# Vals AI benchmark leaderboards

> Current top models on every Vals AI benchmark, regenerated with each results export. Each benchmark page has the full leaderboard, cost, latency, and methodology.

Site index: https://www.vals.ai/llms.txt

## Vals Index

- Page: https://www.vals.ai/benchmarks/vals_index
- Category: Index
- Updated: 2026-10-18
- Models evaluated: 50

A single measure of AI's potential economic impact — agentic model performance across finance, coding, legal, and tax tasks, weighted by each sector's share of U.S. GDP.

Top 10 by Accuracy:

1. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 68.90%
2. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 67.04%
3. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 66.97%
4. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 65.83%
5. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 64.57%
6. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 63.67%
7. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 63.13%
8. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 61.39%
9. [GPT-6.1 Sol](https://www.vals.ai/models/openai_gpt-6.1-sol): 61.15%
10. [Wonder](https://www.vals.ai/models/meta_muse-wonder): 61.06%

## Vals Multimodal Index

- Page: https://www.vals.ai/benchmarks/vals_multimodal_index
- Category: Index (archived: no longer run on new models)
- Updated: 2026-08-11
- Models evaluated: 33

Benchmark consisting of a weighted performance across finance, coding, and education tasks. Showing the potential impact that LLMs can have on the economy.

Top 10 by Accuracy:

1. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 74.15%
2. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 73.90%
3. [Kimi K3](https://www.vals.ai/models/kimi_kimi-k3): 73.42%
4. [GPT-5.6 Sol](https://www.vals.ai/models/openai_gpt-5.6-sol): 72.64%
5. [Claude Opus 4.8](https://www.vals.ai/models/anthropic_claude-opus-4-8): 70.89%
6. [Muse Spark 1.2](https://www.vals.ai/models/meta_muse_spark_1_2): 69.80%
7. [GPT-5.6 Luna](https://www.vals.ai/models/openai_gpt-5.6-luna): 69.06%
8. [Claude Sonnet 5](https://www.vals.ai/models/anthropic_claude-sonnet-5): 68.83%
9. [GPT 5.5](https://www.vals.ai/models/openai_gpt-5.5): 68.07%
10. [Claude Opus 4.7](https://www.vals.ai/models/anthropic_claude-opus-4-7): 67.36%

## Vals RSI Index

- Page: https://www.vals.ai/benchmarks/rsi_index
- Category: Index
- Updated: 2026-10-15
- Models evaluated: 25

Can a model do the research that builds the next model? Autonomous LLM R&D scored against published, human, and frontier-model references.

Top 10 by Index score:

1. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 0.373
2. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 0.361
3. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 0.330
4. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 0.305
5. [GPT-6 Sol](https://www.vals.ai/models/openai_gpt-6-sol): 0.281
6. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 0.271
7. [GPT-6.1 Sol](https://www.vals.ai/models/openai_gpt-6.1-sol): 0.266
8. [Grok 4.6](https://www.vals.ai/models/grok_grok-4.6): 0.255
9. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 0.244
10. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 0.243

## Vals Science Index

- Page: https://www.vals.ai/benchmarks/science_index
- Category: Index
- Updated: 2026-10-16
- Models evaluated: 21

A single measure of AI's ability to do science — agentic model performance on biology data investigation, mechanism discovery, and research workflows, weighted equally.

Top 10 by Accuracy:

1. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 65.09%
2. [GPT-6.1 Sol](https://www.vals.ai/models/openai_gpt-6.1-sol): 59.63%
3. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 59.12%
4. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 58.65%
5. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 55.36%
6. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 48.54%
7. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 47.93%
8. [GPT-6 Sol](https://www.vals.ai/models/openai_gpt-6-sol): 45.00%
9. [GPT-5.6 Sol](https://www.vals.ai/models/openai_gpt-5.6-sol): 41.48%
10. [Wonder](https://www.vals.ai/models/meta_muse-wonder): 39.12%

## Vals Web Search Index

- Page: https://www.vals.ai/benchmarks/web_search
- Category: Index
- Updated: 2026-07-16
- Models evaluated: 4 (8 model and search tool pairs)

Comparing native provider search against independent web-search tools on legal-research and finance-analysis tasks

Top 8 by Score:

1. [Claude Fable 5 (Exa)](https://www.vals.ai/models/anthropic_claude-fable-5): 48.45%
2. [Claude Fable 5 (Native)](https://www.vals.ai/models/anthropic_claude-fable-5): 46.94%
3. [GPT-5.6 Sol (Exa)](https://www.vals.ai/models/openai_gpt-5.6-sol): 45.24%
4. [GPT-5.6 Sol (Native)](https://www.vals.ai/models/openai_gpt-5.6-sol): 43.58%
5. [Gemini 3.5 Flash (Exa)](https://www.vals.ai/models/google_gemini-3.5-flash): 41.36%
6. [Grok 4.5 (Exa)](https://www.vals.ai/models/grok_grok-4.5): 38.75%
7. [Gemini 3.5 Flash (Native)](https://www.vals.ai/models/google_gemini-3.5-flash): 37.04%
8. [Grok 4.5 (Native)](https://www.vals.ai/models/grok_grok-4.5): 35.54%

## CaseLaw v2

- Page: https://www.vals.ai/benchmarks/case_law_v2
- Category: Legal (archived: no longer run on new models)
- Updated: 2026-05-04
- Models evaluated: 54

Private question-answer benchmark over Canadian court-cases.

Top 10 by Accuracy:

1. [Grok 4.3](https://www.vals.ai/models/grok_grok-4.3): 79.31%
2. [GPT 5.1](https://www.vals.ai/models/openai_gpt-5.1-2025-11-13): 73.42%
3. [GPT 4.1](https://www.vals.ai/models/openai_gpt-4.1-2025-04-14): 69.88%
4. [GPT 5 Mini](https://www.vals.ai/models/openai_gpt-5-mini-2025-08-07): 68.49%
5. [Claude Opus 4.7](https://www.vals.ai/models/anthropic_claude-opus-4-7): 68.38%
6. [GPT 5](https://www.vals.ai/models/openai_gpt-5-2025-08-07): 66.45%
7. [GPT 5.5](https://www.vals.ai/models/openai_gpt-5.5): 66.24%
8. [GPT 5.2](https://www.vals.ai/models/openai_gpt-5.2-2025-12-11): 66.02%
9. [Grok 4](https://www.vals.ai/models/grok_grok-4-0709): 65.81%
10. [Grok 4 Fast (Reasoning)](https://www.vals.ai/models/grok_grok-4-fast-reasoning): 65.70%

## Harvey's Legal Agent Benchmark

- Page: https://www.vals.ai/benchmarks/hlab
- Category: Legal
- Updated: 2026-10-18
- Models evaluated: 81

Tests an agent's ability to complete legal work using documents, spreadsheets, presentations, and file-system tools.

