Vals AI provides independent AI evaluation and benchmarking services, emphasizing unbiased benchmark results. Its core test subjects are real tasks with high economic value, covering fields such as finance, software, cybersecurity, and healthcare, and also involving frontier risks such as recursive self-improvement and mental health.
Main capabilities include: running its own evaluations and building multiple benchmarks; publishing data such as rankings, scores, and per-test costs of models on various benchmarks; evaluating major AI model releases, such as Meta's Wonder Full and Wonder, zAI's GLM Legal, Anthropic's Claude Haiku 5.5 and Claude Sonnet 5.5, Google's Gemini 4 Argon, OpenAI's GPT-6 Sol and GPT-6 Luna, etc.
Examples of benchmarks involved on the platform: Vals Index, CyberBench, Public Benefits Bench, Harvey's Legal Agent Benchmark, LegalBench, Legal Research Bench, Vibe Code Bench, SAFE-Teen, MedScribe, Code Migration, EMB, SRE Bench, Terminal-Bench Science, SAGE, BioMysteryBench, MysteryMechanism, Finance Agent v2, IOI, Tax Agent Bench, ProgramBench, ProofBench v1.1, CUA-bench, SNAP, MedCode, etc. Some benchmarks are built by the platform itself, for example SAFE-Teen is used to test the safety of AI chatbots' responses when teenagers seek mental health support.
Vals AI has also planned or launched the Vals Web Search Index, with Exa, Keenable, Parallel, Tavily, etc. as partners, running independently using unified settings, scoring agents' final answers on private legal and financial expert tasks, changing only the search tool while keeping the model and execution framework fixed.
In terms of usage, the materials do not provide specific instructions for registration, invocation, or access interfaces, and only provide the website link https://www.vals.ai/home。评估通常由平台自行运行，并记录推理努力、温度等设置；部分模型测试使用回退模型处理拒答，并将拒答或提供方策略阻断计为失败任务。
Applicable scenarios include: developers or enterprises comparing model performance and costs on different real tasks during selection; researchers learning about model benchmark rankings in professional fields such as finance, law, programming, security, and healthcare; and people concerned about AI safety and mental health risks viewing relevant evaluation results.
Limitations: the materials do not provide whether the platform charges fees, whether it opens an API, specific usage steps, or the platform's support scope. Some model data is exemplary release content, and the actual latest results should be based on what the platform publishes.
