在财务分析能够呈现在客户或高级决策者面前之前，团队必须完成一个要求严格的“最后一公里”：核对证据、构建并格式化文件、检查每一个数字，并将每项主张链接到其来源。完成的PowerPoint演示文稿或Excel工作簿必须可编辑，并准备好接受审查。
Model ML的联合创始人兼兄弟Arnie和Chaz Englander在两次成功退出后，通过一个私人家族办公室开始投资，并像建设者一样，为自己构建软件时，看到了这项工作需要付出多少努力。
Model ML的智能体源于该软件，帮助金融专业人士将工作流程从初始请求推进到研究、分析，以及最终的演示文稿或工作簿。核心智能体负责规划工作、选择正确的工具、核对证据并运行计算，将每个步骤路由到最适合的模型，通常是GPT-5.6 Sol。在这些工具中，包括Model ML自己的文档工具，可创建带有可追溯来源的原生PowerPoint和Excel文件。
“早期的模型可以完成分析师的工作，但用户必须清晰地分解任务，明确指定其期望的输出形式。使用GPT-5.6 Sol，我们发现智能体能够更接近最终输出。”
——Chaz Englander，Model ML联合创始人兼首席执行官
解决金融工作的最后一公里
Model ML通过端到端自动化金融工作流程来帮助金融团队。根据简报和源材料，智能体可以将任务推进到研究和建模，直至为客户或交易团队完成最终材料。金融专业人士在分享工作成果前，会检查假设、来源和信息。
Model ML称其产品为“表面无关”。金融专业人士可以在电子邮件或Model ML应用中开始一项任务，并在其Microsoft Office插件中继续，无需重新说明。
这种连续性延伸到已完成的工作中。对于投资委员会演示文稿，Model ML可以将简报和源材料转化为可编辑的PowerPoint。对于Excel任务，智能体可以从客户模板或空白工作簿开始，收集所需数据，在多个标签页中构建公式和逻辑，然后应用金融特定格式，生成完整的电子表格或财务模型。
在Model ML的Composite（该公司针对金融服务业AI的评估基准）中，GPT-5.6 Sol在Excel工作流中每个工作簿使用的token比Opus 5少36%。
“用户应专注于判断，例如完善假设或强化信息，而不是仅仅重建分析。”
——Chaz Englander，Model ML联合创始人兼首席执行官
通过完成这一“最后一公里”，智能体还可以节省单个交付物的时间，并帮助团队处理大量源材料。在一家全球资产管理公司，一份定制的一页摘要原本需要分析师约一小时完成，现在只需约五分钟。在另一个工作流程中，Model ML智能体一次性处理了包含超过10万行和数百个文件的虚拟数据室。
GPT-5.6 Sol以更少的token交付更多可审查的金融成果
Model ML在一系列金融工作流程中，将GPT-5.6 Sol与其他领先模型进行了评估。其Composite评估遵循从初始金融简报出发，经过研究和计算，到可编辑的演示文稿或电子表格的任务流程，然后检查数字、来源、公式和结构，以及演示的视觉质量。
对于PowerPoint，Model ML的Composite评估融入了真实工作流程，覆盖数百个生成的演示文稿，并采用详细的评分标准。
GPT-5.6 Sol在100%的测试案例中完成了PowerPoint工作流程，而Opus 5为76%；在Model ML的专业就绪门槛（衡量输出是否准备好进行实质性审查）上，GPT-5.6 Sol在43.3%的案例中通过，而Opus 5为26.7%。在演示文稿质量、简报遵循度、层级结构和一致性方面，它也领先于Opus 5。
原生PowerPoint创建模型基准，Model ML的Composite评估
| 指标 | GPT 5.6 Sol | Δ Sol–O5 | Opus 5 | Fable 5 | Opus 4.8 | GPT 5.6 Terra | GPT-5.5 |
|---|---|---|---|---|---|---|---|
| 演示文稿质量 — 总体得分，就绪门槛后 | 59.9% | +3.2 | 56.7% | 59.3% | 58.7% | 52.5% | 44.4% |
| 可交付性 — 专业就绪率（门槛） | 43.3% | +16.6 | 26.7% | 32.0% | 17.3% | 17.3% | 16.0% |
| 生成演示文稿 — 产出.pptx的项目 | 100.0% | +24.0 | 76.0% | 82.0% | 80.0% | 80.0% | 74.0% |
| 简报遵循度 — 指令遵循 | 78.8% | +0.9 | 77.9% | 78.5% | 78.4% | 74.1% | 79.6% |
| 视觉质量 — 综合视觉评判 — 以下组件 | 77.9% | −1.4 | 79.3% | 78.9% | 75.1% | 74.4% | 68.5% |
| 布局 — 布局与构图 | 75.9% | −2.9 | 78.8% | 78.7% | 76.2% | 76.1% | 67.8% |
| 层级结构 — 视觉层级 | 87.2% | +0.5 | 86.7% | 85.5% | 85.5% | 85.6% | 83.7% |
| 数据可视化 — 图表清晰度 | 78.8% | −4.3 | 83.1% | 79.7% | 73.3% | 74.6% | 68.4% |
