Terminal X Says GPT-6.1 Sol Matched Senior Investor Reviews 88.6% of the Time

OpenAI published the Terminal X evaluation on October 20

OpenAI published a customer story on October 20, 2026 describing how Terminal X uses GPT-6.1 Sol and GPT-6 Luna in customized investment agents. The central new result is narrow and concrete: across 500 scored judgments, Terminal X says an Eval Agent using GPT-6.1 Sol agreed with a senior human reviewer 88.6% of the time. The unnamed next-best model reached 86.4%, making the reported difference 2.2 percentage points, or 220 basis points.

This is not a new model launch or an independent benchmark. GPT-6.1 Sol has been available since September under the API identifier gpt-6.1-sol. The October story adds customer-reported production evidence about one evaluation workflow. Terminal X is available through a book-a-demo sales motion and says its customer-specific agents can be deployed on-premise or in a customer cloud; neither reviewed company page publishes a self-serve Terminal X price.

The agent grades outputs against 54 firm-specific criteria

Terminal X describes an orchestrated Eval Agent that reviews every agent answer against 54 pass-or-fail criteria. The checks cover whether figures, comparables, consensus data and quotations came from appropriate sources; whether the narrative is coherent; and whether the recommendation fits the firm’s mandate and house view. A failed check includes a short reason plus a proposed source or corrected text.

The surrounding system is customized rather than generic. Terminal X says its staff reconstruct how a customer’s investment teams screen deals and write memos, encode those processes into agents and sub-agents, and retain analyses, committee decisions, instructions and house views in a versioned “Capital Memory.” That context makes the result more relevant to bespoke investment review and less portable to an off-the-shelf financial chatbot.

The relative gains are promising but underspecified

Terminal X says GPT-6.1 Sol caught about 25% more recommendations that drifted outside a fund’s mandate than the next-best model. It also reports about 30% lower cost per query at similar speed. The story attributes the cost result to Sol’s token efficiency and a deeper review sequence: test the thesis against the mandate, compare it with recent deals, rerun uncertain verdicts and reconcile results.

Those percentages cannot be turned into absolute error or dollar rates from the published material. OpenAI and Terminal X do not identify the comparator, count mandate-drift cases, publish the confusion matrix, disclose how many senior reviewers labeled the cases, state whether labels were blinded, report repeated-run variance or provide token totals. “25% more” could describe different absolute gains depending on prevalence; “30% lower cost” depends on model, prompt, reasoning effort, cache behavior, price band and retry policy.

The 88.6% figure is agreement, not investment accuracy

The measured outcome is agreement with a senior reviewer across scored judgments. It does not establish that 88.6% of investment recommendations were correct, that returns improved, or that the system predicts markets. A reviewer can be internally consistent without being objectively right, and agreement can hide different error costs. Missing a prohibited strategy, approving a fabricated quotation and disagreeing on writing style should not be averaged into one deployment threshold without severity weights.

The article separately reports company-level operating figures: 1,700 firm Memories, more than 1 million custom workflows, an 81% weekly-to-monthly active-user ratio, and $2.4 million in annualized savings for one private-equity customer. These are Terminal X claims carried in an OpenAI customer story. No customer identity, calculation period, counterfactual labor baseline, realized investment outcome or independent audit is provided.

Current API pricing does not prove the reported query cost

OpenAI’s current Standard short-context prices for GPT-6.1 Sol are $2 per million input tokens, $0.10 cached input, $2.50 cache writes and $10 output. The model page documents a 1,050,000-token context window and 128,000 maximum output, while requests above 272,000 input tokens use higher rates for the full request. Batch, Flex, Fast mode and regional processing can also change the bill.

The customer story does not publish prompts, token classes, tool charges, reasoning settings, retries or the comparison model’s rates. Buyers therefore cannot reproduce the 30% figure from list prices. Measure cost per accepted review on the complete workflow, including retrieval, evaluation passes, human corrections and failed runs, rather than comparing only headline token rates.

Use the pattern, then rebuild the evidence on your cases

A useful pilot starts with a frozen set of real, permission-cleared historical cases and a rubric written before model selection. Name every comparator, blind reviewers where practical, preserve source documents and traces, separate factual checks from mandate and style checks, and report counts as well as percentages. Evaluate held-out cases after every change to model, memory, retrieval or instructions so the system is not merely learning the test set.

Trial the pattern for analyst-assistance and quality-control tasks where every claim remains source-linked and a qualified person reviews the output. Constrain authority, data access and retention. Wait if the vendor cannot provide task-level error evidence, reproducible cost logs or deployment controls. Reject autonomous investment authority: the published result supports a bounded evaluator experiment, not delegation of fiduciary judgment.

Disclosures

AccessAllGPT received no OpenAI or Terminal X account, demo, API credit, customer data, benchmark files, early access, briefing, review, payment or compensation for this article. OpenAI and Terminal X did not sponsor, review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with OpenAI, Terminal X or any investment firm. Publication-wide relationships are listed on the disclosures page.