Top 10 by Task Pass Rate:

1. [Muse Spark 1.2](https://www.vals.ai/models/meta_muse_spark_1_2): 25.42%
2. [Muse Spark 1.3 Max](https://www.vals.ai/models/meta_muse_spark_1_3_max): 23.75%
3. [Muse Spark 1.3](https://www.vals.ai/models/meta_muse_spark_1_3): 22.92%
4. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 21.67%
5. [Muse Spark 1.1](https://www.vals.ai/models/meta_muse-spark-1.1): 20.00%
6. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 19.58%
7. [GLM Legal](https://www.vals.ai/models/zai_glm-legal): 19.17%
8. [Wonder](https://www.vals.ai/models/meta_muse-wonder): 17.08%
9. [Muse Spark Open](https://www.vals.ai/models/meta_muse-spark-open): 16.67%
10. [Mistral Large 4](https://www.vals.ai/models/mistralai_mistral-large-4): 15.83%

## Legal Research Bench

- Page: https://www.vals.ai/benchmarks/legal_research
- Category: Legal
- Updated: 2026-10-18
- Models evaluated: 80

Evaluating agents on legal research tasks across diverse areas of US law

Top 10 by All-Pass:

1. [Muse Spark 1.3 Max](https://www.vals.ai/models/meta_muse_spark_1_3_max): 55.29%
2. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 55.29%
3. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 55.29%
4. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 54.81%
5. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 53.37%
6. [Wonder](https://www.vals.ai/models/meta_muse-wonder): 50.48%
7. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 50.48%
8. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 49.52%
9. [GLM 5.3](https://www.vals.ai/models/zai_glm-5.3): 49.04%
10. [Grok 4.6](https://www.vals.ai/models/grok_grok-4.6): 48.08%

## LegalBench

- Page: https://www.vals.ai/benchmarks/legal_bench
- Category: Legal
- Updated: 2026-10-13
- Models evaluated: 151

Evaluating language models on a wide range of open source legal reasoning tasks.

Top 10 by Accuracy:

1. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 88.56%
2. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 88.51%
3. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 88.30%
4. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 87.42%
5. [Gemini 3.1 Pro Preview (02/26)](https://www.vals.ai/models/google_gemini-3.1-pro-preview): 87.40%
6. [Gemini 3.7 Flash](https://www.vals.ai/models/google_gemini-3.7-flash): 87.26%
7. [Gemini 3 Pro (11/25)](https://www.vals.ai/models/google_gemini-3-pro-preview): 87.03%
8. [Gemini 3.8 Flash](https://www.vals.ai/models/google_gemini-3.8-flash): 86.99%
9. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 86.97%
10. [GPT-5.6 Sol](https://www.vals.ai/models/openai_gpt-5.6-sol): 86.97%

## CorpFin v2

- Page: https://www.vals.ai/benchmarks/corp_fin_v2
- Category: Finance (archived: no longer run on new models)
- Updated: 2026-08-12
- Models evaluated: 134

A private benchmark evaluating understanding of long-context credit agreements

Top 10 by Accuracy:

1. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 73.19%
2. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 71.83%
3. [Kimi K3](https://www.vals.ai/models/kimi_kimi-k3): 71.56%
4. [Muse Spark 1.1](https://www.vals.ai/models/meta_muse-spark-1.1): 71.29%
5. [Muse Spark 1.2](https://www.vals.ai/models/meta_muse_spark_1_2): 70.94%
6. [Inkling Small](https://www.vals.ai/models/thinkingmachines_inkling-small): 69.62%
7. [Inkling](https://www.vals.ai/models/thinkingmachines_inkling): 68.57%
8. [Grok 4.3](https://www.vals.ai/models/grok_grok-4.3): 68.53%
9. [GPT 5.5](https://www.vals.ai/models/openai_gpt-5.5): 68.42%
10. [Kimi K2.5](https://www.vals.ai/models/kimi_kimi-k2.5-thinking): 68.26%

## Excel Modeling Benchmark

- Page: https://www.vals.ai/benchmarks/emb
- Category: Finance
- Updated: 2026-10-18
- Models evaluated: 76

Evaluating agents on Excel-based financial modeling tasks

Top 10 by Accuracy:

1. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 76.67%
2. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 75.94%
3. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 75.71%
4. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 75.24%
5. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 73.67%
6. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 73.56%
7. [GPT-5.6 Sol](https://www.vals.ai/models/openai_gpt-5.6-sol): 72.34%
8. [Gemini 3.8 Flash](https://www.vals.ai/models/google_gemini-3.8-flash): 72.20%
9. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 71.92%
10. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 71.70%

## Finance Agent v2

- Page: https://www.vals.ai/benchmarks/fabv2
- Category: Finance
- Updated: 2026-10-18
- Models evaluated: 80

Evaluating agents on core financial analyst tasks

Top 10 by Accuracy:

1. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 65.40%
2. [Gemini 3.8 Flash](https://www.vals.ai/models/google_gemini-3.8-flash): 61.44%
3. [Muse Spark Open](https://www.vals.ai/models/meta_muse-spark-open): 61.34%
4. [Muse Spark 1.2](https://www.vals.ai/models/meta_muse_spark_1_2): 60.60%
5. [Muse Spark 1.3 Max](https://www.vals.ai/models/meta_muse_spark_1_3_max): 59.96%
6. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 59.06%
7. [Gemini 3.7 Flash](https://www.vals.ai/models/google_gemini-3.7-flash): 59.04%
8. [Muse Spark 1.3](https://www.vals.ai/models/meta_muse_spark_1_3): 58.90%
9. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 58.88%
10. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 58.63%

## MortgageTax

- Page: https://www.vals.ai/benchmarks/mortgage_tax
- Category: Finance (archived: no longer run on new models)
- Updated: 2026-09-01
- Models evaluated: 98

Evaluating reading and understanding tax certificates as images

Top 10 by Accuracy:

1. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 72.06%
2. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 70.79%
3. [Claude Opus 4.7](https://www.vals.ai/models/anthropic_claude-opus-4-7): 70.27%
4. [Claude Sonnet 5](https://www.vals.ai/models/anthropic_claude-sonnet-5): 70.03%
5. [Claude Opus 4.8](https://www.vals.ai/models/anthropic_claude-opus-4-8): 69.91%
6. [Gemini 3.1 Pro Preview (02/26)](https://www.vals.ai/models/google_gemini-3.1-pro-preview): 69.40%
7. [Gemini 3 Pro (11/25)](https://www.vals.ai/models/google_gemini-3-pro-preview): 69.08%
8. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 68.92%
9. [Gemini 2.5 Pro](https://www.vals.ai/models/google_gemini-2.5-pro): 68.92%
10. [GPT 5.5](https://www.vals.ai/models/openai_gpt-5.5): 68.76%

