EMB
Whether an agent can build a complete Excel financial model — tabs, structure, formulas and formatting — from a written business scenario and a set of source files, the work junior investment-banking and private-equity analysts do.
How it is run
Seven model categories (LBO, DCF, M&A, operating models, comps, dataroom summaries and misc) with tasks, gold-standard models and rubrics authored and peer-reviewed by financial experts. Each task runs in Template mode, filling in a provided skeleton and graded by exact cell match, and in Scratch mode, built from nothing and graded by a partial-credit rubric over numerical, formula and presentation checks.
What it cannot tell you
Partial credit flatters the field: the leader at 76.67% still leaves a substantial-but-incomplete model rather than a client-ready deliverable. Numerical checks are the bottleneck — the leading run passes 89% of formula checks and only 64% of numerical ones — and LBO and DCF tasks, where early errors cascade, are consistently the hardest.
- Leader
- Claude Fable 5.176.67% · Vals AI's own list
- Top-ten spread
- 5.14%Leader minus tenth place, or last place in smaller fields.
- Median
- 57.85%of the 69 rows published
- No published result
- 9of 46 models on this site
Independent leaderboard · Vals AI
Top 69 of 69 runs by Vals AI, one harness.
| # | Model | Score |
|---|---|---|
| 1 | Claude Fable 5.1 | 76.67% |
| 2 | Claude Opus 5.5 | 75.94% |
| 3 | Claude Sonnet 5.5 | 75.71% |
| 4 | Gemini 4 Argon | 75.24% |
| 5 | Claude Fable 5 | 73.67% |
| 6 | Claude Opus 5 | 73.56% |
| 7 | GPT-5.6 Sol | 72.34% |
| 8 | Gemini 3.8 Flash | 72.20% |
| 9 | GPT-6 Astra | 71.70% |
| 10 | GPT-6 Sol | 71.53% |
| 11 | Gemini 3.7 Flash | 71.33% |
| 12 | GPT-6.1 Sol | 70.77% |
| 13 | Claude Opus 4.8 | 69.37% |
| 14 | GPT-6 Luna | 68.52% |
| 15 | Muse Spark 1.3 Max | 67.43% |
| 16 | GPT-5.6 Luna | 67.12% |
| 17 | Grok 4.7 | 66.99% |
| 18 | Kimi K3 | 66.40% |
| 19 | Claude Sonnet 5 | 66.32% |
| 20 | GPT-5.6 Terra | 66.20% |
| 21 | MiMo-V2.6-Flash | 65.46% |
| 22 | Gemini 3.6 Flash | 65.41% |
| 23 | GPT-5.5 | 64.54% |
| 24 | Claude Opus 4.7 | 63.73% |
| 25 | Gemini 3.5 Flash | 63.55% |
| 26 | Ember 1 | 63.37% |
| 27 | MiMo V2.6 Pro | 62.86% |
| 28 | Grok 4.6 | 62.73% |
| 29 | Muse Spark 1 3 | 62.71% |
| 30 | Glm 5.2 | 61.53% |
| 31 | Claude Sonnet 4 6 | 60.15% |
| 32 | Qwen3.8 Max | 60.07% |
| 33 | Qwen3.8 27b | 59.66% |
| 34 | Hunyuan Hy4 Preview | 58.08% |
| 35 | Kimi K2.6 | 57.85% |
| 36 | DeepSeek V4.1 Flash | 57.21% |
| 37 | Qwen3.7 Max | 57.04% |
| 38 | Muse Spark 1.2 | 56.98% |
| 38 | Deepseek V4 Flash 0731 | 56.98% |
| 40 | Muse Spark 1 1 | 56.36% |
| 41 | GLM 5.3 | 56.34% |
| 42 | GLM 5.3 Flash | 55.93% |
| 43 | Mimo V2.5 Pro | 55.23% |
| 44 | Mimo V2.5 | 55.09% |
| 45 | Grok 4.5 | 52.91% |
| 46 | DeepSeek V4 Pro | 52.80% |
| 47 | Gemini 3.1 Pro | 52.62% |
| 48 | Deepseek V4 Pro | 51.62% |
| 49 | Qwen3.7 Plus | 49.34% |
| 50 | MiniMax M3 | 47.76% |
| 51 | Gpt 5.4 Mini 2026 03 17 | 45.43% |
| 52 | Gpt 5.4 Nano 2026 03 17 | 44.75% |
| 53 | Gemini 3.5 Flash Lite | 43.23% |
| 54 | Inkling | 38.36% |
| 55 | Ling 3.0 Flash Af Rc3 | 36.83% |
| 56 | Qwen3.6 Plus | 32.87% |
| 57 | Inkling Small | 31.76% |
| 58 | Nemotron 3 Ultra 550b A55b | 31.73% |
| 59 | Kimi K2.5 Thinking | 28.47% |
| 60 | MiniMax M2.7 | 28.45% |
| 61 | Gemini 3 Flash Preview | 26.02% |
| 62 | Ling 3.0 Flash 2607 | 25.01% |
| 63 | Claude Haiku 4 5 20251001 Thinking | 22.66% |
| 64 | Grok 4.3 | 18.39% |
| 65 | Nemotron Lightning 3p5 30b A3b | 16.99% |
| 66 | Grok 4.20 0309 Reasoning | 11.74% |
| 67 | Mistral Medium 3.5 | 11.19% |
| 68 | Mercury 2.5 | 8.79% |
| 69 | Gemini 3.1 Flash Lite Preview | 8.63% |
No published result tracked
9 of the 46 models on this site have no published number on this benchmark, so their cells in the home table stay blank rather than reading as zero.
Sources
Every page used for this benchmark, read on 17 October 2026.
- Benchmark site or the page that publishes itwww.vals.ai/benchmarks/emb
- Independent resultswww.vals.ai/benchmarks/emb