Model ML has integrated GPT‑5.6 Sol into its agent-based finance workflows to produce native, editable PowerPoint decks and Excel workbooks with traceable sources. In the company’s Composite benchmark the model completed PowerPoint workflows in 100% of test cases and reduced token usage compared with competing models.
The problem and the solution
Preparing financial analysis for clients or senior decision‑makers requires a labor‑intensive final stage: reconciling evidence, formatting files, checking every number, and linking claims to sources so the finished PowerPoint or Excel file is editable and reviewable. Arnie and Chaz Englander, cofounders of Model ML, built software to automate these steps after encountering the workload themselves while investing through a private family office.
From that software, Model ML’s agents now carry an assignment from a brief through research and modeling to a completed, editable deck or workbook. A central agent plans the work, selects tools, reconciles evidence, runs calculations, and routes each step to the model best suited for it — frequently GPT‑5.6 Sol. Model ML’s document tooling produces native PowerPoint and Excel files whose charts, tables, and formulas remain editable and whose numbers are traceable back to sources.
How the platform works in practice
The product is described as "surface‑agnostic": a finance professional can start a task in email or the Model ML app and continue it in Microsoft Office plug‑ins without re‑explaining the assignment. The system preserves the original brief in context and visually reviews every slide before returning the file.
For presentations, Model ML can turn brief and source material into an editable PowerPoint suitable for an investment committee. For Excel tasks, the agent can begin from a client template or blank workbook, gather data, construct formulas and logic across tabs, and apply finance‑specific formatting to produce a complete spreadsheet or model.
Performance and benchmark results
Model ML’s Composite benchmark follows a finance assignment end to end, then checks numbers, sources, formulas, structure, and visual quality. Key findings include:
- GPT‑5.6 Sol completed PowerPoint workflows in 100% of test cases; Opus 5 completed them in 76% of cases.
- Under Model ML’s professional‑readiness metric, GPT‑5.6 Sol produced outputs ready for substantive review in 43.3% of cases, compared with 26.7% for Opus 5.
- GPT‑5.6 Sol used about 21% fewer tokens per PowerPoint deck than Fable 5.
- In Excel workflows, GPT‑5.6 Sol used 36% fewer tokens per workbook than Opus 5.
Practical impacts observed: at one global asset manager a bespoke tearsheet that previously took an analyst about one hour to assemble now takes roughly five minutes. In another use case, Model ML agents processed virtual data rooms containing more than 100,000 rows and hundreds of files in a single pass.
Technical approach and integration
The agent runs inside a harness and is equipped with toolkits for data integrations, document editing, and code execution environments. The system loads the appropriate toolkit when needed so the agent stays focused and has exactly the tools required to work with the user’s documents.
Model ML reached this setup after on‑site sessions with OpenAI, during which teams traced how the agent planned presentations, chose tools, and maintained context. Those findings were used to refine agent instructions and determine when each toolkit should load.
Security, interactivity and the future of outputs
Model ML’s platform can produce secure, interactive outputs that either update continuously or are frozen to a particular moment. A reviewer can open an investment summary, click through to the financial model behind a figure, and continue working with the agent from the same page.
As Arnie Englander put it: “Ready for real work means the user can move directly into real review. The numbers trace back, the workbook recalculates, the slide is editable.” He argues that PowerPoint, Excel, and Word were designed for a manual knowledge‑work world, and that AI is changing that assumption — and the software itself — fundamentally.
Conclusion
Based on Composite results, Model ML has expanded GPT‑5.6 Sol in production, including workflows formerly handled by Opus 4.8. According to the company’s internal evaluations, GPT‑5.6 Sol delivers higher completion rates, improved professional readiness, and lower token consumption for producing editable, source‑traceable financial presentations and workbooks.
