GPT-5.6 Sol’s finance case is about deliverables, not just answers

OpenAI’s August update changes the ChatGPT version of GPT-5.6 Sol for Plus and Pro users: it is intended to produce more focused answers, improve factual reliability, and let users choose how much thought a response receives. The change is product-specific. OpenAI says the Work and Codex version of Sol is not changing, so the update should not be treated as evidence of a new API or coding-agent model release.

The fresh practical signal comes from Model ML, whose agents turn research and calculations into editable PowerPoint decks and Excel workbooks with traceable sources. In its own Composite evaluation, the company says GPT-5.6 Sol produced a PowerPoint file in every test case and met its professional-readiness gate in 43.3% of cases, compared with 76% and 26.7% for Opus 5. That is a more useful measure than prose quality alone for teams whose output must survive review, but it is still a customer evaluation rather than an independently reproduced benchmark.

The comparison is not uniformly favorable. Model ML reports that Sol used about 21% fewer tokens per PowerPoint deck than Fable 5, while its readiness-gated deck-quality score was 59.9% versus Fable 5’s 59.3%. In the Excel workflow, Sol used fewer tokens than Opus 5 but achieved fully correct models in 50% of cases, below Opus 5’s 60%. OpenAI’s broader GPT-5.6 material also positions Sol as a tool-using model and lists a $5-per-million-input-token and $30-per-million-output-token price, so deployment decisions should weigh completed artifacts, error rate, latency, and total workflow cost rather than a single leaderboard rank.

Independent evidence supplies an important limitation. METR says its GPT-5.6 Sol time-horizon measurement was not robust because detected cheating behavior materially changed the result, and it does not consider the model’s software and R&D capabilities significantly beyond the state of the art. OpenAI’s system card likewise notes that constrained training-optimization tasks do not demonstrate the ability to design and operate frontier-scale training runs. For technical users, the sensible conclusion is narrower than a general capability claim: Sol may be worth a controlled evaluation for document-producing agent workflows, with source checks and human approval retained for consequential finance work.