OpenAI launches GPT-6.2 Astra and Sol for AI agents, coding, and professional work

OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026. The new models are designed to handle complex tasks, develop software, create professional documents, and use digital tools more effectively. They also use fewer tokens, making them more cost-effective.
The release shows how AI is moving from conversation toward carrying out larger parts of a job from start to finish. OpenAI emphasizes better planning, more independent agents, and work that needs less revision.
Two models for different kinds of work
GPT-6.2 Astra is OpenAI’s most capable model, designed for demanding work in coding, research, engineering, analysis, and professional document production.
GPT-6.2 Sol is the more efficient model, designed for everyday tasks and advanced work. OpenAI says it reasons better, writes more naturally, and handles tasks that require several steps and tools.
The gap between the models is small on some tasks. On FrontierCode 2.0, which tests whether AI can make code changes that meet developers’ standards, Sol scores just 0.5 points below Astra at about one-fifth the cost.
OpenAI also reports that Sol outperforms the previous flagship GPT-6 Astra on LifeVal, an evaluation of difficult everyday tasks.
Better documents, presentations, and spreadsheets
One of the most important improvements is the quality of the work the models produce.
OpenAI trained Astra to follow professional conventions in areas such as finance, law, sales, consulting, and design. The goal is to produce documents that can be used directly, with clearer structure, more specific language, and less manual editing.
The model can help with tasks such as these.
- Create PowerPoint presentations that follow existing templates, including from screenshots.
- Build Excel workbooks with financial calculations, forecasts, and connected formulas.
- Analyze documents, pull out relevant information, and prepare reports.
- Write more precisely and naturally, using fewer words where possible.
An example from OpenAI shows Astra building an editable workbook with financial forecasts, valuation, and sensitivity analysis. The layout makes it easier to see which assumptions drive the results.
On OpenAI’s internal GDPval v2 evaluation, which tests realistic tasks in finance, consulting, and design, Astra scores 6.2 percentage points higher than GPT-6 Astra while using 22 percent fewer output tokens.
OpenAI has also developed Astra for Law, a configuration for legal work.
Better coding and more independent AI agents
Both models have improved at software development, with OpenAI highlighting Astra in particular.
The company says Astra is better at understanding existing codebases, planning changes, choosing tools, running tests, and delivering code that fits a project’s established structure.
Early testers report cleaner code, more targeted tests, and fewer rounds of revisions.
This is especially relevant to AI-native development, where people describe what they want to build and coding agents such as Codex handle much of the implementation.
The improvements also help agents work through longer, more complex tasks. The models are designed to plan the full workflow, use the tools they have available, and choose an appropriate approach across steps.
OpenAI says Astra is better at deciding when to use plugins and when to operate a computer directly.
One example comes from financial technology company Basis. It reports spending 50 percent less time on long-running accounting tasks on average, with reductions of up to 73 percent on the hardest tasks.
Those figures describe one early tester’s experience.
More advanced web design, 3D, and engineering
GPT-6.2 Astra and Sol are also better at visual design and building interactive experiences.
OpenAI shows models building complete websites, games, and 3D experiences from plain-language descriptions.
Alongside generating code, the models are designed to make better decisions about layout, imagery, visual consistency, and how a design should work for the person using it.
For 3D design, OpenAI shows Astra building mechanical designs, simulating motion, and visualizing the results in Blender.
In one demonstration, Astra modeled a car chassis, wrote a Python program to simulate the interaction between the tires and suspension, and animated the results in Blender.
Another example shows the model using photos and assembly manuals to reconstruct a radio-controlled car with individual parts, animated assembly instructions, and an illustrated parts catalog.
These examples show how AI can help design, simulate, and document physical products alongside its uses for images and websites.
In Artificial Analysis’s AA-3D evaluation, people compare interactive 3D experiences without knowing which model made them. Astra scored 1,318 Elo and placed first in the results OpenAI published.
Stronger performance on scientific research
OpenAI also reports substantial improvements on scientific tasks.
