AI RADAR / PRACTICAL GUIDE
GPT‑6.2 Astra and Sol: what changes for creative work, and how should you test them?
A creative-production perspective on GPT‑6.2 Astra and Sol: what official examples actually show, how design and motion teams can run a useful trial, and how to measure cost and delivery accuracy. Includes sourced explanations, a reusable brief and a checklist.
- 01Verify access and sources in the correct environment.
- 02Choose a small real assignment and acceptance criteria.
- 03Compare Sol and Astra using the same inputs.
- 04Open the deliverable and record corrections and cost.
Start with the date and access: regular Chat and Work are different surfaces
The first question about a new model need not be its benchmark score. For a designer, video editor or small production team, a more useful question is which steps change when turning an actual assignment into a deliverable you can verify. That is our approach to GPT‑6.2 Astra and Sol. This is not an independent performance test: we distinguish verified product information, provider demonstrations and workflows we propose that readers try.
OpenAI’s API changelog records the release of GPT‑6.2 Astra and GPT‑6.2 Sol on 17 November 2026. The ChatGPT release notes say the new models are in the Work model picker; this rollout does not add either model to regular Chat. Not seeing the name in Chat therefore does not invalidate the announcement. Check the appropriate surface and your account’s available models first. Workspace administrator settings also matter in managed environments.
This distinction affects production planning. API availability does not mean every ChatGPT account has identical tools, file access and usage limits. Record the environment alongside the model name: Work, Codex or your own API application. Otherwise, differences between two setups can be mistaken for differences between models alone.
What has actually been demonstrated?
OpenAI’s announcement shows editable Excel workbooks, presentations based on a brand template, interactive web and CAD examples. The financial example uses the fictional Alderfen Field Systems company; it is not an actual investment result. The lamp example runs from visual references through CAD to an object assembled by hand. These are examples selected by the provider, not AI Radar’s independent tests or guarantees for every assignment.
Assess more than the final frame of a demo. What was the input file, what deliverable was requested, which tool was used and what did a human do at the end? Creating a CAD image and verifying the measurements of a manufacturable part require different acceptance criteria. Likewise, an attractive presentation preview does not establish that fonts transfer correctly, charts remain editable or sources are disclosed.
A proposed designer’s test: from client brief to presentation
Our first proposed experiment is a small deliverable you already produce regularly, such as a six-page product presentation. Provide an authorized client brief, usable product photos and your own presentation template within the same task. Before starting, define the six pages’ functions: problem, audience, solution, product difference, evidence and next step. Explicitly instruct the model to leave unsupported numerical claims unfilled.
Make three separate checks. For content, identify the input supporting each claim. For design, inspect heading hierarchy, overflowing text, image proportions and brand consistency. For delivery, open the file in your own presentation application and edit its text and charts. A good-looking PDF is a different deliverable from a presentation the client can edit.
Do not count time to the first draft as the entire benefit. Record the interventions needed to make it client-ready. Replacing an incorrect reference and shortening a heading are different kinds of correction. This reveals where the model helps and where an experienced eye is still required. This section proposes an experiment; it is not a client project conducted by AI Radar.
Source: Using GPT-6.2
A proposed video and motion workflow: evidence for every scene
Instead of giving a long source file the instruction “make an impressive video,” request a scene-by-scene production table. Each row should contain one idea, its narration, a source link, the real footage to use, the motion direction and a verification criterion. In an API-pricing scene, for example, the source can be a pricing document, the screen text specifies the billing unit, and the animation explains why input and output rates differ.
Real product footage and explanatory motion have different jobs. Footage establishes what is being shown; motion directs attention to the relevant detail. An arrow diagram can clarify a brief’s workflow, but a drawn interface must not be passed off as a product recording. Credit official demonstrations and label your own suggested method as a proposed workflow.
Do not stop at a scene list. Play the actual MP4 from beginning to end: is the topic clear in the first frame, does screen text coincide with narration, is the last word intact, and can viewers read the title and credit at phone size? Less movement sometimes improves comprehension. If viewers need to read a price or step, let the text stay still once it has settled.
Source: Using GPT-6.2
Software and interactive design: test behavior after appearance
For a team trying an interactive website, a narrowly scoped product page makes a useful starting point. Supply brand colors, authorized images, sample copy and the page’s primary purpose. Then make acceptance criteria concrete: no mobile overflow, the main button leads to the right section, incomplete form data receives a useful explanation, and the page is usable with a keyboard.
When the preview opens, do more than look at it. Submit invalid information and empty fields, try the menu on a small screen, open links and verify that the requested files actually exist. When changing an existing project, review the modified files and run the relevant project checks. Record the difference between code being written and the requested behavior being verified.
