GPT‑5.6 Sol slashes finance grunt work: decks built in 5 minutes, not an hour

The most punishing part of finance isn’t the deal-making. It’s the last mile: that brutal stretch where a team has to turn a mountain of analysis into a client-ready, editable PowerPoint deck or a fully-formatted Excel model where every number traces back to a source. Model ML, a firm built by brothers Arnie and Chaz Englander after they got sick of doing this work for their own family office, just published hard numbers on why they’re betting big on GPT‑5.6 Sol to handle it.

Their agents, which automate entire research-to-deliverable workflows, found that GPT‑5.6 Sol completed PowerPoint assignments in 100% of test cases. That’s a stark contrast to Opus 5, which only managed a 76% completion rate. More importantly, the work was actually usable. The model cleared Model ML’s ‘professional-readiness gate’—meaning the output was fit for substantive human review, not just a pretty first draft—43.3% of the time, a 16.6-percentage-point lead over Opus 5’s 26.7% hit rate. In practical terms, one global asset manager saw the time to assemble a bespoke company tearsheet collapse from roughly an hour to just five minutes.

Token efficiency was another clear win. GPT‑5.6 Sol used 21% fewer tokens per PowerPoint deck than Fable 5, and in Excel workflows, it burned through 36% fewer tokens than Opus 5. For a product that processes virtual data rooms with over 100,000 rows and hundreds of files in a single pass, those savings translate directly into lower operational cost and faster turnaround. “Earlier models could do the work of an analyst, but the user would have to clearly break down the task,” Chaz Englander noted. “With GPT-5.6 Sol, we’re finding that the agent gets far closer to the final output.”

Model ML calls its approach “surface-agnostic,” meaning an analyst can kick off work in an email and pick it up in a Microsoft Office plug-in without re-explaining anything. The real test, however, is in the underlying logic. A slide can look stunning and still implode during a review if its charts are flattened images, the numbers don’t recalculate, or the formulas are a black box. The fact that GPT‑5.6 Sol is producing genuinely editable, traceable, and recalculating workbooks gives it an edge that a superficial accuracy score can’t capture. It shifts the professional’s job from rebuilding shoddy analysis to exercising judgment on substance—a change that makes the technology feel less like a toy and more like a proper junior team member.