← Back to feed

Terminal X is a startup whose forward-deployed engineers — former JPMorgan and Blackstone professionals — spend weeks on-site with investment firms reverse-engineering how their best people actually make decisions. Then they encode that judgment into agents. The problem they are solving has a name: memory debt, the institutional reasoning that disappears when experienced people leave.

OpenAI published a case study on October 20 detailing results from Terminal X’s Capital Memory operating system running on GPT-6.1 Sol and GPT-6 Luna.

Evaluation Numbers

Terminal X ran 500 scored investment judgments through its Eval Agent — a workflow on GPT-6.1 Sol designed to check every agent output against the firm’s standards the way a VP checks an analyst’s work. Each output is scored against a 54-criterion rubric.

Across those 500 judgments, the Eval Agent on GPT-6.1 Sol agreed with the senior human reviewer 88.6% of the time. The next-best model scored 86.4%, a gap of 220 basis points. GPT-6.1 Sol also caught about 25% more cases where an agent’s recommendation had drifted from the fund’s mandate — the style-drift problem, where market conditions push an agent toward a strategy the fund is not authorized to pursue.

The same workflows ran at roughly 30% lower cost per query on GPT-6.1 Sol than on the next-best model, at comparable speed. One private equity customer reported $2.4 million in annualized savings from automated due diligence and investment memo workflows.

The Architecture

Terminal X uses GPT-6.1 Sol for the judgment-heavy stages: synthesizing investment theses, identifying key considerations, assessing fit with a fund’s mandate. GPT-6 Luna handles high-volume steps where token efficiency matters more than depth.

Each Eval Agent failure — a missed source, a wrong comparable, a thesis that drifts from the house view — becomes a Memory: a new workflow instruction or test case for that desk. Every fail includes a one-line reason and a corrected text or better source. The system gets more accurate the longer a firm uses it.

As of the case study date, Terminal X has built 1,700 firm Memories across its customer base, run more than 1 million custom workflows, and reached a weekly active user count equal to 81% of monthly active users — high retention for an enterprise tool.

Why This Architecture Favors GPT-6.1 Sol

GPT-6.1 Sol’s long-context handling is the differentiating factor for the Eval Agent’s deeper review workflow. Terminal X’s process tests a thesis first against the fund’s mandate, then against its recent deals. Uncertain verdicts are re-run, and the agent reconciles the results before reporting. A single-pass architecture on a shorter-context model would flag fewer deviations — which is where the 25% improvement in mandate drift detection comes from.

The pricing also matters. GPT-6.1 Sol was launched at $2 per million input tokens and $10 per million output tokens, roughly one-fifth the price of GPT-6 Astra, which is why the 30% lower cost figure reflects a real advantage rather than a minor increment.

Terminal X’s next step is extending firm Memories from desks to individuals — capturing how each investment professional makes decisions, not just how a desk collectively operates.