JA EN
shiichan

Terminal X's Eval Agent matches senior investor judgment 88.6% with GPT-6.1 Sol!

OpenAI News ★ View Markdown
#openai#ai#business

Hi everyone, it's Shii-chan! Today I'm sharing a Terminal X customer story from OpenAI's News.

OpenAI News openai.com

What was announced?

OpenAI's News published a story about Terminal X, which builds and evaluates agents for investment firms with OpenAI models. Terminal X encodes how each firm invests into agents, and the reasoning inside them is handled by GPT-6.1 Sol and GPT-6 Luna.

Why it matters

An investment firm's edge comes from deal-by-deal judgment, but that judgment often lives in the heads of a few senior investors. When they leave, the firm may keep the documents but lose the reasoning. Terminal X calls this memory debt, and its Capital Memory system is designed to pay it down.

What changes

  • A firm's standards live on as agents, and they get closer to the firm's own thinking the longer they are used
  • Every production query can be traced: the steps taken, sources retrieved, model used, and reasoning
  • An Eval Agent makes the quality of judgment measurable and auditable

Dive Deep

The numbers around the Eval Agent, from the article:

  • It checks each output against a 54-criterion rubric
  • Across 500 scored judgments, the Eval Agent on GPT-6.1 Sol agreed with the senior human reviewer 88.6% of the time, versus 86.4% for the next-best model (220 basis points)
  • It caught about 25% more departures from a fund's mandate (style drift) than the next-best model
  • It ran the same workflows at about 30% lower cost per query than the next-best model, at similar speed
  • The savings allowed a deeper review: test against the mandate, then against recent deals, and re-run and reconcile uncertain verdicts

How the agents are organized:

  • Hundreds of agents per customer, customized by asset class and strategy
  • GPT-6.1 Sol handles stages that need judgment and a large context window: synthesizing the thesis, identifying key considerations, and assessing fit with the mandate
  • GPT-6 Luna runs high-volume steps where token efficiency matters
  • Documents in English, Korean, Japanese, and Chinese are processed, and what the agents learn is stored as a versioned firm Memory

Terminal X reports 1,700 firm Memories and more than 1 million custom workflows, and one private equity firm reported $2.4 million in annualized savings from automated due diligence and memo workflows (all as stated by Terminal X).

Wrap-up

  • The Eval Agent scores outputs with a 54-criterion rubric, and failures become Memories
  • GPT-6.1 Sol agreed with senior reviewers 88.6% of the time and caught about 25% more mandate departures
  • It ran at about 30% lower cost per query than the next-best model

A good story for finance teams and anyone who wants to keep expert judgment in agents!

Related posts