## Tax Agent Bench

- Page: https://www.vals.ai/benchmarks/tax_agent_bench
- Category: Finance
- Updated: 2026-10-18
- Models evaluated: 71

Evaluating agents on research-grade US corporate tax questions

Top 10 by Accuracy:

1. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 49.19%
2. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 45.57%
3. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 45.55%
4. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 44.56%
5. [DeepSeek V4.1 Flash 1017](https://www.vals.ai/models/deepseek_deepseek-v4.1-flash-1017): 44.35%
6. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 43.00%
7. [Muse Spark 1.3 Max](https://www.vals.ai/models/meta_muse_spark_1_3_max): 42.44%
8. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 41.96%
9. [GLM 5.3](https://www.vals.ai/models/zai_glm-5.3): 39.70%
10. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 38.28%

## TaxEval v2

- Page: https://www.vals.ai/benchmarks/tax_eval_v2
- Category: Finance (archived: no longer run on new models)
- Updated: 2026-09-01
- Models evaluated: 145

A Vals-created set of questions and responses to tax questions

Top 10 by Accuracy:

1. [Muse Spark 1.2](https://www.vals.ai/models/meta_muse_spark_1_2): 80.38%
2. [Muse Spark 1.1](https://www.vals.ai/models/meta_muse-spark-1.1): 79.72%
3. [Muse Spark](https://www.vals.ai/models/meta_muse_spark): 77.68%
4. [Claude Sonnet 4.6](https://www.vals.ai/models/anthropic_claude-sonnet-4-6): 77.11%
5. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 76.94%
6. [GPT-5.6 Terra](https://www.vals.ai/models/openai_gpt-5.6-terra): 76.17%
7. [GPT-5.6 Luna](https://www.vals.ai/models/openai_gpt-5.6-luna): 76.17%
8. [Claude Opus 4.6 (Thinking)](https://www.vals.ai/models/anthropic_claude-opus-4-6-thinking): 75.96%
9. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 75.96%
10. [Grok 3](https://www.vals.ai/models/grok_grok-3): 75.88%

## MedCode

- Page: https://www.vals.ai/benchmarks/medcode
- Category: Healthcare
- Updated: 2026-10-13
- Models evaluated: 107

Can models support the medical billing process?

Top 10 by Accuracy:

1. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 63.57%
2. [Gemini 3.1 Pro Preview (02/26)](https://www.vals.ai/models/google_gemini-3.1-pro-preview): 59.06%
3. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 58.80%
4. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 56.07%
5. [Gemini 3 Flash (12/25)](https://www.vals.ai/models/google_gemini-3-flash-preview): 55.92%
6. [Gemini 3.5 Flash](https://www.vals.ai/models/google_gemini-3.5-flash): 55.83%
7. [Claude Opus 4.7](https://www.vals.ai/models/anthropic_claude-opus-4-7): 54.86%
8. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 53.51%
9. [Gemini 3.7 Flash](https://www.vals.ai/models/google_gemini-3.7-flash): 53.39%
10. [Claude Opus 4.8](https://www.vals.ai/models/anthropic_claude-opus-4-8): 53.22%

## MedQA

- Page: https://www.vals.ai/benchmarks/medqa
- Category: Healthcare (archived: no longer run on new models)
- Updated: 2026-04-16
- Models evaluated: 95

Evaluating language model bias in medical questions.

Top 10 by Accuracy:

1. [o1](https://www.vals.ai/models/openai_o1-2024-12-17): 96.52%
2. [GPT 5.1](https://www.vals.ai/models/openai_gpt-5.1-2025-11-13): 96.38%
3. [Gemini 3.1 Pro Preview (02/26)](https://www.vals.ai/models/google_gemini-3.1-pro-preview): 96.37%
4. [GPT 5](https://www.vals.ai/models/openai_gpt-5-2025-08-07): 96.32%
5. [GPT 5.4 (xhigh)](https://www.vals.ai/models/openai_gpt-5.4-2026-03-05): 96.09%
6. [o3](https://www.vals.ai/models/openai_o3-2025-04-16): 96.06%
7. [GPT 5 Mini](https://www.vals.ai/models/openai_gpt-5-mini-2025-08-07): 96.06%
8. [Gemini 3 Pro (11/25)](https://www.vals.ai/models/google_gemini-3-pro-preview): 96.03%
9. [o4 Mini](https://www.vals.ai/models/openai_o4-mini-2025-04-16): 96.02%
10. [Claude Opus 4.5 (Thinking)](https://www.vals.ai/models/anthropic_claude-opus-4-5-20251101-thinking): 95.88%

## MedScribe

- Page: https://www.vals.ai/benchmarks/medscribe
- Category: Healthcare
- Updated: 2026-10-13
- Models evaluated: 109

Can models support doctors with their administrative work?

Top 10 by Accuracy:

1. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 91.43%
2. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 91.29%
3. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 91.10%
4. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 90.98%
5. [Muse Spark 1.2](https://www.vals.ai/models/meta_muse_spark_1_2): 90.06%
6. [Claude Haiku 5.5](https://www.vals.ai/models/anthropic_claude-haiku-5-5): 89.54%
7. [Grok 4.7](https://www.vals.ai/models/grok_grok-4.7): 89.38%
8. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 89.33%
9. [GLM 5.3 Flash](https://www.vals.ai/models/zai_glm-5.3-flash): 88.94%
10. [Muse Spark 1.1](https://www.vals.ai/models/meta_muse-spark-1.1): 88.89%

## AIME

- Page: https://www.vals.ai/benchmarks/aime
- Category: Math (archived: no longer run on new models)
- Updated: 2026-04-16
- Models evaluated: 96

Challenging national math exam given to top high-school students

Top 10 by Accuracy:

1. [Gemini 3.1 Pro Preview (02/26)](https://www.vals.ai/models/google_gemini-3.1-pro-preview): 98.13%
2. [GPT 5.2](https://www.vals.ai/models/openai_gpt-5.2-2025-12-11): 96.88%
3. [Muse Spark](https://www.vals.ai/models/meta_muse_spark): 96.88%
4. [Gemini 3 Pro (11/25)](https://www.vals.ai/models/google_gemini-3-pro-preview): 96.68%
5. [GPT 5.4 (xhigh)](https://www.vals.ai/models/openai_gpt-5.4-2026-03-05): 96.67%
6. [Grok 4.20 (Reasoning)](https://www.vals.ai/models/grok_grok-4.20-0309-reasoning): 96.46%
7. [Claude Opus 4.7](https://www.vals.ai/models/anthropic_claude-opus-4-7): 96.25%
8. [Gemini 3 Flash (12/25)](https://www.vals.ai/models/google_gemini-3-flash-preview): 95.63%
9. [Claude Opus 4.6 (Thinking)](https://www.vals.ai/models/anthropic_claude-opus-4-6-thinking): 95.63%
10. [GPT 5.4 Mini](https://www.vals.ai/models/openai_gpt-5.4-mini-2026-03-17): 95.63%

