European B2B software investment benchmark

Investment judgement,benchmarked.

Not “can the model answer?” but “can the model help decide?”

View leaderboard Inspect scored case

Observed 18 July 20269 requested8 ranked

Methodology v0.4-mini

Observed standings from a synthetic diligence case.

8 ranked / 9 requested / 2 samples

#1Kimi K3Moonshot AI96.3Range 95.6 to 96.9Provisional tieComplete

Dimension scoresMView model record

Grounding85.4

Commercial judgement98.7

Scepticism97.7

Numerical sanity100

Risk discovery96.6

Question generation96.1

European context100

Output usefulness99.7

Closed-bookn=2/2Route caveat

#2Grok 4.5xAI96.3Range 95.5 to 97.0Provisional tieComplete

Dimension scoresView model record

Grounding89.4

Commercial judgement99.1

Scepticism100

Numerical sanity100

Risk discovery99.6

Question generation96.1

European context70

Output usefulness99.7

Closed-bookn=2/2

#3GPT-5.6 SolOpenAI95.5Range 95.1 to 95.9Provisional tieComplete

Dimension scoresOAIView model record

Grounding87.3

Commercial judgement96.9

Scepticism96.5

Numerical sanity100

Risk discovery94.2

Question generation94.5

European context100

Output usefulness99.6

Closed-bookn=2/2

#4Gemini 3.5 FlashGoogle89.1Range 86.9 to 91.3Provisional tieComplete

Dimension scoresView model record

Grounding86.7

Commercial judgement81.1

Scepticism88.8

Numerical sanity100

Risk discovery91.7

Question generation92.1

European context96.7

Output usefulness74.6

Closed-bookn=2/2

#5Claude Opus 4.8Anthropic88.8Range 80.2 to 97.4Provisional tieComplete

Dimension scoresView model record

Grounding91.3

Commercial judgement77.7

Scepticism96.5

Numerical sanity100

Risk discovery94.2

Question generation66.5

European context100

Output usefulness85.7

Closed-bookn=2/2

#6DeepSeek V4 ProDeepSeek88.6Range 87.9 to 89.4Provisional tieComplete

Dimension scoresView model record

Grounding82

Commercial judgement90.7

Scepticism87.2

Numerical sanity100

Risk discovery81.1

Question generation91.3

European context85

Output usefulness91.5

Closed-bookn=2/2

#7GLM 5.2Z.ai88.2Range 86.6 to 89.7Provisional tieComplete

Dimension scoresZView model record

Grounding85.9

Commercial judgement95.9

Scepticism88.4

Numerical sanity90.7

Risk discovery83.2

Question generation87.3

European context70

Output usefulness90.4

Closed-bookn=2/2

#8Qwen3.7 MaxAlibaba65.5Range 44.6 to 86.4Complete

Dimension scoresView model record

Grounding66.5

Commercial judgement78.4

Scepticism60.8

Numerical sanity57.5

Risk discovery45.9

Question generation80.4

European context70

Output usefulness73.5

Closed-bookn=2/2

	Model		Range									Samples	Mode
1	MKimi K3Moonshot AIProvisional tieRoute caveat	96.3	95.6 to 96.9	85.4	98.7	97.7	100.0	96.6	96.1	100.0	99.7	2/2	Closed-book
2	Grok 4.5xAIProvisional tie	96.3	95.5 to 97.0	89.4	99.1	100.0	100.0	99.6	96.1	70.0	99.7	2/2	Closed-book
3	OAIGPT-5.6 SolOpenAIProvisional tie	95.5	95.1 to 95.9	87.3	96.9	96.5	100.0	94.2	94.5	100.0	99.6	2/2	Closed-book
4	Gemini 3.5 FlashGoogleProvisional tie	89.1	86.9 to 91.3	86.7	81.1	88.8	100.0	91.7	92.1	96.7	74.6	2/2	Closed-book
5	Claude Opus 4.8AnthropicProvisional tie	88.8	80.2 to 97.4	91.3	77.7	96.5	100.0	94.2	66.5	100.0	85.7	2/2	Closed-book
6	DeepSeek V4 ProDeepSeekProvisional tie	88.6	87.9 to 89.4	82.0	90.7	87.2	100.0	81.1	91.3	85.0	91.5	2/2	Closed-book
7	ZGLM 5.2Z.aiProvisional tie	88.2	86.6 to 89.7	85.9	95.9	88.4	90.7	83.2	87.3	70.0	90.4	2/2	Closed-book
8	Qwen3.7 MaxAlibaba	65.5	44.6 to 86.4	66.5	78.4	60.8	57.5	45.9	80.4	70.0	73.5	2/2	Closed-book

1 requested model was excluded from ranking because a complete two-sample result was not available. Inspect the full roster

Scoring shape

Judgement carries the most weight.

A grounded but commercially naive model should not lead an investment benchmark. Overall is derived from eight dimension scores, never stored directly.

Grounding

15%

Claims are supported by packet evidence or clearly marked as inference.

Commercial judgement

20%

Identifies what matters economically and strategically to the deal.

Scepticism

15%

Challenges weak claims, inflated metrics and convenient narratives.

Numerical sanity

15%

Interprets and reconciles financial and operating data correctly.

Risk discovery

15%

Finds hidden and second-order risks across the packet.

Question generation

10%

Asks the next questions most likely to change the recommendation.

European context

Accounts for fragmented markets, regulation, language and procurement.

Output usefulness

Produces material a deal team can reuse with limited editing.

Interpretation

Workflow usefulness, not general intelligence.

This board reports observed OpenRouter outputs and evaluator scores for one fresh, synthetic company. It measures decision usefulness on this workflow, not general model intelligence.

Interrogate the method