The API documentation says Astra and Sol accept text and image inputs and produce text. Web search, code execution and computer use happen through tools; API tool calling uses Responses. These models should not be described as direct video or audio generation models. Results depend on the connected tools and the permissions granted to them as well as the model.
Source: GPT-6.2 Astra model and pricing
Astra or Sol? Define the same assignment first
The official usage guide recommends Astra for the most demanding work and Sol for complex work at lower cost. Our starting suggestion is to try a small, repeatable, readily checked assignment with Sol, then compare Astra on the same inputs for the part involving difficult reasoning, many dependencies or demanding delivery constraints. This does not claim Sol is always sufficient or Astra is always better.
Change as little as possible besides the model name. Use the same source files, permissions, delivery format and assessment list. Comparing a detailed brief in one run with a single sentence in another does not isolate the model difference. Do not generalize one assignment’s result to all your work, either: slide layout, data checking and code fixes involve different errors.
A simple experiment log can record the task, model and environment, initial delivery time, correction time, substantive errors, missing files and total cost. Measure the cost of a usable final result rather than how impressive the first draft feels. That log may tell you more directly than a promotional score which model should handle the next assignment.
Source: Using GPT-6.2
Read pricing correctly: subscriptions, tokens and task costs differ
At our 18 November 2026 check, Standard short-context API rates per million input / cached input / output tokens are $8 / $0.40 / $40 for Astra and $2 / $0.10 / $10 for Sol. These are neither ChatGPT monthly subscription prices nor fixed video-production fees. Usage, long context, speed tier, cache writes and tool fees can affect the bill.
Budget for more than the successful first request. Consider retries, unnecessarily resending the same files and the labor of checking the output. A job that consumes fewer tokens may still be less economical for your team if it requires extensive manual correction. Measuring your own workload is a more reliable decision method than assuming a universal savings percentage.
Source: GPT-6.2 Sol model and pricing
A useful brief names the output, boundaries and completion criteria
OpenAI’s usage guide recommends specifying the completed outcome and the decisions that require user input for longer tasks. It also recommends defining visual direction and typography in design work. In practical terms, state the deliverable up front instead of asking vaguely for something attractive.
A proposed reusable brief:
Adapt this brief to the job. File format can be central to a product presentation, while narration endings and screen-text timing matter to motion work. State the authorization boundary as well: permission to edit files does not necessarily grant permission to publish immediately on a social account. Completion should be observable, such as a file opening successfully or a link actually working.
EXAMPLE PROMPT / ADAPT TO YOUR BRIEF
Using the authorized sources I have attached, create [deliverable] for [audience]. Preserve the brand template and typefaces. Cite the source of every numerical claim and do not invent missing information. After drafting, open the file and check content, overflow, links and editability. List completed files, verification results and unfinished steps separately. Apart from [specified decisions], make reasonable assumptions about reversible details and finish the work. Do not perform publication or a new paid action unless I have authorized it.
Source: Using GPT-6.2
Why discuss safety and errors in this article?
The system card includes evaluations of whether an obstacle is disclosed to the user, including missing files and unavailable search tools. These evaluations do not mean the same failure happens in every real task. They do, however, give a useful reason to match a completion statement with an existing file, an accessible source and a verification result.
Reduce your own check to three questions: does the requested output exist, is it consistent with the source, and are incomplete steps explicit? Ask the model to disclose unavailable tools instead of hiding the problem. Do not ask it to invent a precise number when a source is missing. For a client document, financial workbook or public post, do not leave responsibility inside a vague claim that AI handled everything.
Source: GPT-6.2 System Card
Finish your first experiment this way
Choose the next recurring assignment, prepare a clear input of only a few files and write the acceptance criteria. Take one model through to completion, then try the other under the same conditions. Record final accuracy, correction needs and cost. Base the choice on which model delivers a verified result for the task rather than the appeal of the newest name.
Source: Using GPT-6.2
Before you deliver
- Record the model name, environment and test date.
- Establish permission for inputs and sources for claims.
- Label real demos, original diagrams and proposed methods separately.
- Actually open the requested editable file.
- Check links, calculations or interactions as appropriate to the assignment.
- Verify readable text and product detail at phone size.
- Disclose unfinished steps and human-completed work.
- Record correction time and total cost.
Sources and method
This guide combines the official documentation below with original workflow suggestions from AI Radar. Example prompts are not production test results. Prepared with AI assistance; this is not a claim of independent product testing or human review.
Sources checked: 18/11/2026
Share your discovery.
Send this story to a friend.