## MATH 500

- Page: https://www.vals.ai/benchmarks/math500
- Category: Math (archived: no longer run on new models)
- Updated: 2026-01-09
- Models evaluated: 60

Academic math benchmark on probability, algebra, and trigonometry

Top 10 by Accuracy:

1. [Gemini 3 Pro (11/25)](https://www.vals.ai/models/google_gemini-3-pro-preview): 96.40%
2. [Grok 4](https://www.vals.ai/models/grok_grok-4-0709): 96.20%
3. [GPT 5](https://www.vals.ai/models/openai_gpt-5-2025-08-07): 96.00%
4. [Claude Opus 4.1 (Thinking)](https://www.vals.ai/models/anthropic_claude-opus-4-1-20250805-thinking): 95.40%
5. [Gemini 2.5 Pro Exp](https://www.vals.ai/models/google_gemini-2.5-pro-exp-03-25): 95.20%
6. [GPT 5 Mini](https://www.vals.ai/models/openai_gpt-5-mini-2025-08-07): 94.80%
7. [GPT OSS 120B](https://www.vals.ai/models/fireworks_gpt-oss-120b): 94.80%
8. [o3](https://www.vals.ai/models/openai_o3-2025-04-16): 94.60%
9. [Qwen 3 (235B)](https://www.vals.ai/models/fireworks_qwen3-235b-a22b): 94.60%
10. [o4 Mini](https://www.vals.ai/models/openai_o4-mini-2025-04-16): 94.20%

## MGSM

- Page: https://www.vals.ai/benchmarks/mgsm
- Category: Math (archived: no longer run on new models)
- Updated: 2026-01-09
- Models evaluated: 75

A multilingual benchmark for mathematical questions.

Top 10 by Accuracy:

1. [Claude Opus 4.5 (Thinking)](https://www.vals.ai/models/anthropic_claude-opus-4-5-20251101-thinking): 95.20%
2. [Claude Opus 4.5 (Nonthinking)](https://www.vals.ai/models/anthropic_claude-opus-4-5-20251101): 94.76%
3. [Claude Opus 4.1 (Thinking)](https://www.vals.ai/models/anthropic_claude-opus-4-1-20250805-thinking): 94.44%
4. [Claude Sonnet 4.5 (Thinking)](https://www.vals.ai/models/anthropic_claude-sonnet-4-5-20250929-thinking): 94.33%
5. [Claude Opus 4.1 (Nonthinking)](https://www.vals.ai/models/anthropic_claude-opus-4-1-20250805): 94.22%
6. [GPT 5.2](https://www.vals.ai/models/openai_gpt-5.2-2025-12-11): 94.00%
7. [Gemini 3 Pro (11/25)](https://www.vals.ai/models/google_gemini-3-pro-preview): 93.93%
8. [Claude Opus 4 (Nonthinking)](https://www.vals.ai/models/anthropic_claude-opus-4-20250514): 93.78%
9. [o4 Mini](https://www.vals.ai/models/openai_o4-mini-2025-04-16): 93.42%
10. [Gemini 3 Flash (12/25)](https://www.vals.ai/models/google_gemini-3-flash-preview): 93.31%

## ProofBench v1.1

- Page: https://www.vals.ai/benchmarks/proof_bench
- Category: Math
- Updated: 2026-10-12
- Models evaluated: 50

Can models write math proofs that are formally verified?

Top 10 by Accuracy:

1. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 100.00%
2. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 100.00%
3. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 100.00%
4. AlephProver: 100.00%
5. [GPT-6.1 Sol](https://www.vals.ai/models/openai_gpt-6.1-sol): 99.00%
6. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 99.00%
7. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 99.00%
8. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 99.00%
9. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 95.00%
10. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 94.00%

## BioMysteryBench

- Page: https://www.vals.ai/benchmarks/biomysterybench
- Category: Science
- Updated: 2026-10-12
- Models evaluated: 27

Can agents solve agentic biology data investigations?

Top 10 by Accuracy:

1. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 81.11%
2. [GPT-6.1 Sol](https://www.vals.ai/models/openai_gpt-6.1-sol): 79.63%
3. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 79.26%
4. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 79.26%
5. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 79.26%
6. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 76.30%
7. [GPT-6 Sol](https://www.vals.ai/models/openai_gpt-6-sol): 74.81%
8. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 73.33%
9. [Kimi K3](https://www.vals.ai/models/kimi_kimi-k3): 72.59%
10. [Grok 4.6](https://www.vals.ai/models/grok_grok-4.6): 72.22%

## MysteryMechanism

- Page: https://www.vals.ai/benchmarks/mysterymechanism
- Category: Science
- Updated: 2026-10-12
- Models evaluated: 27

Can agents rediscover sealed mathematical mechanisms through bounded experiments?

Top 10 by Accuracy:

1. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 53.15%
2. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 49.55%
3. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 49.10%
4. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 47.75%
5. [GPT-6.1 Sol](https://www.vals.ai/models/openai_gpt-6.1-sol): 46.40%
6. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 45.49%
7. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 37.39%
8. [Gemini 3.8 Flash](https://www.vals.ai/models/google_gemini-3.8-flash): 36.49%
9. [Muse Spark 1.3 Max](https://www.vals.ai/models/meta_muse_spark_1_3_max): 36.04%
10. [Claude Haiku 5.5](https://www.vals.ai/models/anthropic_claude-haiku-5-5): 35.13%

## Robo-bench

- Page: https://www.vals.ai/benchmarks/robo_bench
- Category: Science
- Updated: 2026-10-13
- Models evaluated: 7

Can a model control a robot arm and a humanoid in a real-time physics simulation, from motor commands and camera images alone? Most tasks are held out.

