What Is GPT-6 Astra? Differences from GPT-5.6 Sol, Capabilities, Pricing, and Availability
When a new model arrives, the practical questions are what it can do, whether your plan gets it, and what it costs. This guide takes GPT-6 Astra in that order.
On September 3, 2026, OpenAI announced GPT-6 Astra. Beyond answering questions, Astra is designed to carry out multi-step work, including using computer interfaces, writing and running code, and creating documents.
This article explains how Astra differs from the previous model, GPT-5.6 Sol, what it can do, where it is available, how much it costs, and what OpenAI’s “Critical” designation means.
Key Points
- Astra is a model with stronger capabilities for multi-step work: computer use, coding, document creation, and scientific tasks
- In standard ChatGPT chat, the Astra-based model is offered as “GPT-6 Pro”; in ChatGPT Work, Codex, and the API, it is called “GPT-6 Astra”
- Availability in ChatGPT depends on both your plan and the product, and API usage is billed separately from your ChatGPT subscription
- “Critical” is a cybersecurity capability threshold in OpenAI’s internal framework—not an overall safety rating for Astra or ChatGPT
1. What GPT-6 Astra Is
GPT-6 Astra is the GPT-6 generation model that OpenAI describes as its most capable. In its public materials, OpenAI presents Astra as built for the hardest sequences of work, citing complex reasoning, coding, computer use, research, and document creation.
One point to clarify at the outset is that the naming differs by product. In standard ChatGPT chat, the Astra-based model is offered as “GPT-6 Pro.” In ChatGPT Work, Codex, and the API, it is called “GPT-6 Astra.” The available features and usage limits also differ by product.
Terms: agents and computer use
An agent is a system that works toward a given goal by assembling a sequence of steps and using tools along the way.
Computer use is one way an agent can act. It involves interacting with browser or application interfaces—clicking, typing, and checking results.
Selecting Astra does not, by itself, give the model unrestricted access to your computer. What it can access and control depends on the product you are using, the tools you connect, and the permissions you grant.
Not every task calls for Astra. For short questions and everyday writing, earlier models are often sufficient. OpenAI positions Astra for long-running, complex tasks.
Astra is strongest at complex reasoning and work that spans several steps. One of its main strengths is the ability to carry out chains of actions—searching for information, opening resources, entering data, verifying results, and producing a summary—with fewer interventions along the way.
2. What Changed From GPT-5.6 Sol
Here are five representative evaluation results published by OpenAI.
| Area | GPT-6 Astra | GPT-5.6 Sol | What it means for you |
|---|---|---|---|
| Computer use (OSWorld 2.0) | 72.6% | 65.7% | Better at working through multiple steps while reading the screen |
| Workflow automation (AutomationBench) | 41.4% | 18.1% | A large gap on an evaluation of chaining business steps together |
| Coding (Terminal-Bench 4.0) | 57.9% | 37.3% | Better at practical development work using terminals and tools |
| Scientific workflows (Terminal-Bench Science 0.1) | 64.6% | 22.4% | Better at data analysis, simulation, and model fitting |
| Long context (MRCR 512K–1M) | 96.3% | 73.8% | Better at locating needed information within very long contexts |
Term: benchmark
A benchmark is a fixed set of tasks given to a model, scored on accuracy or a similar measure.
Each name above is such a test: OSWorld covers operating a computer interface, and Terminal-Bench covers development work at the terminal.
These are test scores, not a direct measure of quality on your own work.
On computer use, OpenAI reports that in an OSWorld 2.0 latency simulation, Astra scored higher than Sol while taking roughly 47% less time per task.
The results point to four practical improvements: stronger and faster performance on screen-based tasks; larger gains on multi-step workflows; better handling of development work across terminals and test suites; and an improved ability to maintain context and constraints across long documents, conversation histories, and tool outputs.
On safety, OpenAI reports that Astra improved over Sol in robustness to jailbreaks (prompts designed to bypass safety limits), in staying within instructed boundaries, and in safety behavior in browser and workplace environments.
Note: how to read these results
These are evaluation results published by OpenAI. They do not guarantee the same margins on any given piece of real work.
Evaluation environments and the standard ChatGPT product can differ in system configuration and available tools.
MIF has not independently reproduced these numbers.
3. What You Can Do With GPT-6 Astra
Computer use
The examples OpenAI cites in its announcement include filling in online forms, updating customer records in a CRM, organizing a calendar, conducting online research, drafting summaries for emails and documents, installing software and verifying that it runs, and front-end QA on a website.
These tasks are not guaranteed to succeed every time. Website restrictions, sign-in requirements, confirmation steps, and the need for human approval may still apply.
In practice, the best fit is work that is tedious and time-consuming for a person but whose steps are well defined—checking conditions across several sites and reformatting the results into an internal template, for example. Work where the criteria for judgment are not written down, or where a mistake is hard to undo, is better designed with a human check in the loop.
Coding
Astra can carry a task across terminals, browsers, and test suites as a single workflow. It is designed for longer-running tasks such as debugging, refactoring, and database migrations, including work that requires understanding a large codebase.