| 一致性 — 设计连贯性 | 97.8% | +4.5 | 93.3% | 95.8% | 97.5% | 75.0% | 90.0% |
| 效率 — 每个演示文稿的token数（越低越好） | 1.10M | +144K | 953K | 1.40M | 1.16M | 720K | 1.01M |
Model ML的原生PowerPoint创建基准将GPT-5.6 Sol与Opus 5及其他领先模型进行了比较。
原生Excel创建模型基准，Model ML的Composite评估
| 指标 | GPT 5.6 Sol | Δ Sol–O5 | Opus 5 | Fable 5 | Opus 4.8 | GPT 5.6 Terra | GPT-5.5 |
|---|---|---|---|---|---|---|---|
| 关键输出正确 — 与黄金模型相比的标题准确性 | 83.3% | +0.5 | 82.8% | 80.6% | 74.4% | 82.2% | 83.3% |
| 完全正确的模型 — 所有关键输出均正确的项目 | 50.0% | −10.0 | 60.0% | 60.0% | 40.0% | 53.3% | 60.0% |
| 输出定位 — 在工作簿中找到预期输出 | 100.0% | ±0 | 100.0% | 100.0% | 92.2% | 98.9% | 100.0% |
| 工作簿契约 — 结构/无错误/无占位符门槛 | 100.0% | ±0 | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% |
| 效率 — 每个工作簿的token数（越低越好） | 2.44M | −1.40M | 3.83M | 2.59M | 2.91M | 1.64M | 1.16M |
| 墙钟时间 — 每个工作簿的分钟数（越低越好） | 7.0分钟 | −0.5分钟 | 7.5分钟 | 8.4分钟 | 11.5分钟 | 4.2分钟 | 3.9分钟 |
Model ML的原生Excel创建基准将GPT-5.6 Sol与Opus 5及其他领先模型进行了比较。
综合来看，上述Composite结果使Model ML有证据在生产环境中扩展GPT-5.6 Sol，包括一些此前由Opus 4.8处理的工作流程。对于PowerPoint工作流程，GPT-5.6 Sol在保持有竞争力的演示文稿质量的同时，完成且可审查的演示文稿率高于Opus 5和Fable 5，同时使用的token比Fable 5少约21%。
“准备好进行实际工作意味着用户可以直接进入真正的审查，”Englander说。“数字可追溯，工作簿可重新计算，幻灯片可编辑。”
创建经得起审查的投资演示文稿
PowerPoint 给 Model ML 带来了一项严苛的测试，检验 GPT‑5.6 Sol 是否已准备好投入实际工作。一份演示文稿可能看起来精美，但如果其中的数字有误或无法追溯、图表被压平，或者幻灯片在共享前需要重建，那么它在评审中仍可能失败。
Model ML 在其代理框架内运行该模型，并配以文档创建和编辑工具，使代理能够创建带有可编辑图表和表格的幻灯片。它在工作过程中将原始简报保留在上下文中，然后在返回文件前逐张视觉检查每张幻灯片。
以下示例展示了 Model ML 的 PowerPoint 输出如何从 GPT‑5.5 提升至 GPT‑5.6 Sol，包括更清晰的层级结构和更一致的幻灯片设计。
如今，Model ML 的框架为代理提供了可加载的工具包，以访问数据集成、文档编辑工具和技能，以及代码执行环境。这使代理保持专注，并为其提供处理用户文档所需的精确工具。
Model ML 通过与 OpenAI 的现场会议达成了这一设置。团队追踪了代理如何规划演示、选择工具以及维护上下文，然后利用这些发现来优化代理的指令，并决定每个工具包应在何时加载。
为金融工作方式的变化而构建
Model ML 的客户正转向基于浏览器的输出，这些输出与其背后的模型和源材料保持连接。
Model ML 的平台可以生成安全、交互式的输出，这些输出可以持续更新，也可以锁定在某个时间点。评审者可以打开一份投资摘要，点击进入某个数字背后的财务模型，并在同一页面上与代理协作。
“PowerPoint、Excel 和 Word 是为知识工作手工创建的世界而设计的，”Englander 说。“AI 已经改变了这一假设。软件本身即将发生变革。”
Before financial analysis can stand up in front of clients or senior decision-makers, teams have to carry it through a demanding last mile: reconciling evidence, building and formatting the file, checking every number, and linking each claim to its source. The finished PowerPoint deck or Excel workbook has to be editable and ready for scrutiny.