Top 7 by Accuracy:

1. [GPT-6.1 Sol](https://www.vals.ai/models/openai_gpt-6.1-sol): 28.33%
2. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 23.33%
3. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 14.58%
4. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 10.42%
5. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 5.83%
6. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 0.83%
7. [Grok 4.7](https://www.vals.ai/models/grok_grok-4.7): 0.21%

## Terminal-Bench-Science

- Page: https://www.vals.ai/benchmarks/terminal-bench-science
- Category: Science
- Updated: 2026-10-14
- Models evaluated: 41

Research workflow tasks contributed by practicing scientists across five domains

Top 10 by Score:

1. [GPT-6 Astra (Mini-SWE-agent)](https://www.vals.ai/models/openai_gpt-6-astra): 62.86%
2. [GPT-6.1 Sol (Mini-SWE-agent)](https://www.vals.ai/models/openai_gpt-6.1-sol): 52.86%
3. [Claude Opus 5.5 (Mini-SWE-agent)](https://www.vals.ai/models/anthropic_claude-opus-5-5): 47.14%
4. [Claude Sonnet 5.5 (Mini-SWE-agent)](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 47.14%
5. [Gemini 4 Argon (Mini-SWE-agent)](https://www.vals.ai/models/google_gemini-4-argon): 44.29%
6. [Claude Fable 5.1 (Mini-SWE-agent)](https://www.vals.ai/models/anthropic_claude-fable-5-1): 40.00%
7. [Wonder Full (Mini-SWE-agent)](https://www.vals.ai/models/meta_muse-wonder-full): 37.14%
8. [GPT-6 Sol (Mini-SWE-agent)](https://www.vals.ai/models/openai_gpt-6-sol): 30.00%
9. [Claude Opus 5 (Mini-SWE-agent)](https://www.vals.ai/models/anthropic_claude-opus-5): 27.14%
10. [Wonder (Mini-SWE-agent)](https://www.vals.ai/models/meta_muse-wonder): 22.86%

## GPQA Diamond

- Page: https://www.vals.ai/benchmarks/gpqa
- Category: Academic (archived: no longer run on new models)
- Updated: 2026-09-01
- Models evaluated: 138

Graduate-level Google-Proof Q&A benchmark evaluating models on questions that require deep reasoning.

Top 10 by Accuracy:

1. [Gemini 3.1 Pro Preview (02/26)](https://www.vals.ai/models/google_gemini-3.1-pro-preview): 95.45%
2. [GPT-5.6 Sol](https://www.vals.ai/models/openai_gpt-5.6-sol): 95.20%
3. [Grok 4.6](https://www.vals.ai/models/grok_grok-4.6): 94.70%
4. [Gemini 3.8 Flash](https://www.vals.ai/models/google_gemini-3.8-flash): 94.44%
5. [Gemini 3.7 Flash](https://www.vals.ai/models/google_gemini-3.7-flash): 93.94%
6. [Qwen 3.8 Max](https://www.vals.ai/models/alibaba_qwen3.8-max): 93.69%
7. [Gemini 3.6 Flash](https://www.vals.ai/models/google_gemini-3.6-flash): 93.43%
8. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 93.43%
9. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 93.43%
10. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 93.18%

## MMLU Pro

- Page: https://www.vals.ai/benchmarks/mmlu_pro
- Category: Academic (archived: no longer run on new models)
- Updated: 2026-09-01
- Models evaluated: 138

Academic multiple-choice benchmark covering 14 subjects including STEM, humanities, and social sciences.

Top 10 by Accuracy:

1. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 92.38%
2. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 91.59%
3. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 91.50%
4. [Gemini 3.1 Pro Preview (02/26)](https://www.vals.ai/models/google_gemini-3.1-pro-preview): 90.99%
5. [Gemini 3.8 Flash](https://www.vals.ai/models/google_gemini-3.8-flash): 90.22%
6. [Gemini 3.7 Flash](https://www.vals.ai/models/google_gemini-3.7-flash): 90.12%
7. [Gemini 3 Pro (11/25)](https://www.vals.ai/models/google_gemini-3-pro-preview): 90.10%
8. [Claude Opus 4.7](https://www.vals.ai/models/anthropic_claude-opus-4-7): 89.87%
9. [Claude Opus 4.8](https://www.vals.ai/models/anthropic_claude-opus-4-8): 89.58%
10. [Gemini 3.5 Flash](https://www.vals.ai/models/google_gemini-3.5-flash): 89.52%

## MMMU Pro

- Page: https://www.vals.ai/benchmarks/mmmu
- Category: Academic (archived: no longer run on new models)
- Updated: 2026-09-01
- Models evaluated: 93

Multimodal Multi-task Benchmark

Top 10 by Accuracy:

1. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 90.64%
2. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 89.88%
3. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 89.31%
4. [Gemini 3.8 Flash](https://www.vals.ai/models/google_gemini-3.8-flash): 89.08%
5. [Gemini 3.7 Flash](https://www.vals.ai/models/google_gemini-3.7-flash): 88.96%
6. [GPT-5.6 Sol](https://www.vals.ai/models/openai_gpt-5.6-sol): 88.84%
7. [Gemini 3.6 Flash](https://www.vals.ai/models/google_gemini-3.6-flash): 88.38%
8. [Gemini 3.5 Flash](https://www.vals.ai/models/google_gemini-3.5-flash): 88.27%
9. [GPT 5.5](https://www.vals.ai/models/openai_gpt-5.5): 88.27%
10. [Gemini 3.1 Pro Preview (02/26)](https://www.vals.ai/models/google_gemini-3.1-pro-preview): 88.21%

## SAGE

- Page: https://www.vals.ai/benchmarks/sage
- Category: Education
- Updated: 2026-10-08
- Models evaluated: 91

Student Assessment with Generative Evaluation

Top 10 by Accuracy:

1. [Claude Opus 4.7](https://www.vals.ai/models/anthropic_claude-opus-4-7): 56.10%
2. [Gemma 4 31B IT](https://www.vals.ai/models/google_gemma-4-31b-it): 55.03%
3. [Claude Opus 4.8](https://www.vals.ai/models/anthropic_claude-opus-4-8): 54.79%
4. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 53.65%
5. [Kimi K3](https://www.vals.ai/models/kimi_kimi-k3): 52.78%
6. [GPT-5.6 Sol](https://www.vals.ai/models/openai_gpt-5.6-sol): 52.56%
7. [Qwen 3.8 27B](https://www.vals.ai/models/alibaba_qwen3.8-27b): 52.40%
8. [Claude Opus 4.5 (Thinking)](https://www.vals.ai/models/anthropic_claude-opus-4-5-20251101-thinking): 52.09%
9. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 51.89%
10. [Gemini 3 Flash (12/25)](https://www.vals.ai/models/google_gemini-3-flash-preview): 51.85%

## VoiceCodeBench

- Page: https://www.vals.ai/benchmarks/voice-code-bench
- Category: Voice
- Updated: 2026-09-24
- Models evaluated: 20

How good are speech-to-text models at handling exact structured values in English workplace speech?

