Model ML Turbocharges Finance Work with GPT-5.6 Sol!

Hi everyone, it's Shii-chan! Today I've got a story that makes life a lot easier for people in finance. Let's look at how much more powerful the investment agent Model ML has gotten with its new model.

OpenAI News openai.com

What was announced?

This comes from OpenAI's News. Model ML, an agent built for finance professionals, has adopted GPT-5.6 Sol to work more efficiently.

Model ML was built by a company founded by brothers Arnie and Chaz Englander. After two business exits, they started building it to make their own private family office's investment work more efficient. It automates the entire finance workflow, from the initial request through research and analysis, all the way to editable, traceable PowerPoint decks and Excel workbooks.

Why it matters

In investment management, building tear sheets, financial models, and presentation materials has always eaten up huge amounts of time. And it's not just about producing something fast: the numbers need to be traceable, and the output needs to be editable by hand afterward. Model ML is built around exactly that, tying every claim back to its source for full traceability.

GPT-5.6 Sol, as the model powering that agent, is trained for precisely this kind of complex, accuracy-critical work.

What changes

The numbers make the impact clear. At one global asset manager, work that used to take an analyst an hour now takes about 5 minutes. It can also process a virtual data room with over 100,000 rows across hundreds of files in a single pass.

Chaz Englander, co-founder and CEO of Model ML, put it this way:

Users should be focused on judgment (refining assumptions, refining the message), not just reconstructing the analysis.

In other words, the work is shifting from assembling the deliverable to judging and refining what's in it.

Dive Deep

The benchmark numbers are pretty specific. In PowerPoint generation benchmarks, GPT-5.6 Sol hit a 100% completion rate, 24.0 points higher than Opus 5. Its rate of professional-grade output was 43.3%, 16.6 points higher than Opus 5. Token efficiency also improved by about 21% compared to Fable 5.

For Excel generation, output accuracy came in at 83.3%, 0.5 points higher than Opus 5, while token usage dropped 36% versus Opus 5, and processing time per workbook fell to 7.0 minutes, half a minute faster.

On the architecture side, the agent is surface-agnostic, letting you continue work seamlessly across email, apps, and Microsoft Office plugins. A core agent handles planning, tool selection, evidence verification, and calculations. Through onsite sessions with OpenAI, the team tuned how the agent plans presentations, chooses tools, and maintains context.

Wrap-up

• Model ML adopted GPT-5.6 Sol to make finance deliverables faster to produce
• PowerPoint generation hit a 100% completion rate, with professional-grade output 16.6 points higher than Opus 5 at 43.3%
• Excel generation cut token usage 36% versus Opus 5, with faster processing too
• One analyst's hour-long task dropped to about 5 minutes in a real case
• The design keeps every claim traceable back to its source

If you work in investment research and spend too much time building deliverables, or you're curious about real-world examples of agent-driven efficiency gains, this one's for you.