SlackBench - MCP - Hallucination: leaderboard

Agents answer questions across realistic Slack workplace reenactments and associated evolving code repositories. Mean category pass score (%). SlackBench140tasks:58AgenticSearch,48Hallucination,14Summarization,20CircumstanceInference; three runs,20min/query; GPT-5.4 judge. Summarization earns checklist partial credit, other categories binary pass; Overall is the unweighted category mean. Harness, MCP/CLI access and reasoning effort remain separate in row identities. Native first-party model APIs.

Metric: Mean category pass score (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 33 models tracked.

Top models

#ModelScoreOverall rank
1Gemini 4 Argon (Benchmark Sweep Run: source=SlackBench; harness=Antigravity; access=MCP; effort=medium)90.28
2Claude Opus 5.5 (Benchmark Sweep Run: source=SlackBench; harness=Claude Code; access=MCP; effort=medium)75
3Claude Fable 5.1 (Benchmark Sweep Run: source=SlackBench; harness=Claude Code; access=MCP; effort=medium)71.53
4GLM-5.3 (Benchmark Sweep Run: source=SlackBench; harness=Claude Code; access=MCP; effort=medium)70.83
5Muse Wonder (Benchmark Sweep Run: source=SlackBench; harness=Claude Code; access=MCP; effort=medium)66.67
6Muse Wonder (Benchmark Sweep Run: source=SlackBench; harness=Muse Code; access=MCP; effort=medium)62.5
7Claude Opus 4.7 (Benchmark Sweep Run: source=SlackBench; harness=Claude Code; access=MCP; effort=medium)60.42
8Grok 4.7 (Benchmark Sweep Run: source=SlackBench; harness=Grok Build; access=MCP; effort=medium)56.94
9GPT-6.1 Terra (Benchmark Sweep Run: source=SlackBench; harness=Codex CLI; access=MCP; effort=medium)55.56
10GPT-6 Astra (Benchmark Sweep Run: source=SlackBench; harness=Codex CLI; access=MCP; effort=medium)54.17

Interactive version: theaggregate.ai/benchmark?slug=slackbench-mcp-hallucination · How It Works · Data refreshed daily, snapshot 2026-11-06.