Top 10 by TSR ↑:

1. [GPT Live Transcribe](https://www.vals.ai/models/openai_gpt-live-transcribe): 67.67%
2. [Grok Voice Transcribe 2.0](https://www.vals.ai/models/grok_grok-voice-transcribe-2.0): 65.33%
3. [Cartesia Ink 2](https://www.vals.ai/models/cartesia_ink-2): 62.00%
4. [GPT 4o Transcribe](https://www.vals.ai/models/openai_gpt-4o-transcribe): 59.67%
5. [Deepgram Nova-3 (Streaming)](https://www.vals.ai/models/deepgram_nova-3-streaming): 59.33%
6. [Reson8 Resonant-1](https://www.vals.ai/models/reson8_resonant-1): 59.00%
7. [Inworld STT 1](https://www.vals.ai/models/inworld_inworld-stt-1): 57.33%
8. [Whisper Large-v3 Turbo](https://www.vals.ai/models/groq_whisper-large-v3-turbo): 57.00%
9. [Muse Voice Transcribe](https://www.vals.ai/models/meta_muse_voice_transcribe): 57.00%
10. [Deepgram Flux General English](https://www.vals.ai/models/deepgram_flux-general-en): 55.67%

## Code Migration

- Page: https://www.vals.ai/benchmarks/code-migration
- Category: Coding
- Updated: 2026-10-18
- Models evaluated: 79

Can language models reimplement working programs in another language?

Top 10 by Accuracy:

1. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 69.83%
2. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 68.17%
3. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 67.74%
4. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 66.65%
5. [GPT-6.1 Sol](https://www.vals.ai/models/openai_gpt-6.1-sol): 65.12%
6. [Wonder](https://www.vals.ai/models/meta_muse-wonder): 62.55%
7. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 61.03%
8. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 57.47%
9. [GPT-6 Sol](https://www.vals.ai/models/openai_gpt-6-sol): 57.20%
10. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 55.06%

## IOI

- Page: https://www.vals.ai/benchmarks/ioi
- Category: Coding
- Updated: 2026-10-14
- Models evaluated: 44

International Olympiad in Informatics

Top 10 by Accuracy:

1. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 100.00%
2. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 100.00%
3. [GPT-6.1 Sol](https://www.vals.ai/models/openai_gpt-6.1-sol): 96.89%
4. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 95.06%
5. [GPT-5.6 Sol](https://www.vals.ai/models/openai_gpt-5.6-sol): 91.17%
6. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 90.78%
7. [GPT-5.6 Terra](https://www.vals.ai/models/openai_gpt-5.6-terra): 87.61%
8. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 84.33%
9. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 83.06%
10. [GPT-6 Sol](https://www.vals.ai/models/openai_gpt-6-sol): 82.61%

## LiveCodeBench

- Page: https://www.vals.ai/benchmarks/lcb
- Category: Coding (archived: no longer run on new models)
- Updated: 2026-09-01
- Models evaluated: 143

Our Implementation of the LiveCodeBench benchmark

Top 10 by Accuracy:

1. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 90.52%
2. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 89.78%
3. [Gemini 3.8 Flash](https://www.vals.ai/models/google_gemini-3.8-flash): 89.48%
4. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 89.03%
5. [Gemini 3.7 Flash](https://www.vals.ai/models/google_gemini-3.7-flash): 88.65%
6. [Gemini 3.1 Pro Preview (02/26)](https://www.vals.ai/models/google_gemini-3.1-pro-preview): 88.48%
7. [Grok 4.6](https://www.vals.ai/models/grok_grok-4.6): 88.22%
8. [Gemini 3.6 Flash](https://www.vals.ai/models/google_gemini-3.6-flash): 88.08%
9. [GPT 5.2 Codex](https://www.vals.ai/models/openai_gpt-5.2-codex): 87.99%
10. [Qwen 3.8 Max](https://www.vals.ai/models/alibaba_qwen3.8-max): 87.85%

## ProgramBench

- Page: https://www.vals.ai/benchmarks/programbench
- Category: Coding
- Updated: 2026-10-12
- Models evaluated: 64

Can language models rebuild programs from scratch?

Top 10 by Accuracy:

1. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 18.50%
2. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 7.00%
3. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 6.50%
4. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 5.50%
5. [GPT-6.1 Sol](https://www.vals.ai/models/openai_gpt-6.1-sol): 3.00%
6. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 3.00%
7. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 2.50%
8. [Claude Haiku 5.5](https://www.vals.ai/models/anthropic_claude-haiku-5-5): 2.50%
9. [Muse Spark 1.3 Max](https://www.vals.ai/models/meta_muse_spark_1_3_max): 2.50%
10. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 2.00%

## SkillsBench

- Page: https://www.vals.ai/benchmarks/skillsbench
- Category: Coding
- Updated: 2026-09-27
- Models evaluated: 34 (35 model and harness pairs)

How important are skills for agents?

Top 10 by Score:

1. [DeepSeek V4.1 Flash (OpenHands)](https://www.vals.ai/models/deepseek_deepseek-v4.1-flash): 69.80%
2. [Grok 4.5 (OpenHands)](https://www.vals.ai/models/grok_grok-4.5): 66.03%
3. [Gemini 3.7 Flash (OpenHands)](https://www.vals.ai/models/google_gemini-3.7-flash): 65.89%
4. [GPT 5.5 (Codex)](https://www.vals.ai/models/openai_gpt-5.5): 62.55%
5. [GPT 5.5 (OpenHands)](https://www.vals.ai/models/openai_gpt-5.5): 62.21%
6. [Claude Fable 5.1 (OpenHands)](https://www.vals.ai/models/anthropic_claude-fable-5-1): 61.55%
7. [GPT-5.6 Luna (OpenHands)](https://www.vals.ai/models/openai_gpt-5.6-luna): 60.45%
8. [Claude Opus 5 (OpenHands)](https://www.vals.ai/models/anthropic_claude-opus-5): 60.44%
9. [Claude Opus 4.8 (OpenHands)](https://www.vals.ai/models/anthropic_claude-opus-4-8): 59.23%
10. [Muse Spark 1.1 (OpenHands)](https://www.vals.ai/models/meta_muse-spark-1.1): 59.19%

## SWE-bench Multimodal

- Page: https://www.vals.ai/benchmarks/swebench-multimodal
- Category: Coding
- Updated: 2026-10-16
- Models evaluated: 20

Do AI Systems Generalize to Visual Software Domains?

