Terminal X matches senior investor judgment 88.6% with GPT-6.1

In an OpenAI customer story, Terminal X says its Eval Agent running on GPT-6.1 Sol agreed with a senior human reviewer 88.6% of the time across 500 scored judgments, versus 86.4% for the next-best model. It also says the setup caught about 25% more recommendations that drifted from a fund's mandate and ran at about 30% lower cost per query.

The bigger idea is what the company calls memory debt: firms keep the documents when senior investors leave, but lose the reasoning. Terminal X embeds former finance professionals with clients to encode that reasoning into agents the firm owns and runs on-premise or in its own cloud.

We like the shape of this more than the headline number. The design that stands out is the Eval Agent: every output gets checked against a 54-criterion rubric covering sources, comps, narrative and fit with the house view, each check pass or fail with a one-line reason, and every failure becomes a new instruction or test case. Every production query keeps a full trail of steps, sources, model and reasoning. That's what useful AI in a regulated, high-stakes business should look like: measurable, auditable and correctable.

But keep the numbers in perspective. This is a vendor case study published by the model maker, and the 88.6% figure measures agreement between an evaluation agent and one senior reviewer, not investment returns. The gap over the next-best model is 2.2 points on 500 judgments. Meaningful, perhaps, but not a revolution, and we don't know which model came second. The single dollar figure, $2.4 million in annualized savings at one private equity firm, is attributed to Terminal X and reported by the customer.

The style-drift example is the strongest practical case. A buyout fund with a take-private mandate can find its deal flow sliding toward distressed credit when markets turn. An automated reviewer that flags those departures is a real governance tool, the kind of discipline investment committees often apply unevenly.

The longer-term trajectory deserves a harder look. Terminal X's next step is to extend firm Memories to individual investors, capturing each professional's standard as it is practiced. That's powerful for firms and uneasy for people. If a firm owns its judgment as data, the leverage of the senior investor who built that judgment shrinks. Codified judgment can also harden into codified blind spots, and a system that learns from analyst corrections will learn the firm's mistakes as faithfully as its insights. I'm watching whether the people whose expertise gets encoded share in its value.

Terminal X uses GPT-6.1 Sol for judgment-heavy, long-context stages such as synthesizing a thesis and assessing mandate fit, and GPT-6 Luna for high-volume steps. Its agents work in a private data room and handle documents in English, Korean, Japanese and Chinese. The company says it has built 1,700 firm Memories and run more than 1 million custom workflows. Its co-founder and CEO is Hyun Hong; co-founder and CTO is Ken Lee.