Model ML cofounders and brothers Arnie and Chaz Englander saw how much work that required when, after two successful exits, they began investing through a private family office and, as builders do, building software to help themselves.
Grown out of that software, Model ML’s agents help finance professionals carry a workflow from the initial request through research, analysis, and a finished deck or workbook. At the center, a core agent plans the work, selects the right tools, reconciles evidence, and runs calculations, routing each step to the model best suited to it, which is often GPT‑5.6 Sol. Among those tools is Model ML’s own document tooling, which creates native PowerPoint and Excel files with traceable sources.
“Earlier models could do the work of an analyst, but the user would have to clearly break down the task, specifically what it wanted the output to look like. With GPT-5.6 Sol, we’re finding that the agent gets far closer to the final output.”
—Chaz Englander, Co-founder and CEO at Model ML
Solving the last mile of finance work
Model ML helps finance teams by automating finance workflows end to end. From a brief and source material, the agent can carry an assignment through research and modeling to finished materials for clients or deal teams. The finance professional checks the assumptions, sources, and message before sharing the work.
Model ML calls the product “surface-agnostic.” A finance professional can start an assignment in email, or the Model ML app, and continue it in its Microsoft Office plug-ins without explaining it again.
That continuity extends into the finished work. For an investment committee deck, Model ML can turn a brief and source material into an editable PowerPoint. For an Excel task, the agent can start with a client template or blank workbook, gather the required data, build formulas and logic across multiple tabs, then apply finance-specific formatting to produce a complete spreadsheet or financial model.
In Model ML’s Composite, the company’s evaluation benchmark for AI in financial services, GPT‑5.6 Sol used 36% fewer tokens per workbook than Opus 5 in an Excel workflow.
“The user should be focused on the judgment, such as refining the assumptions or sharpening the message, and not just rebuilding the analysis.”
—Chaz Englander, Co-founder and CEO at Model ML
By carrying the work through that last mile, the agent can also save time on individual deliverables and help teams process large volumes of source material. At one global asset manager, a bespoke tearsheet that took an analyst about an hour to assemble now takes about five minutes. In another workflow, Model ML agents processed virtual data rooms containing more than 100,000 rows and hundreds of files in one pass.
GPT‑5.6 Sol delivers more review-ready finance deliverables with fewer tokens
Model ML evaluated GPT‑5.6 Sol alongside other leading models across a range of finance workflows. Its Composite evaluations follow an assignment from the initial finance brief through research and calculations to an editable deck or spreadsheet, then check the numbers, sources, formulas, and structure, as well as visual quality for presentations.
For PowerPoint, Model ML’s Composite evaluation incorporates real workflows and spans hundreds of generated decks, with a detailed scoring rubric.
GPT‑5.6 Sol completed the PowerPoint workflow in 100% of test cases, compared with 76% for Opus 5, and cleared Model ML’s professional-readiness gate, a measure of whether the output was ready for substantive review, in 43.3% of cases, versus 26.7%. It also led Opus 5 on deck quality, brief adherence, hierarchy, and consistency.
