Multi-domain agent tasks
The Vals Index: agentic tasks in finance, coding, legal and tax, weighted by each sector’s share of U.S. GDP.
Highest reported score
- Vals Index: Claude Sonnet 5.5 · 70.85%
Public AI benchmarks, with sources
Which AI is best for your work?
Compare published benchmark scores against cost, then put your chosen model to work.
Explore the benchmarks
Vals Index
Agentic performance on finance, coding, legal and tax tasks, weighted by each sector’s share of U.S. GDP.
Loading chart…
Each square is one model, marked with its provider’s initial. Higher is better; further left costs less per test.
Highest reported score: Claude Sonnet 5.5 (Anthropic) at 70.85%.
Accuracy, highest first. Position is the source’s published rank.
- 1Claude Sonnet 5.5 Anthropic70.85%
- 2Gemini 4 Carbon Google70.45%
- 3Claude Opus 5.5 Anthropic70.18%
- 4Gemini 4 Argon Google69.24%
- 5Wonder Full Meta66.79%
- 6GPT-6 Astra OpenAI65.02%
- 7Wonder Meta64.88%
- 8GPT-6.1 Sol OpenAI63.94%
- 9Kimi K3.1 Pro Moonshot62.52%
- 10DeepSeek V4.1 Flash 1017 DeepSeek61.26%
- 11GPT-6.1 Terra OpenAI60.54%
- 12GLM 5.3 Z.ai59.73%
- 13Hy4 Other59.35%
- 14Qwen 3.8 Max 0902 Alibaba59.05%
- 15Claude Haiku 5.5 Anthropic58.73%
- 16Grok 4.7 xAI58.56%
- 17MiMo V2.6 Pro Xiaomi57.81%
- 18Mistral Large 4 Mistral56.53%
- 19Inkling 2 Small Other56.08%
- 20Gemini 3.8 Flash Google55.95%
- 21MiMo V2.6 Flash Xiaomi54.83%
- 22Step 5 Preview Other54.76%
- 23MiniMax-M3.1-Flash MiniMax53.78%
- 24Hy4 Preview Other53.21%
- 25DeepSeek V4 Pro 0813 DeepSeek52.97%
| 1 | Claude Sonnet 5.5 | Anthropic | 70.85% | $16 |
|---|---|---|---|---|
| 2 | Gemini 4 Carbon | 70.45% | $11 | |
| 3 | Claude Opus 5.5 | Anthropic | 70.18% | $25 |
| 4 | Gemini 4 Argon | 69.24% | $13 | |
| 5 | Wonder Full | Meta | 66.79% | $6.81 |
| 6 | GPT-6 Astra | OpenAI | 65.02% | $21 |
| 7 | Wonder | Meta | 64.88% | $11 |
| 8 | GPT-6.1 Sol | OpenAI | 63.94% | $3.61 |
| 9 | Kimi K3.1 Pro | Moonshot | 62.52% | $6.44 |
| 10 | DeepSeek V4.1 Flash 1017 | DeepSeek | 61.26% | $0.66 |
| 11 | GPT-6.1 Terra | OpenAI | 60.54% | $2.94 |
| 12 | GLM 5.3 | Z.ai | 59.73% | $8.52 |
| 13 | Hy4 | Other | 59.35% | $1.72 |
| 14 | Qwen 3.8 Max 0902 | Alibaba | 59.05% | $8.73 |
| 15 | Claude Haiku 5.5 | Anthropic | 58.73% | $3.17 |
| 16 | Grok 4.7 | xAI | 58.56% | $12 |
| 17 | MiMo V2.6 Pro | Xiaomi | 57.81% | $0.43 |
| 18 | Mistral Large 4 | Mistral | 56.53% | $3.72 |
| 19 | Inkling 2 Small | Other | 56.08% | $3.51 |
| 20 | Gemini 3.8 Flash | 55.95% | $5.03 | |
| 21 | MiMo V2.6 Flash | Xiaomi | 54.83% | $0.22 |
| 22 | Step 5 Preview | Other | 54.76% | $2.73 |
| 23 | MiniMax-M3.1-Flash | MiniMax | 53.78% | $0.66 |
| 24 | Hy4 Preview | Other | 53.21% | $1.57 |
| 25 | DeepSeek V4 Pro 0813 | DeepSeek | 52.97% | $2.50 |
Source: Vals AI · leaderboard dated 5 Nov 2026 · checked 8 Nov 2026
Open the full Vals Index leaderboard
Prompts to try with these models
All prompts →ReviewsClaudeCritique AI outputs and tag failuresEvaluating two answers from an AI agent to improve it.View prompt →ResearchClaudeCompetitor and market scanMapping who else plays in your space and who could be a partner or acquirer.View prompt →ResearchClaudeExplain a concept, applied to my businessLearning something new and immediately seeing what it means for you.View prompt →
Results from published AI leaderboards. Compare models within each benchmark; scores from different tests aren’t directly comparable. How we collect these results
Every field and leaderboard
Pick a field, then a benchmark. Where the source publishes a cost per test, you can see score against cost; otherwise you get the ranking.
