ApprenticeBench - API (MCP Tools): leaderboard
NeoCognition's ApprenticeBench: an agent joins a simulated construction company as an accounts-payable apprentice and processes 100 vendor bills in order in an Odoo ERP with only the handbook, six months of history and mentor feedback. This track replaces the GUI with MCP tools over Odoo's backend API at feature parity, inside a Codex or Claude Code harness; score is the cumulative share of bills passed, higher is better.
Metric: Cumulative Success Rate (%). Source: neocognition.io. Saturation forecast: Around August 2027. 27 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5.1 (Max) | 70 | #10 (Claude Fable 5.1) |
| 2 | GPT-6 Astra (Max) | 65 | #6 (GPT-6 Astra) |
| 3 | Claude Opus 5 (Max) | 49 | #16 (Claude Opus 5) |
| 4 | Claude Fable 5 (Max) | 45 | #20 (Claude Fable 5) |
| 5 | Gemini 3.8 Flash | 34 | #30 |
| 6 | Gemini 3.7 Flash | 32 | #24 |
| 7 | GPT-5.6 Sol (Max) | 30 | #22 (GPT-5.6 Sol) |
| 8 | Claude Opus 4.8 (Max) | 28 | #39 (Claude Opus 4.8) |
| 9 | Claude Sonnet 5 (Max) | 26 | #65 (Claude Sonnet 5) |
| 10 | Kimi K3 (Max) | 25 | #35 (Kimi K3) |
| 11 | Muse Spark 1.3 (Max) | 24 | #31 (Muse Spark 1.3) |
| 12 | Qwen 3.8 Max (Max) | 24 | #34 (Qwen 3.8 Max) |
| 13 | DeepSeek V4 Pro (Max) | 21 | #112 (DeepSeek V4 Pro) |
| 14 | GLM-5.3 Flash (Max) | 21 | #77 (GLM-5.3 Flash) |
| 15 | GPT-5.6 Terra (Max) | 21 | #40 (GPT-5.6 Terra) |
No result here: #3 Claude Opus 5.5, #7 GPT-6.1 Sol, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=apprenticebench-api-mcp-tools · How It Works · Data refreshed daily, snapshot 2026-11-05.