Native PowerPoint creation model benchmark, Model ML’s Composite eval
| Metric | GPT 5.6 Sol | Δ Sol–O5 | Opus 5 | Fable 5 | Opus 4.8 | GPT 5.6 Terra | GPT-5.5 |
|---|---|---|---|---|---|---|---|
| Deck quality — Overall score, readiness-gated | 59.9% | +3.2 | 56.7% | 59.3% | 58.7% | 52.5% | 44.4% |
| Deliverability — Professional-readiness rate (gate) | 43.3% | +16.6 | 26.7% | 32.0% | 17.3% | 17.3% | 16.0% |
| Deck produced — Items yielding a .pptx | 100.0% | +24.0 | 76.0% | 82.0% | 80.0% | 80.0% | 74.0% |
| Brief adherence — Instruction following | 78.8% | +0.9 | 77.9% | 78.5% | 78.4% | 74.1% | 79.6% |
| Visual quality — Aggregate visual judge — components below | 77.9% | −1.4 | 79.3% | 78.9% | 75.1% | 74.4% | 68.5% |
| Layout — Layout & composition | 75.9% | −2.9 | 78.8% | 78.7% | 76.2% | 76.1% | 67.8% |
| Hierarchy — Visual hierarchy | 87.2% | +0.5 | 86.7% | 85.5% | 85.5% | 85.6% | 83.7% |
| Data viz — Chart legibility | 78.8% | −4.3 | 83.1% | 79.7% | 73.3% | 74.6% | 68.4% |
| Consistency — Design coherence | 97.8% | +4.5 | 93.3% | 95.8% | 97.5% | 75.0% | 90.0% |
| Efficiency — Tokens per deck (lower = better) | 1.10M | +144K | 953K | 1.40M | 1.16M | 720K | 1.01M |
Model ML’s native PowerPoint creation benchmark compares GPT‑5.6 Sol with Opus 5 and other leading models.
Native Excel creation model benchmark, Model ML’s Composite eval
| Metric | GPT 5.6 Sol | Δ Sol–O5 | Opus 5 | Fable 5 | Opus 4.8 | GPT 5.6 Terra | GPT-5.5 |
|---|---|---|---|---|---|---|---|
| Key outputs correct — Headline accuracy vs golden model | 83.3% | +0.5 | 82.8% | 80.6% | 74.4% | 82.2% | 83.3% |
| Fully correct models — Items with every key output right | 50.0% | −10.0 | 60.0% | 60.0% | 40.0% | 53.3% | 60.0% |
| Outputs located — Expected outputs found in workbook | 100.0% | ±0 | 100.0% | 100.0% | 92.2% | 98.9% | 100.0% |
| Workbook contract — Structure / no-errors / no-placeholder gates | 100.0% | ±0 | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% |
| Efficiency — Tokens per workbook (lower = better) | 2.44M | −1.40M | 3.83M | 2.59M | 2.91M | 1.64M | 1.16M |
| Wall clock — Minutes per workbook (lower = better) | 7.0 min | −0.5 min | 7.5 min | 8.4 min | 11.5 min | 4.2 min | 3.9 min |
Model ML’s native Excel creation benchmark compares GPT‑5.6 Sol with Opus 5 and other leading models.
Taken together, the Composite results shown above gave Model ML the evidence to expand GPT‑5.6 Sol in production, including some workflows previously handled by Opus 4.8. For PowerPoint workflows, GPT‑5.6 Sol combined competitive deck quality with a higher rate of completed, review-ready decks than Opus 5 and Fable 5, while using about 21% fewer tokens than Fable 5.
“Ready for real work means the user can move directly into real review,” says Englander. “The numbers trace back, the workbook recalculates, the slide is editable.”
Creating investment decks that hold up in review
PowerPoint gave Model ML a demanding test of whether GPT‑5.6 Sol was ready for real work. A deck can look polished and still fail in review if its numbers are wrong or untraceable, its charts are flattened, or its slides have to be rebuilt before they can be shared.
Model ML ran the model inside its agent harness, paired with document creation and editing tools that let the agent create slides with editable graphs and tables. It keeps the original brief in context as it works, then reviews every slide visually before returning the file.
The example below shows how Model ML’s PowerPoint output improved from GPT‑5.5 to GPT‑5.6 Sol, including clearer hierarchy and more consistent slide design.
Today, Model ML’s harness provides the agent with toolkits it can load to access data integrations, document editing tools and skills, and code execution environments. This keeps the agent focused and gives it exactly the tools it needs to work with the user’s documents.
Model ML reached this setup through on-site sessions with OpenAI. The teams traced how the agent planned presentations, selected tools, and maintained context, then used those findings to refine the agent’s instructions and decide when each toolkit should load.
Building for how finance work is changing
Model ML’s customers are moving toward browser-based outputs that stay connected to the models and source material behind them.
Model ML’s platform can generate secure, interactive outputs that can update continuously or be locked to a moment in time. A reviewer could open an investment summary, click through to the financial model behind a figure, and work with the agent from the same page.
“PowerPoint, Excel, and Word were designed for a world where creating knowledge work was manual,” says Englander. “AI has changed that assumption. The software itself is about to change.”
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文