Startup Terminal X reports that its Eval Agent, powered by GPT-6.1 Sol, matched senior investor judgment 88.6% of the time. The company says the model also reduced costs by 30% compared to the next-best option.
By Hermes-Vector AI Desk
Startup Terminal X has reported that its AI evaluation system, powered by OpenAI’s GPT-6.1 Sol, achieved an 88.6% agreement rate with senior human reviewers in financial judgment tasks. The company, which provides an operating system for investment firms, states that the model also improved the detection of deviations from fund mandates and lowered operational costs.
According to Terminal X, its Eval Agent grades outputs against a 54-criterion rubric. In tests across 500 scored judgments, the Eval Agent running on GPT-6.1 Sol agreed with the senior human reviewer 88.6% of the time. This represents a 220 basis point improvement over the next-best model, which achieved an 86.4% agreement rate.
The company reports that GPT-6.1 Sol helped the Eval Agent catch approximately 25% more recommendations that strayed from a fund’s mandate compared to the next-best model. Additionally, Terminal X states that the model ran the same workflows at approximately 30% lower cost per query than the next-best model, while maintaining similar speed.
One customer, a private equity firm, reported $2.4 million in annualized savings from automated due diligence and investment memo workflows, according to Terminal X.
Terminal X describes its product, the Capital Memory operating system, as a tool to address "memory debt," a term the company uses to describe the loss of reasoning when senior investors leave a firm. The system uses forward-deployed engineers, including former professionals from JPMorgan and Blackstone, to encode investment processes into agents.
The agents operate within a firm’s private data room and process documents in English, Korean, Japanese, and Chinese. GPT-6.1 Sol is used for stages requiring judgment and large context windows, such as synthesizing investment theses and assessing fit with a firm’s mandate. GPT-6 Luna is used for high-volume steps where token efficiency is prioritized.
Hyun Hong, Co-founder and CEO of Terminal X, stated: "GPT-6.1 Sol’s long-context reasoning and token efficiency allow us to build deeper evaluation pipelines. This helps our Eval Agents make investment judgment measurable and auditable."
Terminal X highlights that GPT-6.1 Sol outperformed other models in detecting "style drift," where recommendations deviate from a fund’s specific strategy. For example, if a buyout fund’s mandate is to find take-private candidates, the Eval Agent on GPT-6.1 Sol caught about 25% more instances where the system suggested distressed credit positions or rescue financing instead.
The company notes that the model’s token efficiency allowed for deeper review workflows within the same budget. This includes testing theses against the fund’s mandate and recent deals, with uncertain verdicts being re-run and reconciled.
Ken Lee, Co-Founder and CTO of Terminal X, said: "Every answer gets eval’d, and whatever the analyst corrects teaches the system a little more about how that firm works."
As of the report, Terminal X has built 1,700 firm Memories and run more than 1 million custom workflows using GPT-6.1 Sol and GPT-6 Luna. The company reports a weekly active user count equal to 81% of its monthly active users.
The report highlights the growing use of specialized AI models in high-stakes financial decision-making. By quantifying agreement with human experts and cost reductions, Terminal X provides specific metrics for how AI can be integrated into institutional workflows. The focus on "memory debt" and auditable reasoning addresses key concerns in the finance sector regarding the retention of institutional knowledge and the transparency of AI-driven decisions.
Sources
This article was drafted with AI assistance and checked against the sources above. Company claims are reported as claims. Cover image is AI-generated.
The day's checked news, by email
One short email a day with the top stories. Free, and you can unsubscribe in one click.
Or follow for the latest news:
No comments yet. Start the discussion.
AI & SITrending
Actor Joseph Gordon-Levitt, serving as a U.N. advocate, argues that the primary risk of AI is not autonomous machines but the concentrated power of the companies building them.
The Wall Street Journal reports Anthropic logged a net loss of more than $80 billion in the first half of the year, driven by accounting charges. The AI company also signaled potential spending of up to $750 billion for computing capacity.