Top 10 by Accuracy:

1. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 70.21%
2. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 61.25%
3. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 58.33%
4. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 56.04%
5. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 47.29%
6. [DeepSeek V4.1 Flash](https://www.vals.ai/models/deepseek_deepseek-v4.1-flash): 46.46%
7. [Wonder](https://www.vals.ai/models/meta_muse-wonder): 45.63%
8. [Kimi K3](https://www.vals.ai/models/kimi_kimi-k3): 44.17%
9. [Claude Haiku 5.5](https://www.vals.ai/models/anthropic_claude-haiku-5-5): 43.96%
10. [GPT-6.1 Sol](https://www.vals.ai/models/openai_gpt-6.1-sol): 43.54%

## SWE-bench Verified

- Page: https://www.vals.ai/benchmarks/swebench
- Category: Coding (archived: no longer run on new models)
- Updated: 2026-09-01
- Models evaluated: 85 (88 model and harness pairs)

Solving production software engineering tasks

Top 10 by Score:

1. [Claude Opus 5 (Mini-SWE-agent)](https://www.vals.ai/models/anthropic_claude-opus-5): 97.00%
2. [DeepSeek V4 Pro 0813 (Mini-SWE-agent)](https://www.vals.ai/models/deepseek_deepseek-v4-pro-0813): 96.40%
3. [GPT-5.6 Sol (Mini-SWE-agent)](https://www.vals.ai/models/openai_gpt-5.6-sol): 96.20%
4. [Grok 4.6 (Mini-SWE-agent)](https://www.vals.ai/models/grok_grok-4.6): 95.60%
5. [GPT-5.6 Terra (Mini-SWE-agent)](https://www.vals.ai/models/openai_gpt-5.6-terra): 95.40%
6. [GLM 5.3 (Mini-SWE-agent)](https://www.vals.ai/models/zai_glm-5.3): 95.40%
7. [Claude Fable 5 (Mini-SWE-agent)](https://www.vals.ai/models/anthropic_claude-fable-5): 95.00%
8. [Kimi K3 (Mini-SWE-agent)](https://www.vals.ai/models/kimi_kimi-k3): 93.40%
9. [GPT-5.6 Luna (Mini-SWE-agent)](https://www.vals.ai/models/openai_gpt-5.6-luna): 93.00%
10. [GLM 5.3 Flash (Mini-SWE-agent)](https://www.vals.ai/models/zai_glm-5.3-flash): 92.00%

## Terminal-Bench 4.0

- Page: https://www.vals.ai/benchmarks/terminal-bench-4
- Category: Coding
- Updated: 2026-10-18
- Models evaluated: 50

Frontier-difficulty terminal tasks across software, science, ML, operations, hardware, security, and media

Top 10 by Score:

1. [Claude Opus 5.5 (Mini-SWE-agent)](https://www.vals.ai/models/anthropic_claude-opus-5-5): 65.15%
2. [Claude Sonnet 5.5 (Mini-SWE-agent)](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 64.14%
3. [GPT-6 Astra (Mini-SWE-agent)](https://www.vals.ai/models/openai_gpt-6-astra): 59.60%
4. [Claude Fable 5.1 (Mini-SWE-agent)](https://www.vals.ai/models/anthropic_claude-fable-5-1): 58.08%
5. [Wonder Full (Mini-SWE-agent)](https://www.vals.ai/models/meta_muse-wonder-full): 58.08%
6. [Gemini 4 Argon (Mini-SWE-agent)](https://www.vals.ai/models/google_gemini-4-argon): 57.58%
7. [GPT-6.1 Sol (Mini-SWE-agent)](https://www.vals.ai/models/openai_gpt-6.1-sol): 55.05%
8. [Claude Opus 5 (Mini-SWE-agent)](https://www.vals.ai/models/anthropic_claude-opus-5): 53.53%
9. [Wonder (Mini-SWE-agent)](https://www.vals.ai/models/meta_muse-wonder): 49.49%
10. [GPT-6 Sol (Mini-SWE-agent)](https://www.vals.ai/models/openai_gpt-6-sol): 44.44%

## Vibe Code Bench 1-100

- Page: https://www.vals.ai/benchmarks/vcb-1-100
- Category: Coding
- Updated: 2026-10-14
- Models evaluated: 25

Can models extend a working web application across future user requests?

Top 10 by Accuracy:

1. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 33.98%
2. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 30.36%
3. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 28.53%
4. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 28.00%
5. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 27.64%
6. [Wonder](https://www.vals.ai/models/meta_muse-wonder): 26.55%
7. [GPT-5.6 Luna](https://www.vals.ai/models/openai_gpt-5.6-luna): 22.59%
8. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 21.20%
9. [Muse Spark 1.3 Max](https://www.vals.ai/models/meta_muse_spark_1_3_max): 20.46%
10. [Claude Haiku 5.5](https://www.vals.ai/models/anthropic_claude-haiku-5-5): 20.41%

## Vibe Code Bench v1.1

- Page: https://www.vals.ai/benchmarks/vibe-code
- Category: Coding
- Updated: 2026-10-18
- Models evaluated: 110 (114 model and harness pairs)

Can models build web applications from scratch?

Top 10 by Score:

1. [Claude Sonnet 5.5 (OpenHands)](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 92.39%
2. [Gemini 4 Argon (OpenHands)](https://www.vals.ai/models/google_gemini-4-argon): 91.91%
3. [Claude Haiku 5.5 (OpenHands)](https://www.vals.ai/models/anthropic_claude-haiku-5-5): 90.44%
4. [Claude Fable 5 (OpenHands)](https://www.vals.ai/models/anthropic_claude-fable-5): 90.35%
5. [Claude Opus 5.5 (OpenHands)](https://www.vals.ai/models/anthropic_claude-opus-5-5): 90.29%
6. [Claude Fable 5.1 (OpenHands)](https://www.vals.ai/models/anthropic_claude-fable-5-1): 90.26%
7. [GPT-6 Astra (OpenHands)](https://www.vals.ai/models/openai_gpt-6-astra): 89.59%
8. [GPT-6.1 Sol (OpenHands)](https://www.vals.ai/models/openai_gpt-6.1-sol): 88.93%
9. [Claude Opus 5 (OpenHands)](https://www.vals.ai/models/anthropic_claude-opus-5): 88.40%
10. [GPT-6 Sol (OpenHands)](https://www.vals.ai/models/openai_gpt-6-sol): 87.82%

## CyberBench v1.1

- Page: https://www.vals.ai/benchmarks/cyber
- Category: Cyber
- Updated: 2026-10-12
- Models evaluated: 47

Can AI agents find security bugs and patch them?

