OpenAI's 17 November 2026 launch pairs two models with different jobs: GPT-6.2 Astra for difficult projects and GPT-6.2 Sol for everyday work. The three official videos make that launch concrete. One stages an office scramble over invoices and a workbook. Another rebuilds a room in Blender. The third introduces a Sales plugin that brings customer context into ChatGPT. Together they show three different things: generating a work product, building an application, and connecting existing tools.
Start with Sol for a recurring task with clear inputs and a result you can check. Test Astra when you need a more involved build, detailed 3D work, or a deliverable that has to reconcile several sources. Judge the output and review time before changing your default.
Watch: Models Built for the Real World
The office film has three jobs running against the clock: update Friday's payables from new invoices, finish a slide for the weekly presentation, and make a workbook for a one-year delay in expected integration savings. The workbook request compares net debt through 2031 and next year's earnings per share against plan. In the script, the final answer is $32 million more in 2031 debt and $0.30 less in next-year EPS. Those are figures from the film, not a verified forecast or a measurement of the model's accuracy or speed.
For a real pilot, keep the invoice total traceable to approved sources, check duplicates and reconcile the worksheet before posting any changes. For a scenario model, change only the agreed assumption, keep the baseline intact and have a reviewer trace the result through the formulas. A draft for approval is a useful output. It isn't authorization to send a payment or publish a forecast.
Model Access, Standard API Rates and Ultrafast
| Model and API ID | Standard USD per 1M tokens | Where to start |
|---|---|---|
GPT-6.2 Astragpt-6.2-astra | $8 input / $0.40 cached input / $40 output | Complex analysis, multi-file builds and detailed 3D prototypes. |
GPT-6.2 Solgpt-6.2-sol | $2 input / $0.10 cached input / $10 output | Repeatable office work, first drafts and well-scoped coding jobs. |
These are short-context Standard rates. OpenAI's model docs list different rates for cache writes, requests above 272,000 input tokens, and faster processing tiers. Astra's standard input and output rates are 20% below GPT-6 Astra's; Sol retains GPT-6.1 Sol's rates. Both models can cost less on a task if they use fewer tokens, but a list rate doesn't tell you how many attempts, tool calls or editing rounds a finished result will need.
At launch, OpenAI listed both models in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu, with admin enablement for eligible Enterprise and Edu workspaces. Pro 500 gets Ultrafast for both models; Pro 200 starts with Sol. The API Ultrafast tier supports both models at a premium with separate limits. It is a speed setting, not a different model or a guarantee that an entire workflow finishes sooner.
OpenAI highlights template matching for slides and professional documents in its own launch tests. In a separate evaluation, Artificial Analysis placed Astra first on its human-judged interactive 3D test at launch but found it level with the previous Astra on its Coding Agent Index in Codex at max effort. That's a reason to test the actual workflow rather than treating one ranking as a verdict across every kind of work.
Watch: From Two Video Frames to a Blender Set
At 00:09, the presenter reveals that the room behind him is a digital set. Two still frames went into Codex, and Astra iterated on a Blender recreation with requests to match camera blur and studio lighting. The video then changes the decor, adds a walkable view and shows a virtual webcam background. A close match from the reference angle is a useful result. Two views cannot verify the dimensions or geometry of parts of the room they don't show.
At 01:04, the app adjusts the set lighting to match the webcam. The Decisions API documentation lists gpt-6-luna as the supported model for that separate API. Astra is used to build the set in Codex; the lighting adaptation is a feature of the resulting app. Before relying on a reconstructed set, inspect it from a new angle, try different lighting and check what webcam data the app sends off-device.
Bonus: ChatGPT for Sales
ChatGPT for Sales is a separate launch, not a third model. OpenAI introduced its Sales plugin and more than 30 new or expanded partner plugins. The practical starting point is a call brief: pull approved account history and meeting notes, list open questions, and prepare a draft follow-up. Examples in the announcement include Salesforce and HubSpot for account context and Gong or Granola for call notes.
The setup guide lists the Sales plugin on all ChatGPT plans, including Free, while partner access varies by plan, region, workspace settings and connected accounts. Start with read-only access where available. Check facts and commitments against the source records, then have a person approve outbound messages or CRM changes separately.
A First Pilot You Can Actually Grade
Run the same synthetic invoice packet with Sol and Astra. Set the expected row count, totals by currency and known duplicates beforehand. Ask a reviewer to score the results against those answers and log correction time. If Sol passes and Astra doesn't materially reduce review time, keep Sol for that job. Don't use live payables data until the read-only workflow and review process are approved.
Compare models on a draft payables sheet
Synthetic inputs, traceable rows and no payment actions.
Launch Film Timestamps
| Time | Scene |
|---|---|
| 00:01 | Twenty invoices arrive before the meeting |
| 00:13 | The weekly presentation needs a slide |
| 00:32 | Request a one-year delay scenario |
| 00:41 | Ask for the Friday payables update |
| 00:51 | Track net debt and next-year EPS against plan |
| 01:39 | Deliver the scenario readout |
Official Sources and Limits
The three embedded videos are credited to OpenAI and presented as vendor demonstrations. The launch announcements and API documentation support model access, pricing and plugin details; Artificial Analysis reports its separate benchmark. JQ AI SYSTEMS did not replicate the office, Blender or sales demos. The pilot prompt is an original exercise, not a result from the launch. Evidence cutoff: 18 November 2026.