Legal
Drafting, extraction, legal reasoning and US legal research.
Highest reported score
- Contract Workflows: Joint leaders: Claude Opus 5.5, Claude Fable 5.1 and Muse Spark 1.3 · 48.4%
- Data Extraction: Claude Opus 5.5 · 76.7%
- LegalBench: Claude Fable 5 · 88.56%
- Legal Research Bench: Gemini 4 Carbon · 55.77%
Health & medical
Physician-graded health conversations, medical coding and exam questions.
Highest reported score
- HealthBench (independent run): Claude Sonnet 5.5 (max) · 62.1%
- MedCode: Claude Opus 5 · 63.57%
- MedQA: o1 · 96.52%
Finance
Analyst research on SEC filings and long credit agreements.
Highest reported score
- Finance Agent: Claude Opus 5.5 · 67.85%
- CorpFin v2: Claude Opus 5 · 73.19%
Coding
Long-horizon engineering tasks, real GitHub issues and fresh programming problems.
Highest reported score
- DeepSWE: GPT-6 Astra (xhigh) · 74.12%
- SWE-bench Verified: Claude Opus 5 · 97.0%
- LiveCodeBench: Claude Fable 5.1 · 90.52%
Agents & tool use
Multi-step tasks an agent completes alone in a terminal.
Highest reported score
- Terminal-Bench 4.0: Claude Opus 5.5 · 65.15%
- Terminal-Bench 2.1: Claude Opus 5.5 · 87.64%
Reasoning & knowledge
Expert exams, graduate science, contest maths and broad knowledge.
Highest reported score
- Humanity's Last Exam: GPT 6 Astra · 54.80%
- GPQA Diamond: Gemini 3.1 Pro Preview (02/26) · 95.45%
- AIME: Gemini 3.1 Pro Preview (02/26) · 98.13%
- MMLU Pro: Claude Fable 5.1 · 92.38%
Chat preference
Which answers real people prefer in blind side-by-side votes.
Highest reported score
- LMArena Text Arena: Gemini 4 Carbon (High) · 1527
Multilingual
Reasoning in languages other than English.
Highest reported score
- MGSM: Claude Opus 4.5 (Thinking) · 95.20%
Put your chosen AI to work
Found a model worth testing? Start with copy-ready prompts for decks, emails and landing pages, for ChatGPT, Claude, Gemini, Grok Bot and more.
EmailsChatGPTMeeting follow-up emailRight after a customer or partner meeting, when you have your notes and a clear next step.View prompt →SlidesChatGPTCreate a customer proposal deckYou are pitching an AI platform to a specific customer and have reference material to hand.View prompt →WebsitesChatGPTWrite a landing pageLaunching an offer and starting from a blank page.View prompt →SlidesChatGPTImprove a single slideA slide is crowded or the message is muddy.View prompt →SlidesChatGPTMake a deck suitable for senior managementA working deck needs to be reshaped for the CEO, business leaders and CTO.View prompt →SlidesChatGPTMake a deck more visualYour deck is text-heavy and you want the customer journey to be easy to picture.View prompt →
Private team libraries
Keep the prompts that work in one place
Save the prompts your team relies on in a private, searchable library, and invite colleagues to copy and reuse them. Free for up to 5 members and 50 prompts per team.
How we collect these results
- We don’t run these tests. Every score is copied as published by the leaderboard that ran it, with a link to the source.
- Leaderboard dated is the update date the source shows. When a source doesn’t show one, we say “Leaderboard date not shown”.
- Last checked is when we last read the source. Scores only change here when we refresh from it.
- Highest reported score is the top entry on the source’s leaderboard. Field pages show up to 29 entries per leaderboard, with a link to the full list.
- Compare within a benchmark. Each leaderboard uses its own tasks, settings and scoring, so a score on one test isn’t directly comparable to a score on another.
- No longer updated means the source says it has stopped running that benchmark on new model releases, so recent models may be missing.
- Model names appear as each source lists them. Provider badges are coloured initials, not logos.
- Spotted a mistake? Tell us.