Vals AI: The Cheaper Model That Costs More, and the Team That Cut $1.5M Token Spend by 90%

AI Engineer · youtube · 2026-10-30

Vals AI CEO Rayan Krishnan explains why single model-card scores hide a jagged frontier:

• The best model varies by task — in finance work like tax and Excel LBO modeling, no single winner exists
• A model cheaper on paper can cost more overall by burning extra tokens while scoring lower
• Benchmark lessons: choosing the right accuracy metric, reward hacking, and multi-turn evals are hard; agents solve the test, not the problem
• Case study: a team spent $1.5M on tokens in one month, then cut spend 90% using internal evals
• A firm's own work history is its best eval data for model selection

Original post →

More from Venture

• Investor's Last Exam Benchmark: Top AI Models Ace Research but Fail at Investment Judgment — Huangfei Jiang · 2026-10-30
• Kuaishou Team Proposes GR4AD-X, Upgrading Generative Ad Recommendation with Multi-Path Semantic IDs — _reachsumit · 2026-10-30
• Mintive Raises $11M to Give Every AI Agent Its Own Identity and Permissions — socialwithaayan · 2026-10-30
• Domingos snarks: Anthropic smart to IPO before Claude Code goes the way of AOL — pmddomingos · 2026-10-30
• Team wins $10M Laude Institute Moonshot prize to build open health AI infrastructure with JupyterHealth — irenetrampoline · 2026-10-30
• House Democrats push resolution against AI bailouts as industry debt nears TARP scale — aakashgupta · 2026-10-30

Venture channel →