In one early demo, GPT-6.2 Astra recreated a recording studio from two video frames, turning it into a space you could walk through like a video game. It was less convincing in a prerelease test by Every, a publication that reviews AI tools. The editors asked Astra for help refining an article opening and preferred the previous version’s suggestion. What accounts for such different reactions to OpenAI’s new model? And how did Astra outperform Claude Fable 5.2 in a test where its average cost per task was just 5% of Claude’s?
1. What Does GPT-6.2 Astra Do Better Than Its Predecessor?
On November 17, 2026, OpenAI released two models. GPT-6.2 Astra builds on the GPT-6 Astra model released in September, with improvements for complex work such as analyzing data from multiple sources or making changes to a large application. GPT-6.2 Sol, released alongside it, is a lower-cost model for everyday tasks.
One improvement is how much Astra can handle on its own when preparing a presentation. Given a screenshot of a single slide, it can create more slides using the same layout, colors, and font sizes. In a quarterly earnings presentation OpenAI shared, Astra put the chart’s main takeaway directly in the slide title, “Sales Rose 5.3%.” Readers could see what the data showed at a glance. Prerelease testers noted that titles like this made presentations easier to use with fewer edits.
Presentations are one example of the work OpenAI designed the new Astra to handle. To see how the model performed on a broader range of professional tasks, OpenAI built its GDPval v2 benchmark around real projects in fields including finance and consulting. Astra scored 6.2 percentage points higher than its predecessor while using 22% fewer output tokens.
In the recording studio demo mentioned at the start, Astra helped reconstruct the room from two video frames. The result was an editable 3D scene in Blender, so the presenter could swap the picture on the wall or change the lighting. OpenAI also showed how work on a digital design could lead to a physical object. The team used Astra to design a paper lamp inspired by the balconies of Chicago’s Marina City towers. After a few revisions, they assembled the lamp by hand.
2. GPT-6.2 Astra, Claude Fable 5.2, and Gemini 4 Carbon. Which Performs Best in Tests?
Does the new Astra’s advantage hold when the model has to tackle a professional assignment from start to finish? That is what Agents’ Last Exam 1.1 measures. Each assignment involves several steps, and models earn points for partial completion. GPT-6.2 Astra earned 78.3% of the available points in the results published by OpenAI. Google’s Gemini 4 Carbon came closest at 77.4%. The previous Astra scored 73.1%, and Anthropic’s Claude Fable 5.2 scored 72.8%. The cost difference was particularly large. According to OpenAI, GPT-6.2 Astra’s average cost per task was about 5% of Claude Fable 5.2’s and less than a third of the previous Astra’s. Claude’s results exclude nine tasks where the system switched to a fallback model.
Agents’ Last Exam scores complex assignments as a whole. The independent tests from Artificial Analysis show how Astra and Claude perform in specific areas. The table compares GPT-6.2 Astra and Claude Fable 5.2 on tasks such as building interactive 3D scenes and making code changes. Both models ran at their highest reasoning setting, which gives them the largest compute budget for each task.
| Test and skill measured | GPT-6.2 Astra | Claude Fable 5.2 |
|---|---|---|
| Intelligence Index v5.0.1 – composite score across the benchmark suite | 62 points | 65 points |
| Terminal-Bench-Science 0.1 – scientific work using computational tools | 78.1% | 73.3% |
| FrontierCode-AA 2.0 – code changes that meet a project’s acceptance criteria | 45.4% | 48.8% |
| Harvey LAB-AA v1.1 – legal assignments graded against all requirements | 15.0% | 7.2% |
| AA-3D – interactive 3D scenes rated by people | 1318 Elo | 1271 Elo |
Astra ranked second overall, behind Fable and ahead of Gemini 4 Carbon (60 points) and Claude Opus 5.5 (59 points). It took first place in AA-3D. In that test, participants compared pairs of scenes built from the same prompt and chose the stronger one without knowing which model built it. The Elo ratings reflect those choices. Astra also led Fable in the legal benchmark with a score of 15%. A response earns full credit only if it meets every criterion, and inventing a material fact results in a score of zero for the task.
3. How Much Does GPT-6.2 Astra Cost, and Where Do the Savings Come From?
At standard API rates, GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens. The API is the interface applications use to call the model. Both rates are 20% lower than for GPT-6 Astra. Cached input costs 60% less than it did with the previous Astra. That rate matters when an application repeatedly sends the model the same long set of instructions.
| Model | Uncached input | Cached input | Output |
|---|---|---|---|
| GPT-6.2 Astra | 8.00 | 0.40 | 40.00 |
| GPT-6 Astra | 10.00 | 1.00 | 50.00 |
| GPT-6.2 Sol | 2.00 | 0.10 | 10.00 |
| Claude Fable 5.2 | 8.00 | 0.20 | 40.00 |
Suppose a report requires 100,000 uncached input tokens and 10,000 output tokens. At that usage, the report would cost $1.20 with either the new Astra or Fable, $1.50 with the previous Astra, and $0.30 with Sol. That estimate assumes a standard request without any additional paid tools.
In practice, one model may open the same file several times to find the data it needs, while another finds it on the first pass. Models can also use different numbers of tokens to work out a response. On the Intelligence Index suite at maximum reasoning effort, the weighted average cost per task was $4.00 for Astra and $26.54 for Fable, which scored higher overall. Astra’s settings also affected the cost. Lowering reasoning effort from max to xhigh brought its average cost down to $2.52 per task, with the rounded score holding at 62 points.
