Astra’s Financial Workbook Leaves Its $32 Million Conclusion Unreconciled
OpenAI is positioning GPT-6.2 Astra for complex professional projects and GPT-6.2 Sol for everyday work, arguing that both deliver stronger results with fewer output tokens. Its examples show Astra updating a payables schedule and drafting a meeting slide, but the financial workbook’s displayed figures do not explain how they support the spoken conclusion. The launch announcement presents Sol as the more efficient, versatile option; the article’s examples focus on Astra rather than comparing the models side by side.

The spoken conclusion and workbook show different figures, but their relationship is unexplained
For the Breckenmeyer integration, the spoken conclusion is that a one-year delay in savings would mean owing $32 million more in 2031 and earning 30 cents less per share the following year. The workbook shown alongside that conclusion presents a plan-versus-base comparison: its 2031 row lists plan net debt of $7.1 million and base net debt of $2.1 million, a displayed difference of $(5.0) million. Across 2025–2031, base net debt is below plan in every year. The table gives total net debt of $127.6 million for plan and $111.5 million for base.
| Year | Plan net debt | Base net debt | Difference |
|---|---|---|---|
| 2025 | $32.0M | $31.8M | $(0.2)M |
| 2026 | $25.2M | $24.5M | $(0.7)M |
| 2027 | $21.3M | $20.4M | $(0.9)M |
| 2028 | $17.5M | $15.2M | $(2.3)M |
| 2029 | $14.2M | $11.1M | $(3.1)M |
| 2030 | $10.3M | $6.4M | $(3.9)M |
| 2031 | $7.1M | $2.1M | $(5.0)M |
| Total | $127.6M | $111.5M | $(16.1)M |
The task was to build a workbook for a one-year delay in savings, track net debt through 2031, and show next year’s EPS against plan. The spoken summary supplies the EPS result—30 cents less per share—but the visible table shows net debt figures, not an EPS comparison. The source does not explain how the spoken $32 million figure relates to the table’s $5 million plan-versus-base difference, or whether they describe the same measure or scenario. It therefore leaves the relationship between the spoken conclusion and the displayed figures unresolved.
That limitation matters because the workbook is the example of how GPT-6.2 Astra handles a consequential financial request. The display provides a multiyear comparison and a chart, but does not make the underlying relationship clear enough to reconcile the spoken summary with the table.
Astra turns workplace materials into drafts, while leaving review in the workflow
OpenAI describes GPT-6.2 Astra as its frontier model for demanding projects and says its slides, spreadsheets, and documents follow supplied templates and professional conventions more closely, with outputs that read less like AI-generated work. It also says Astra writes cleaner code and builds better-looking, better-working frontends, particularly for websites, gaming, and 3D design.†
The workplace examples show Astra drawing on company materials to produce or update work products. Asked to update Friday’s payables using invoices in email, Astra reports that the November 13 check run matches QuickBooks at $18,671.30 across 35 unpaid bills, with no payments scheduled or sent. It says it entered 16 new bills totaling $7,163.50 and updated two existing bills: Northline Paper invoice #1048 falls from $604 to $570 for 10 cases delivered, while Folio Copier Service invoice FS-3721 now references PO AC-8056-M and remains $421. The interface says the run was saved in the default schedule and offers a view of the changes.
For the weekly meeting deck, Astra is asked to use material in the Alder Street deal folder. Its draft describes a two-store pilot at Main and Broad and Spruce and Oak, with expected revenue above $25,000 per location based on previous pop-up performance. It notes that extending the option by January 31 would preserve a 20% margin, and that a Form 8-K needs to be filed on the option agreement date. The slide shown on the meeting-room screen condenses those details into three points: the pilot and revenue expectation, the extension date and margin, and the filing requirement.
Together, the examples establish the range of work being demonstrated: updating a payables schedule from invoices and converting deal-folder information into a meeting slide. They also show the limits of what is visible. The payables screen reports changes and explicitly says no payments were scheduled or sent; the slide is a draft based on folder contents. The examples present prepared artifacts, not a demonstration of how their accuracy was checked.
Sol is positioned for everyday work; Astra for harder assignments
OpenAI draws the distinction between GPT-6.2 Sol and Astra around breadth, efficiency, and task difficulty. Astra is the model for the hardest projects. Sol is described as a more efficient, versatile model for everyday work, and as smarter and more efficient than GPT-6.1 Sol. OpenAI says Sol lets users complete more tasks on the same paid plan, handles multistep agentic workflows, writes clearly and naturally, and is intuitive to collaborate with. It also says many OpenAI employees use Sol as their default for technical and non-technical tasks.
The launch description says both models produce better results on demanding tasks with fewer output tokens, which OpenAI says can help users do more within usage limits and make frontier performance more affordable at scale. The examples shown here focus on Astra; the distinction between the models comes from OpenAI’s product description rather than a side-by-side task comparison.
Both models are available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Pro 500 subscribers can use Ultrafast with either model. Pro 200 subscribers receive Ultrafast access starting with Sol, and eligible Enterprise and Edu users receive the same access when a workspace administrator enables it. Developers can access both models through the OpenAI API, where Ultrafast is available for both at launch.