Free ChatGPT and Gemini trail Vals finance benchmark

Vals AI published Financial Coaching Bench, which tests how well AI models coach a household through a financial problem. The project is backed by NextLadder Ventures and advised by the Financial Health Network.

Each of the 256 scored scenarios includes a client's opening message, a hidden fact sheet covering income, debts and deadlines, and a checklist written by practitioners. The model under test acts as the coach and can ask questions and search the web. Claude Sonnet 4.6 plays the client. Borrowing scenarios make up 28% of the set. The test set is private except for two example scenarios.

Claude Opus 5.5 ranks first, but the top 12 models are within 2.4 points of each other. MiMo V2.6 Pro, in fifth place, costs $0.013 per scenario, less than 1% of Opus 5.5's $1.63.

The default models behind free ChatGPT and Google's free AI products rank 25th to 28th, 11 to 19 points below the leader. GPT-6 Luna and Gemini 3.5 Flash-Lite, which took over free ChatGPT and Gemini in October, score lower than the models they replaced. Free Claude ranks 7th. When a practitioner expected a referral to free help, the free ChatGPT and Gemini models gave one in about 70% and 60% of conversations, compared with 75–94% for the leading models.

About 95% of failed checklist items were advice the coach left out or only partly gave. Longer answers also scored higher: within a model, each doubling of response length is associated with about 3 more points. When responses are compared at equal length, GPT-5.6 Sol moves from 8th to 1st.