# ClaimBench v1 (Vals AI) (ClaimBench)

> Vals AI finance benchmark that asks whether a 10-K excerpt supports a claim, contradicts it, or says nothing about it. It compares small decision models with frontier language models.

Canonical page: https://benchlm.ai/benchmarks/claimbench

- Category: [Decision Models](/decision-models)
- Last updated: October 23, 2026

Evaluation results published by [Vals AI](https://www.vals.ai/benchmarks/claimbench). Published leaderboard aggregates and evaluation settings only. Dataset contents are not included.

## About ClaimBench

- Year: 2026
- Tasks: 1,960 claims
- Format: Three-label classification accuracy
- Difficulty: Claim verification against SEC 10-K excerpts
- Paper: [ClaimBench leaderboard and methodology](https://www.vals.ai/benchmarks/claimbench)

Vals released v1 on October 20, 2026. It has 1,960 claims, each paired with a 10-K excerpt and five construction categories of about 392 claims: direct support, subtle contradiction, numeric discrepancy, missing evidence and non-supporting citation. Gold labels come from the construction rule, not from a model. We mirror overall and per-category accuracy as a display-only board, outside overall and category rankings.

ClaimBench is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (25 models)

| Rank | Model | Configuration | Creator | Score |
|------|-------|---------------|---------|-------|
| 1 | [d1-ultra](/models/d1-ultra) | — | Liquid AI | 100.00% |
| 2 | [Claude Sonnet 5.5](/models/claude-sonnet-5-5) | — | Anthropic | 100.00% |
| 3 | [Claude Haiku 5.5](/models/claude-haiku-5-5) | — | Anthropic | 100.00% |
| 4 | [Claude Opus 5.5](/models/claude-opus-5-5) | — | Anthropic | 100.00% |
| 5 | [GPT-6.1 Sol](/models/gpt-6-1-sol) | max reasoning | OpenAI | 100.00% |
| 6 | [Perplexity Decider v1.2 27B](/models/pplx-decider-v1-2) | — | Perplexity | 99.69% |
| 7 | [Drex Max](https://www.vals.ai/models/decision_drex-max) | — | Nace.AI via OpenRouter | 99.18% |
| 8 | [GPT-6 Luna](/models/gpt-6-luna) | max reasoning | OpenAI | 99.18% |
| 9 | [Mercury Decide](/models/mercury-decide) | — | Inception | 99.08% |
| 10 | [Perplexity Decider v1.1 (27B)](/models/decision-index-0-3-perplexity-decider-v1-1-27b) | — | Perplexity | 98.78% |
| 11 | [d1](/models/liquid-d1) | — | Liquid AI | 98.67% |
| 12 | [Microsoft-Decision-1](/models/microsoft-decision-1) | — | Microsoft | 98.21% |
| 13 | [Jev 1.14](/models/jev-1-14) | — | TypeSafe AI | 98.06% |
| 14 | [Drex v1.5](/models/drex-v1-5) | — | Nace.AI | 98.01% |
| 15 | [Jev 1.13.0](/models/jev-1-13-0) | — | TypeSafe AI | 97.65% |
| 16 | [Drex v2.0](/models/drex-v2-0) | — | Nace.AI | 97.30% |
| 17 | [GPT-6 Luna Decisions](https://www.vals.ai/models/decision_gpt-6-luna) | — | OpenAI | 96.89% |
| 18 | [Perplexity Decider Flash v1.2 4B](/models/pplx-decider-flash-v1-2) | — | Perplexity | 96.53% |
| 19 | [Clef](/models/cloudflare-clef) | — | Cloudflare | 95.10% |
| 20 | [Solar Decide](https://www.vals.ai/models/decision_solar-decide) | — | Upstage via OpenRouter | 94.39% |
| 21 | [Tev1-4B-experimental](/models/decision-index-0-3-tev1-4b-experimental) | — | togethercomputer | 89.80% |
| 22 | [Clef-Omni](/models/cloudflare-clef-omni) | — | Cloudflare | 87.70% |
| 23 | [ScaleDown Classify](/models/scaledown-classify) | — | ScaleDown | 85.00% |
| 24 | [Clef-Flash](/models/cloudflare-clef-flash) | — | Cloudflare | 84.85% |
| 25 | [Solar Decide Flash](https://www.vals.ai/models/decision_solar-decide-flash) | — | Upstage via OpenRouter | 82.45% |

## FAQ

### What does ClaimBench measure?

Vals AI finance benchmark that asks whether a 10-K excerpt supports a claim, contradicts it, or says nothing about it. It compares small decision models with frontier language models.

### Which model leads the published ClaimBench snapshot?

d1-ultra currently leads the published ClaimBench snapshot with a score of 100.00%.

### How many models are evaluated on ClaimBench?

The October 23, 2026 contains 25 AI models.

### Does ClaimBench affect BenchLM's overall score?

Not directly. ClaimBench is still displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.