia2 full · Search pass@1 · baseline ReAct · three runs · 2026-10-27
Published model comparison, 2026. Gaia2 · baseline ReAct · three runs. Pass@1 on the full Gaia2 scenario set using common ReAct scaffolding. Overall is the paper’s equal-weight average of seven capability splits; Gaia2-mini and alternative agent scaffolds are distinct evaluations. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.
GAIA-2 · Gaia2 full · Ambiguity pass@1 · baseline ReAct · three runs · 2026-10-27
Published model comparison, 2026. Gaia2 · baseline ReAct · three runs. Pass@1 on the full Gaia2 scenario set using common ReAct scaffolding. Overall is the paper’s equal-weight average of seven capability splits; Gaia2-mini and alternative agent scaffolds are distinct evaluations. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.
GAIA-2 · Gaia2 full · Adaptability pass@1 · baseline ReAct · three runs · 2026-10-27
Published model comparison, 2026. Gaia2 · baseline ReAct · three runs. Pass@1 on the full Gaia2 scenario set using common ReAct scaffolding. Overall is the paper’s equal-weight average of seven capability splits; Gaia2-mini and alternative agent scaffolds are distinct evaluations. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.
GAIA-2 · Gaia2 full · Time pass@1 · baseline ReAct · three runs · 2026-10-27
Published model comparison, 2026. Gaia2 · baseline ReAct · three runs. Pass@1 on the full Gaia2 scenario set using common ReAct scaffolding. Overall is the paper’s equal-weight average of seven capability splits; Gaia2-mini and alternative agent scaffolds are distinct evaluations. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.
GAIA-2 · Gaia2 full · Noise pass@1 · baseline ReAct · three runs · 2026-10-27
Published model comparison, 2026. Gaia2 · baseline ReAct · three runs. Pass@1 on the full Gaia2 scenario set using common ReAct scaffolding. Overall is the paper’s equal-weight average of seven capability splits; Gaia2-mini and alternative agent scaffolds are distinct evaluations. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.
GAIA-2 · Gaia2 full · Agent2Agent pass@1 · baseline ReAct · three runs · 2026-10-27
Published model comparison, 2026. Gaia2 · baseline ReAct · three runs. Pass@1 on the full Gaia2 scenario set using common ReAct scaffolding. Overall is the paper’s equal-weight average of seven capability splits; Gaia2-mini and alternative agent scaffolds are distinct evaluations. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.
GAIA-2 · Gaia2 full · Overall pass@1 · baseline ReAct · three runs · 2026-10-27
Published model comparison, 2026. Gaia2 · baseline ReAct · three runs. Pass@1 on the full Gaia2 scenario set using common ReAct scaffolding. Overall is the paper’s equal-weight average of seven capability splits; Gaia2-mini and alternative agent scaffolds are distinct evaluations. Table 2; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

TRACE THE SCORE · Per-model notes use published breakdowns where available.

What it measures

Next-generation dynamic, stateful multi-turn environment extension of GAIA.

Reading the measurement

The useful outcome is a verified change or deliverable. Inspect the tool interface, initial state, user simulator and policy constraints. An agent can claim completion without reaching the required state, or reach it through an action the protocol forbids.

Published evaluation

Pass@1 on the full Gaia2 scenario set using common ReAct scaffolding. Overall is the paper’s equal-weight average of seven capability splits; Gaia2-mini and alternative agent scaffolds are distinct evaluations.

Where this fits in an evaluation suite

A transcript that says “done” is weaker evidence than the verified final state. Track policy violations, recoveries, and invalid attempts alongside success.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.

PRIMARY EVIDENCE & FURTHER READING

Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments — published result table ↗
Froger, Romain, Andrews, Pierre, Bettini, Matteo, Budhiraja, Amar, Cabral, Ricardo Silveira, Do, Virginie et al. · 2026
All 24 authors
Froger, Romain, Andrews, Pierre, Bettini, Matteo, Budhiraja, Amar, Cabral, Ricardo Silveira, Do, Virginie, Garreau, Emilien, Gaya, Jean-Baptiste, Laurençon, Hugo, Lecanu, Maxime, Malkan, Kunal, Mekala, Dheeraj, Ménard, Pierre, Bertran, Gerard Moreno-Torres, Piterbarg, Ulyana, Plekhanov, Mikhail, Rita, Mathieu, Rusakov, Andrey, Vorotilov, Vladislav, Wang, Mengjue, Yu, Ian, Benhalloum, Amine, Mialon, Grégoire, Scialom, Thomas

Reviewed 2026-10-27. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.

Continue exploring

All agents & tool use benchmarks →

Sources

What it measures

Reading the measurement

Published evaluation

Where this fits in an evaluation suite

A reproducible comparison

Continue exploring

AA-Briefcase v1.1

GDPval-AA v2.1

APEX-Agents-AA

AA-AnalystAgent