OpenAI has also introduced a new experimental mechanism in Codex for retaining and retrieving context during long-running work. OpenAI says it can be enabled in settings at launch and is expected to become the default for Astra.
From a developer’s point of view, the change is that you can hand off work in larger units. You can now more readily delegate a whole task—for example, “fix this bug and verify that the tests pass,” rather than only “fix this function.” That said, reviewing the resulting changes is still part of the job.
Documents, spreadsheets, and presentations
Astra is stronger at producing documents, spreadsheets, and presentations that follow an existing template. OpenAI highlights the model’s ability to match a given writing and visual style, drawing only on relevant context rather than repeating unnecessary information, and holding to the original goal while accommodating mid-course changes and added requirements.
The emphasis is no longer just on generating text, but on producing deliverables that require less rework before practical use. It does not mean the result can be used without review. A person still needs to verify the sources behind figures, organization-specific assumptions, and time-sensitive information.
A realistic division of labor is to supply an existing format and source material, let the model produce the draft, and spend your own time checking the substance.
Science and research
The tasks OpenAI cites include data analysis, simulation, model fitting (adapting a mathematical model to data), verifying data in specialized software, generating plots, and organizing what to investigate next.
The aim is to combine scientific reasoning with computer use to support the practical steps of research. It is not presented as automating scientific discovery or replacing the role of researchers.
Outside research, the same pattern applies to aggregating data an organization has accumulated, identifying trends, and suggesting which metrics to examine next.
4. Who Can Use It—ChatGPT, Work, Codex, and the API
The following table summarizes availability as of September 9, 2026.
| Where | Plans and users | Notes |
|---|---|---|
| Standard ChatGPT chat (GPT-6 Pro) | Pro ($100), Pro ($200), Business, Enterprise | Rolling out. Availability in standard chat is not announced for Plus |
| ChatGPT Work (GPT-6 Astra) | Plus, Pro, Business, Enterprise | Rolling out. Timing may differ from chat and Codex |
| Codex (GPT-6 Astra) | Plus, Pro, Business, Enterprise | Rolling out |
OpenAI API (gpt-6-astra) | API users | Usage-based billing, separate from a ChatGPT subscription |
| Microsoft Azure | Azure users | Availability announced. Check the cloud provider for regions and terms |
| Amazon Bedrock | AWS users | Availability announced. Check the cloud provider for regions and terms |
OpenAI positions standard chat for everyday conversation and questions, ChatGPT Work for long-running multi-step work and finished deliverables, and Codex for software development.
For Enterprise workspaces, actual availability also depends on the administrator’s model access settings. OpenAI has not announced standard Astra access for Free or Go. Because the rollout is staged, users on the same plan may receive access at different times across chat, Work, and Codex.
5. What It Costs
ChatGPT subscriptions and API usage are billed separately.
ChatGPT plans and where Astra is available
| Plan | Price (USD, monthly) | Where Astra / GPT-6 Pro is available |
|---|---|---|
| Plus | $20 | Rolling out in ChatGPT Work and Codex |
| Pro | $100 or $200 | Standard chat (GPT-6 Pro), Work, Codex |
| Business | Standard: $25 per user per month, or $20 per user per month with annual billing / Premium: $125 per user per month, or $100 per user per month with annual billing | Standard chat (GPT-6 Pro), Work, Codex |
| Enterprise | Custom | The above, subject to administrator model access settings |
A ChatGPT Business workspace requires a minimum of two paid seats, in any combination of Standard and Premium. The two Pro tiers share the same core features; the main difference is usage limits.
Prices outside the U.S. may vary because of taxes and currency conversion. Each plan has its own usage limits, and those limits may be set separately for ChatGPT Work, Codex, and standard chat. Limits change often, so check the current guidance.
API pricing
In the API, the model is offered as gpt-6-astra. Standard-tier pricing per 1M tokens is as follows.
| Category | Per 1M tokens |
|---|---|
| Input | $10 |
| Cached input | $1 |
| Cache write | $12.50 |
| Output | $50 |
Terms: tokens and the context window
A token is the unit a model uses to process text. A single character, part of a word, a whole word, or a punctuation mark can each be one token, so token counts do not match character counts. How many tokens a given text requires varies by language, by the text itself, and by the model.
The context window is the maximum number of tokens a model can consider at once in a single exchange. A larger window means more material can be handled together.
API pricing is based on these input and output token counts.
The API is billed by the tokens you use. A ChatGPT subscription does not include API usage; that is billed separately.
For requests with more than 272,000 input tokens, input and cache rates double and the output rate increases by 50% for the entire request. Batch and Fast mode have their own pricing. Using certain tools through the API, such as web search and computer use, can also incur tool charges on top of model token costs.
The main API specifications are as follows.
| Item | Value |
|---|---|
| Context window | 1,050,000 tokens |
| Maximum output | 128,000 tokens |
| Knowledge cutoff | April 30, 2026 |
| Input and output | Text input and output; image input supported |
| Not supported | Audio and video |
6. What “Critical” Means—And How It Relates to Everyday Use
At launch, OpenAI also disclosed that it had assessed Astra’s cybersecurity capability as reaching “Critical” under its own framework. The term sounds broad, but its scope is specific.
