Vals Finance Agent v2.1 (Finance Agent v2.1)
We show this table for reference; we do not rank on it.
Financial research with tools, revised questions and rubrics, a new judge jury, and stricter treatment of hedged answers.
Evaluation results published by Vals AI. Published leaderboard aggregates and evaluation settings only. Dataset contents are not included.
Partial-credit accuracy on Finance Agent v2.1 — October 27, 2026
We mirror the published partial-credit accuracy view for Finance Agent v2.1. Claude Opus 5.5 leads the public snapshot at 67.85%, followed by Gemini 4 Argon (67.66%) and Muse Wonder Full (67.58%). We do not use these results to rank models overall.
Claude Opus 5.5
Anthropic
Gemini 4 Argon
high reasoning
Muse Wonder Full
Meta
max reasoning
33 modelsAgenticCurrentDisplay onlyUpdated October 27, 2026
Partial-credit accuracy table (33 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Claude Opus 5.5Anthropic | 67.85% | Not reported | Closed |
| 2 | Gemini 4 ArgonGooglehigh reasoning | 67.66% | Not reported | Closed |
| 3 | Muse Wonder FullMetamax reasoning | 67.58% | Not reported | Closed |
| 4 | Claude Sonnet 5.5Anthropic | 67.28% | Not reported | Closed |
| 5 | Gemini 4 CarbonGooglehigh reasoning | 67.16% | Not reported | Closed |
| 6 | Gemini 3.8 FlashGooglehigh reasoning | 65.76% | Not reported | Closed |
| 7 | Muse WonderMetamax reasoning | 65.55% | Not reported | Closed |
| 8 | DeepSeek V4.1 Flash 1017DeepSeekmax reasoning | 65.43% | Not reported | Open |
| 9 | Claude Haiku 5.5Anthropic | 65.37% | Not reported | Closed |
| 10 | GLM-5.3Z.AImax reasoning | 65.00% | Not reported | Open |
| 11 | MiMo-V2.6-ProXiaomi | 64.26% | Not reported | Open |
| 12 | MiMo-V2.6-FlashXiaomi | 63.32% | Not reported | Open |
| 13 | DeepSeek V4.1 Flash 1017DeepSeekmax reasoning | 62.93% | Not reported | Open |
| 14 | GPT-6.1 SolOpenAImax reasoning | 61.82% | Not reported | Closed |
| 15 | Hy4 previewTencent | 61.62% | Not reported | Open |
| 16 | Kimi K3.1 ProMoonshot AImax reasoning | 61.49% | Not reported | Pending |
| 17 | Mistral Large 4Mistralhigh reasoning | 61.38% | Not reported | Open |
| 18 | Muse Spark OpenMetaxhigh reasoning | 60.66% | Not reported | Open |
| 19 | GLM-5.3-FlashZ.AImax reasoning | 60.59% | Not reported | Open |
| 20 | Grok 4.7xAIxhigh reasoning | 60.59% | Not reported | Closed |
| 21 | GPT-6 AstraOpenAImax reasoning | 60.43% | Not reported | Closed |
| 22 | Kimi K3Moonshot AImax reasoning | 60.01% | Not reported | Open |
| 23 | Step 5 PreviewStepFun | 58.90% | Not reported | Open |
| 24 | Qwen3.8 Max InternalAlibaba | 58.83% | Not reported | Unknown |
| 25 | MiniMax M3.1 FlashMiniMax | 58.83% | Not reported | Closed |
| 26 | DeepSeek V4 Pro 0813DeepSeekmax reasoning | 57.12% | Not reported | Open |
| 27 | GPT-6 LunaOpenAImax reasoning | 56.84% | Not reported | Closed |
| 28 | GPT-6.1 TerraOpenAImax reasoning | 55.96% | Not reported | Closed |
| 29 | Mercury 3Inceptionhigh reasoning | 54.71% | Not reported | Closed |
| 30 | MiniMax M3MiniMax | 54.34% | Not reported | Open |
| 31 | Beam 1.0Reflection | 51.83% | Not reported | Unknown |
| 32 | InklingThinking Machines Lab0.99 reasoning | 49.87% | Not reported | Open |
| 33 | Inkling-SmallThinking Machines Lab0.99 reasoning | 44.78% | Not reported | Open |
How Finance Agent v2.1 is shown here
Finance Agent v2.1 evaluates financial research with six tools and a two-hour task limit. Its primary score gives partial credit for weighted checks only when every required dealbreaker passes. All-pass requires every check to pass and remains a separate task view.
Version 2.1 revises questions and rubrics, replaces the judge jury, and fails answers that offer multiple candidate values without choosing one. Vals says these scores are not comparable with v2. We keep the editions separate and display only.
Snapshot
The published Finance Agent v2.1 snapshot places Claude Opus 5.5 first at 67.85%. The third row is 0.27 points behind. The broader top-10 range is 2.84 points, so many of the published results sit in a relatively narrow band.
33 models have been evaluated on Finance Agent v2.1. The benchmark falls in the Agentic category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. Finance Agent v2.1 is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About Finance Agent v2.1
Year
2026
Tasks
449 private test questions across nine analyst categories
Format
Dealbreaker-gated partial credit; all-pass reported separately
Difficulty
Professional expert-task agent workflow
The primary metric awards severity-weighted partial credit after all dealbreakers pass. All-pass requires every check. Version 2.1 revises the jury, data, and hedging rules; Vals says scores are not comparable with v2. We retain both editions as separate display-only tables.
Freshness and provenance
Version
Finance Agent v2.1 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does Finance Agent v2.1 measure?
Financial research with tools, revised questions and rubrics, a new judge jury, and stricter treatment of hedged answers.
Which model leads the published Finance Agent v2.1 snapshot?
Claude Opus 5.5 currently leads the published Finance Agent v2.1 snapshot with 67.85% partial-credit accuracy. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on Finance Agent v2.1?
The October 27, 2026 snapshot contains 33 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.