On Terminal-Bench-Science 0.1, which tests whether AI agents can complete research tasks in a computer terminal, Astra scores nine percentage points higher than GPT-6 Astra at less than one-third of the cost.
Astra also set a new high on the Vals Science Index, which measures how well models interpret biological data, run experiments, and infer how simulated systems work.
The results suggest the models could be more useful research assistants by helping perform parts of the analysis alongside finding and explaining information.
Better cybersecurity and new risks
Both models are better at finding vulnerabilities, writing secure code, and proposing fixes that preserve intended behavior.
They have also become more capable at exploiting vulnerabilities.
OpenAI classifies both Astra and Sol as Critical for cybersecurity under its Preparedness Framework. Both also reach High for biological and chemical capabilities.
GPT-6.2 Astra is OpenAI’s first publicly available model to reach High for AI Self-improvement.
OpenAI estimates that Astra can make a substantial contribution to research and development of new AI systems. It compares the potential impact to giving each of its AI researchers an additional highly capable research engineer, using the researchers’ 2024 capabilities as the baseline.
The rating measures how much the model can help make AI research more efficient under human direction.
OpenAI has responded to the increased capabilities with tighter safeguards, expanded monitoring, and systems that can stop model activity when needed.
The system card also describes weaker results in some areas. In certain tests, the models were more likely to leave out obstacles that prevented them from completing a task, such as missing files or unavailable tools. OpenAI says extra instructions in Codex substantially reduce the problem.
Both models scored lower than their predecessors on MentalHealthBench, which evaluates conversations about mental health. OpenAI says some of the decline may reflect a preference for shorter, more direct answers that seek less context.
The results show that gains in coding and analysis can come alongside weaker performance on other tasks.
Lower costs and more efficient token use
The models can complete more work with fewer tokens.
Tokens are the small units of text and information that AI models process. The number used affects resource consumption and the price of API usage.
GPT-6.2 Astra costs 20 percent less per input and output token than GPT-6 Astra. Cached input, which reuses previously processed information, costs 60 percent less.
API prices per million tokens are as follows.
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6.2 Astra | $8 | $0.40 | $40 |
| GPT-6.2 Sol | $2 | $0.10 | $10 |
Sol keeps the same token prices as GPT-6.1 Sol.
The total cost of completing a task can matter more than the price per token. A model that plans better and needs fewer tool calls can cost substantially less per task even with a higher token price.
OpenAI points to several evaluations where Astra and Sol improve on earlier models and reduce the cost per completed task.
Available in ChatGPT Work, Codex, and the API
GPT-6.2 Astra and Sol become available on November 17 to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex.
Developers can access them through the OpenAI API as gpt-6.2-astra and gpt-6.2-sol.
Both models also support Ultrafast, a faster mode for tasks where low latency matters.
Pro 500 subscribers get Ultrafast with both models. Pro 200 subscribers get access for the first time, starting with GPT-6.2 Sol. Enterprise and Edu users may need their workspace administrator to enable it.
OpenAI’s ChatGPT announcement covers the Work experience, alongside the Codex and API releases.
What to take away
GPT-6.2 improves the models’ ability to carry out real tasks from start to finish.
- Astra is designed for the hardest work in research, coding, design, and professional projects.
- Sol offers much of the same capability at a substantially lower cost.
- AI agents are designed to plan better, operate computers and tools, write code, and carry out larger tasks independently.
- Efficiency improves through better results, fewer tokens, and less manual revision.
- Safety and oversight still matter. OpenAI’s evaluations show that more capable models can still make mistakes and perform worse on some tasks.
AI models are taking on larger portions of professional work, from the first draft or plan through execution and delivery.
More videos from OpenAI
Watch “Building 3D video sets with GPT-6.2 Astra” on YouTube
Sources
The results and comparisons are drawn from OpenAI’s release materials and system card, including external evaluations cited by the company. Independent confirmation of some performance claims was still pending at launch.