Top 10 by Accuracy:

1. [Wonder](https://www.vals.ai/models/meta_muse-wonder): 80.36%
2. [GPT-6 Sol](https://www.vals.ai/models/openai_gpt-6-sol): 77.98%
3. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 77.86%
4. [GPT-5.6 Sol](https://www.vals.ai/models/openai_gpt-5.6-sol): 76.31%
5. [GPT-6 Luna](https://www.vals.ai/models/openai_gpt-6-luna): 76.25%
6. [Claude Haiku 5.5](https://www.vals.ai/models/anthropic_claude-haiku-5-5): 75.48%
7. [MiMo V2.6 Flash](https://www.vals.ai/models/xiaomi_mimo-v2.6-flash): 75.36%
8. [Kimi K3](https://www.vals.ai/models/kimi_kimi-k3): 75.24%
9. [Claude Opus 5.5 (CVP)](https://www.vals.ai/models/anthropic_claude-opus-5-5-range): 74.58%
10. [DeepSeek V4.1 Flash](https://www.vals.ai/models/deepseek_deepseek-v4.1-flash): 73.69%

## SRE Bench

- Page: https://www.vals.ai/benchmarks/srebench
- Category: Cyber
- Updated: 2026-10-14
- Models evaluated: 35

Can AI agents work out what a real-world binary does without its source code?

Top 10 by Accuracy:

1. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 56.87%
2. [GPT-6.1 Sol](https://www.vals.ai/models/openai_gpt-6.1-sol): 50.76%
3. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 44.27%
4. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 33.59%
5. [Daybreak Blue](https://www.vals.ai/models/openai_gpt-daybreak-blue): 31.68%
6. [GPT-5.6 Sol](https://www.vals.ai/models/openai_gpt-5.6-sol): 30.53%
7. [GPT-6 Sol](https://www.vals.ai/models/openai_gpt-6-sol): 30.53%
8. [Claude Sonnet 5.5](https://www.vals.ai/models/anthropic_claude-sonnet-5-5): 30.15%
9. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 22.90%
10. [Claude Opus 5.5 (CVP)](https://www.vals.ai/models/anthropic_claude-opus-5-5-range): 21.76%

## CUA-bench

- Page: https://www.vals.ai/benchmarks/cua_bench
- Category: Games
- Updated: 2026-10-13
- Models evaluated: 9

Can an AI agent play six commercial video games with only a keyboard and a mouse? Half of them are held out.

Top 9 by Accuracy:

1. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 19.17%
2. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 14.00%
3. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 13.17%
4. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 9.00%
5. [GPT-5.6 Sol](https://www.vals.ai/models/openai_gpt-5.6-sol): 8.33%
6. [Wonder](https://www.vals.ai/models/meta_muse-wonder): 7.17%
7. [Muse Spark 1.3 Max](https://www.vals.ai/models/meta_muse_spark_1_3_max): 5.83%
8. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 4.83%
9. [Gemini 3.8 Flash](https://www.vals.ai/models/google_gemini-3.8-flash): 4.17%

## Poker Agent

- Page: https://www.vals.ai/benchmarks/poker_agent
- Category: Games
- Updated: 2025-12-23
- Models evaluated: 17

Which model can make the most money playing poker?

Top 10 by TrueSkill Rating:

1. [GPT 5.2](https://www.vals.ai/models/openai_gpt-5.2-2025-12-11): 1131.83 (TrueSkill Rating)
2. [GPT 5](https://www.vals.ai/models/openai_gpt-5-2025-08-07): 1103.18 (TrueSkill Rating)
3. [Gemini 3 Flash (12/25)](https://www.vals.ai/models/google_gemini-3-flash-preview): 1100.21 (TrueSkill Rating)
4. [DeepSeek V3.2 (Thinking)](https://www.vals.ai/models/fireworks_deepseek-v3p2-thinking): 1090.30 (TrueSkill Rating)
5. [Grok 4.1 Fast (Reasoning)](https://www.vals.ai/models/grok_grok-4-1-fast-reasoning): 1079.22 (TrueSkill Rating)
6. [Gemini 3 Pro (11/25)](https://www.vals.ai/models/google_gemini-3-pro-preview): 1078.90 (TrueSkill Rating)
7. [DeepSeek V3.1](https://www.vals.ai/models/fireworks_deepseek-v3p1): 1058.23 (TrueSkill Rating)
8. [Claude Sonnet 4.5 (Nonthinking)](https://www.vals.ai/models/anthropic_claude-sonnet-4-5-20250929): 1055.50 (TrueSkill Rating)
9. [GPT 5.1](https://www.vals.ai/models/openai_gpt-5.1-2025-11-13): 1038.59 (TrueSkill Rating)
10. [Grok 4 Fast (Reasoning)](https://www.vals.ai/models/grok_grok-4-fast-reasoning): 1034.30 (TrueSkill Rating)

## Time Horizon Index: KSP

- Page: https://www.vals.ai/benchmarks/time_horizon_index
- Category: Games
- Updated: 2026-10-07
- Models evaluated: 13

Can an AI agent build and run a space program in Kerbal Space Program?

Top 10 by Overall:

1. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 91.33%
2. [GPT-6 Astra](https://www.vals.ai/models/openai_gpt-6-astra): 90.50%
3. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 63.33%
4. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 52.50%
5. [GPT-5.6 Sol](https://www.vals.ai/models/openai_gpt-5.6-sol): 23.83%
6. [Gemini 3.8 Flash](https://www.vals.ai/models/google_gemini-3.8-flash): 18.83%
7. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 18.83%
8. [Claude Opus 4.8](https://www.vals.ai/models/anthropic_claude-opus-4-8): 11.83%
9. [Kimi K3](https://www.vals.ai/models/kimi_kimi-k3): 10.50%
10. [GPT 5.5](https://www.vals.ai/models/openai_gpt-5.5): 8.33%

## Public Benefits Bench v1.1

- Page: https://www.vals.ai/benchmarks/public-benefits-bench
- Category: Social Mobility
- Updated: 2026-10-12
- Models evaluated: 50

Can AI help people navigate SNAP benefits?

Top 10 by Accuracy:

1. [Claude Opus 5](https://www.vals.ai/models/anthropic_claude-opus-5): 76.93%
2. [Claude Fable 5.1](https://www.vals.ai/models/anthropic_claude-fable-5-1): 74.90%
3. [Wonder Full](https://www.vals.ai/models/meta_muse-wonder-full): 73.75%
4. [Claude Opus 5.5](https://www.vals.ai/models/anthropic_claude-opus-5-5): 70.64%
5. [Claude Fable 5](https://www.vals.ai/models/anthropic_claude-fable-5): 70.43%
6. [Wonder](https://www.vals.ai/models/meta_muse-wonder): 70.23%
7. [Gemini 4 Argon](https://www.vals.ai/models/google_gemini-4-argon): 69.76%
8. [MiMo V2.6 Pro](https://www.vals.ai/models/xiaomi_mimo-v2.6-pro): 68.94%
9. [Hy4 Preview](https://www.vals.ai/models/tencent_hy4-preview): 68.61%
10. [GLM 5.3](https://www.vals.ai/models/zai_glm-5.3): 68.54%