MARCA (Portuguese, Orchestrator): leaderboard

MARCA (2026) tests LLM web search on 52 manually authored multi-entity questions asked in Portuguese, in the Orchestrator framework (the model delegates sub-questions to subagents that call web_search and web_scrape). A GPT-4.1 judge checks each answer against a manually validated checklist; checklist accuracy (%) averaged over three runs, higher is better.

Metric: Checklist accuracy (0-1, times 100): share of a question's checklist items (expected entities and attributes) that a GPT-4.1 judge marks satisfied, averaged over MARCA's 52 Portuguese questions and three runs, Orchestrator framework (the model delegates sub-questions to subagents that call web_search and web_scrape); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro88.9#145
2GPT-5.287.2#185
3GPT-4.182.5#505
4Gemini 3 Flash82#175
5GPT-5 Mini81.9#364
6Sabia-480#440
7Kimi K277.3#443
8Sabiazinho-475.3#668
9GPT-4.1 Mini73.5#789
10Qwen 3 235B A22B67.8#628
11Gemini 2.5 Pro59.8#276
12Gemini 2.5 Flash51.7#463
13Qwen 3 30B A3B49.3#1156
14Gemini 2.5 Flash Lite35.4#964

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=marca-portuguese-orchestrator · How It Works · Data refreshed daily, snapshot 2026-11-21.