Terms: the Preparedness Framework and Critical
The Preparedness Framework is OpenAI's internal framework for assessing, by area, capabilities that could lead to severe harm.
Capability levels are organized around two thresholds: High and Critical.
In Astra's case, the Critical designation applies specifically to cybersecurity capability. It is not an assessment of the overall safety of Astra or ChatGPT.
Nor is it a certification by a third-party body.
One input to that assessment is an evaluation called ExploitBench. According to OpenAI’s announcement, in a research configuration without the protective layers used in production, Astra scored 100% and Sol scored 78.5%.
That does not mean attacks on real systems succeed 100% of the time when the released product is used normally. It measures the level of capability demonstrated under the evaluation conditions. The Critical determination also does not rest on that single score; it is presented as a judgment combining multiple evaluations.
OpenAI also reports that Astra improved over Sol on jailbreak robustness and on alignment (behaving within instructions and safety constraints). For the released version of Astra, OpenAI says the model declines advanced cybersecurity work such as writing proof-of-concept exploits for vulnerabilities, while supporting defensive uses such as secure code review and patch development.
What was reported about monitorability
The system card reports that Astra was harder to monitor than Sol when the monitor relied only on the model’s reasoning trace. OpenAI also evaluated monitoring based on actual inputs, outputs, and tool actions, and describes safeguards that do not rely on reasoning traces alone.
Term: chain-of-thought monitoring
This is monitoring that uses a model's intermediate reasoning as a signal for detecting signs of deviation.
It does not mean reading the model's internal state directly.
A complementary approach uses records of actual inputs, outputs, and tool actions as monitoring signals, and the two are combined in practice.
This is not a report of Astra evading monitoring on its own during normal use. Most of the monitor-evasion results come from adversarial evaluations in which the model was explicitly instructed to evade monitoring.
7. What Companies Should Check Before Switching to Astra
The following reflects MIF’s analysis.
To start with, a company using a finished product or API does not need to reproduce the kinds of controls OpenAI applies internally—encrypting model artifacts, isolating research environments. What carries over is not the individual measures but the sequence: when capability changes, review permissions and monitoring alongside it.
With that said, not every use requires the same depth of review. Drafting text or summarizing public information differs in likely impact from uses such as these:
- Connecting to confidential internal data
- Executing code
- Holding credentials
- Operating external systems
- Processing customer information
- Performing hard-to-undo actions such as publishing, deleting, sending, or paying
For the higher-impact uses, five questions are worth answering when switching models.
- What can the new model now do, and how has its behavior changed?
- Who gets access first, and for which tasks?
- Do existing data connections and operational privileges need to carry over as they are?
- Can tool calls, executed actions, and approval history be reviewed after the fact?
- If something goes wrong, who suspends use, revokes permissions, and rolls changes back?
These five questions are a practical starting point when switching models. They are not a checklist that guarantees safety. How tightly to limit scope and how long to keep the rollout limited depend on the work’s importance and risk profile, applicable regulations, and the level of disruption or loss the organization can tolerate. Add or adapt questions to fit your situation.
Switching to a new model does not mean redoing every prior evaluation. It is better treated as a checkpoint for confirming whether changes in capability or behavior affect your current permission, approval, and monitoring design.
8. Conclusion
A key advance in GPT-6 Astra is its ability to complete multi-step tasks using computer interfaces and tools. Compared with GPT-5.6 Sol, improvements were reported in computer use, workflow automation, coding, long-context handling, and scientific workflows.
Availability and pricing differ across standard ChatGPT, ChatGPT Work, Codex, and the API. ChatGPT subscription fees and API usage are billed separately.
“Critical” means that cybersecurity capability reached a threshold in an internal assessment; what it covers is limited to that area.
For higher-impact enterprise uses that involve connections to external systems or execution privileges, reviewing permissions, monitoring, and rollout plans at the time of a model switch can help identify and manage the risks introduced by the change.
About This Information
This article reflects information available as of September 9, 2026. The primary sources are OpenAI’s announcement “GPT-6 Astra: A new generation of intelligence,” the OpenAI Help Center article “GPT-5.6 and GPT-6 Pro in ChatGPT,” the GPT-6 Astra model page and API pricing page in OpenAI’s developer documentation, the ChatGPT release notes, and the GPT-6 Astra system card on the Deployment Safety Hub.
Benchmark figures are evaluation results published by OpenAI. MIF has not independently reproduced them.
The announcement page describes a rollout across paid plans including Plus. The plan-by-plan availability given in this article follows the Help Center guidance as of September 9, 2026, which distinguishes chat, Work, and Codex.
Availability, pricing, and usage limits may change. Because the rollout is staged, the same plan may gain access at different times.
The company-facing checkpoints reflect MIF’s analysis of the public materials, not OpenAI’s views.
Before making purchasing or deployment decisions, consult the provider’s latest public information.