4. Where Can You Use GPT-6.2 Astra?
Astra is available in ChatGPT Work, Codex, and the OpenAI API. According to the ChatGPT Work documentation, the model is available on Plus, Pro, Business, Enterprise, and Edu. In ChatGPT, open Work and select Astra in the advanced model picker. For Enterprise and Edu workspaces, an administrator enables access. OpenAI’s launch-day list covered paid plans starting with Plus and omitted Free and Go.
In Codex, select the model in the app or use the /model command in the CLI. Teams integrating Astra into their own software can use the API model ID gpt-6.2-astra. The lower-cost model in the same series is available as gpt-6.2-sol.
5. What Did Early GPT-6.2 Astra Testers Like, and Where Did It Fall Short?
- The team at Every, a publication that reviews and tests AI tools, used GPT-6.2 Astra in its day-to-day work for several weeks before launch. To see whether the new model was better at helping edit an article, the testers gave both versions of Astra the same piece about AI spending and asked for feedback on the headline and opening. The older model identified the most interesting idea in the piece, that a company’s AI spending might be too low, and suggested “Your AI bill might be too low.” GPT-6.2 Astra led with a rewritten opening under the headline “What if we’re using too few tokens?” and saved its comments for the end. Every cofounder Dan Shipper preferred the older Astra’s headline. He explained his reservations about the new model’s opening in Every’s recap of its prerelease tests.
- Every also tested Astra on coding tasks. When asked to pick an item from a backlog, it carefully checked whether anyone had already completed it. When building an entire application, it sometimes stopped too early, and changes from Claude Fable 5.2 were easier to merge. In a separate Artificial Analysis index that measures coding agents, the new Astra scored about the same as its predecessor.
- Can Astra adapt a design to its subject? Peter Gostev tested that for Arena by asking the model to create websites and interactive 3D scenes. In the scenes, he singled out the detail on building facades and the way Astra rendered light on water. In the website designs, he noticed the model often reused similar layouts. A website promoting a parade and a sushi restaurant’s site looked strikingly similar despite being built from different prompts.
- Ethan Mollick, a professor at the Wharton School, gave Astra a more complex assignment. He asked it to build a playable simulation of a space elevator ride using Blender and Unity and was pleased with the result. In his other tests, he found Astra harder to steer when he had a specific outcome in mind.
6. Is GPT-6.2 Astra a Breakthrough for Businesses?
In early tests, GPT-6.2 Astra completed complex, multistep tasks at a lower cost than its predecessor. That difference can matter for a company that collects data from several spreadsheets every week, prepares a report, and turns the findings into a presentation using a standard template. Even a small saving on each report can add up over a year. To assess what the company actually gains, factor in the time it takes to review and correct the result. The same applies to software development. If an engineer has to prompt the model several times to finish a change, that extra work can eat into the savings.
That extra time matters when choosing between Astra and the lower-cost GPT-6.2 Sol. Cognition’s FrontierCode 2.0 tests whether code changes meet a project’s requirements. Sol scored less than half a percentage point below Astra. At the highest reasoning setting for both models, Sol’s average cost per task was about one-fifth of Astra’s. In your own project, compare the full cost of getting a change accepted, including the time people spend reviewing and revising the code.
For everyday coding tasks, it is worth comparing Astra with the less expensive GPT-6.2 Sol. On Cognition’s FrontierCode 2.0, Sol scored less than half a percentage point below Astra, and its average cost per task was about one-fifth of Astra’s. Both models ran at maximum reasoning effort. In a company pilot, you can measure what it costs to get a change through testing and accepted into the project, including the engineer’s time reviewing and revising the code. That comparison can help you decide where to use AI agents in software development.
7. Find the Right AI Model for Your Business
Start with a task your team performs regularly and already knows the cost of, such as preparing a monthly sales report. At TTMS, we can help you compare models on the same task and calculate the full cost of producing a report that’s ready to send. We’ll account for both model usage and the time your team spends checking the data and making corrections. We can plan a pilot together as part of TTMS’s AI consulting services.

Does ChatGPT Plus include API access to GPT-6.2 Astra?
OpenAI bills ChatGPT subscriptions and API usage separately. With Plus, you can select Astra in ChatGPT Work and use it within your plan limits. If you use Astra in your own application, for example to generate reports automatically, model calls are billed to your API account. This applies even if you sign in to both services with the same email address.
How long a document can GPT-6.2 Astra analyze?
Through the API, Astra can accept up to 922,000 input tokens in a single request. That limit covers the document text, your instructions, and any earlier messages included in the request. For large sets of documents, there’s also a pricing threshold to consider. Above 272,000 input tokens, the rates for input and cached input double, and the output rate increases by 50%. The higher rates apply to the entire request.
Does Ultrafast speed up an entire task in GPT-6.2 Astra?
According to OpenAI, Astra can generate tokens up to 4.5 times as fast in Ultrafast mode as in Standard mode. You’ll notice the biggest difference when the model spends most of its time generating a response. If it needs to open files or wait for an external tool, the overall speedup may be smaller. In ChatGPT Work and Codex, Ultrafast also uses up your plan allowance or available credits faster.
Can GPT-6.2 Astra create a 3D scene from a photo?
Yes, if it has access to an environment where it can build or render the scene. Astra accepts images as references and returns text, including code, through the API. In the recording studio demo, Codex used Astra to build an editable scene in Blender. The presenter could then change the lighting and view the room from another angle.
How can I check whether Astra used all the files in an analysis?
If you’re putting together a report from several spreadsheets, ask Astra to list the files it used and flag any it couldn’t open. You can also ask for the worksheet name alongside each key figure. In a test designed by OpenAI for this purpose, the new Astra omitted information about obstacles it encountered more often than its predecessor. A file list and references to the relevant worksheets help you check what the model based its conclusions on.