Terminal X says its Eval Agent agreed with senior human reviewers 88.6% of the time on GPT-6.1 Sol, compared with 86.4% for the next-best model across 500 scored judgments, per OpenAI News. As of Oct. 20, no major prediction venue including Polymarket or Kalshi lists odds on this benchmark, so current odds are unquoted, volume is zero and no market moved on the result. The comparison rests on an internal rubric and does not disclose fund types or false-positive rates. Not financial advice. Markets move fast.
The 220 basis point gap looks small until you price it.
Terminal X reports the same Eval Agent on GPT-6.1 Sol caught about 25% more recommendations that strayed from a fund mandate and ran the same workflows at about 30% lower cost per query than the next-best model, per the company. The check is a 54-criterion rubric that scores numbers, comps, consensus figures, quotes, narrative and house-view fit as pass or fail, with each fail stored as a Memory. One private equity customer reported $2.4 million in annualized savings from automated due diligence and memo workflows, according to Terminal X. Hyun Hong, co-founder and CEO of Terminal X, said GPT-6.1 Sol's long-context reasoning and token efficiency let the firm build deeper evaluation pipelines and make judgment measurable and auditable. Ken Lee, co-founder and CTO, said every answer gets evaluated and corrections teach the system how that firm works.
The model choice maps to how OpenAI positioned the GPT-6 pair. OpenAI introduced GPT-6 Sol and Luna as frontier models with different balances of capability and cost, and its Help Center lists GPT-6 Luna as the recommended starting point for migrating existing GPT-4.x use cases. Terminal X said it routes judgment-heavy, long-context steps like thesis synthesis, key considerations and mandate fit to GPT-6.1 Sol, while GPT-6 Luna handles high-volume steps where token efficiency matters more. That deeper review on Sol re-tests theses against mandate and recent deals and reconciles uncertain verdicts before reporting, which the company said flagged deals a single pass had waved through. Agents work inside each firm's private data room and process English, Korean, Japanese and Chinese, with Memories versioned by desk.
What the numbers do not show matters. Terminal X has built 1,700 firm Memories, run more than 1 million custom workflows and reports weekly active users equal to 81% of monthly active users as of the October report, but it does not break out accuracy by asset class, language or deal size and does not publish an independent audit of the 500-judgment sample. Cost per query is cited as about 30% lower at similar speed, without a disclosed baseline. One sentence would change the read: a third-party replication with false-positive and false-negative rates by mandate type would tell buyers if the gains hold outside Terminal X's rubric.
