Terminal X’s result is a workflow-evaluation claim, not proof that GPT-6.1 predicts profitable investments. In an OpenAI customer story published October 20, the company reports that its Eval Agent agreed with a senior human reviewer on 88.6% of 500 scored judgments. The alert reached our monitor October 21; this is not a new model launch.
What Terminal X actually measured
Terminal X builds investment agents around each firm’s documents, committee decisions, workflows and investment mandate. GPT-6.1 Sol handles judgment-heavy work; GPT-6 Luna handles high-volume steps. Its Eval Agent checks outputs against a 54-criterion pass/fail rubric covering sourcing, narrative and alignment with the firm’s house view.
| Measure | Reported result | Evidence boundary |
|---|---|---|
| Senior-reviewer agreement | 88.6% versus 86.4% | 500 scored judgments; comparator unnamed |
| Agreement improvement | 2.2 percentage points | 220 basis points, not 2.2% relative |
| Mandate-drift detection | About 25% more departures caught | Customer-reported, not a general accuracy score |
| Per-query spending | Approximately 30% lower | Same workflows, unnamed next-best model, similar speed |
If both agreement percentages cover the same 500 judgments, they correspond to 443 versus 432 agreements: 11 additional matches. The public story does not provide paired results, confidence intervals or a released evaluation dataset, so it cannot establish statistical significance.
The rubric measures alignment with a firm’s standard—not future returns, universal investment quality or freedom from factual errors. The 54 checks should not be confused with 54 independent test cases for each of the 500 judgments.
One example is detecting recommendations that drift from a buyout fund’s mandate into distressed-credit or rescue-financing strategies. That is a concrete control problem, not a stock-picking benchmark. We have not independently reproduced these results.
What the cost claim means
OpenAI’s story reports that Sol’s token efficiency allowed deeper reviews within the same budget: checking the thesis against the mandate and recent deals, rerunning uncertain verdicts, then reconciling them. The reported saving is a completed-workflow comparison, not a change in the published token rate.
The source does not identify the competing model, disclose absolute query costs or publish a full token-and-tool ledger. It therefore cannot support a claim that Sol is 30% cheaper than Luna, every rival, or your current stack.
The chart, calculator and table above render from our maintained pricing dataset, with recorded source dates. We separately checked both official model pages for this report. These are developer API baselines, not a Terminal X quote; cached input, cache writes, long context, service modes, tools and regional processing can change the bill. The calculator does not price deployment, data ingestion, custom engineering or human review.
The announcement publishes no new Terminal X subscription price, OpenAI rate cut or ChatGPT plan entitlement. Terminal X’s website invites firms to contact the company. Request an itemized platform and implementation quote before budgeting from API tokens alone.
What investment and agent teams should do now
Potential beneficiaries: firms with repeated diligence and memo workflows, explicit mandates and labeled review history. Less directly served: teams without a stable rubric, or buyers expecting a turnkey price or validated return forecast from this announcement.
- Replay your own cases. Include clean work, weak sourcing and subtle mandate drift. Hold the task, retrieved evidence and scoring rules constant across models.
- Measure accepted-result cost. Include all reasoning calls, retries, tools, platform fees and reviewer time. Log false approvals separately from harmless disagreements.
- Test routing, not just one model. Compare Luna on repeatable extraction with Sol on judgment-heavy stages; keep human approval for consequential decisions. The story does not establish Luna’s agreement score.
- Check deployment details. Confirm model endpoints, data flows and contractual terms. Running an orchestration platform on-premise does not itself establish that OpenAI model inference runs locally.
Use the OpenAI pricing page, Anthropic pricing page and token calculator to scope a controlled comparison. Our API cost guide explains how to budget beyond headline token rates.
For an optional open-model evaluation control, compare supported Novita inference endpoints on the same rubric; no Terminal X integration or capability parity is implied.
Affiliate disclosure: We may earn a commission from the sponsored Novita link at no extra cost to you.
Sources checked October 21, 2026
OpenAI’s Terminal X case study, Terminal X, GPT-6.1 Sol documentation and GPT-6 Luna documentation. Benchmark and savings figures are attributed to Terminal X through OpenAI’s customer story.