OpenAI has launched GPT-6 Astra, a frontier model designed to move beyond answering questions and take on complex, multistep work across browsers, desktop applications, codebases, and professional software. Its benchmark results are striking, but the larger story is the shift from prompting an AI assistant to supervising an AI operator.
OpenAI introduced GPT-6 Astra on September 3, positioning it as the company’s most capable and aligned model to date. The initial rollout is limited, with access expected to expand over the coming days to ChatGPT Plus, Pro, Business, and Enterprise subscribers, as well as developers using the OpenAI API, Microsoft Azure, and Amazon Web Services Bedrock.
The company has not announced availability for free ChatGPT accounts.
While OpenAI’s launch materials emphasize gains across coding, science, cybersecurity, and professional work, computer use is the defining feature of the release. Astra is intended to operate software through the same interfaces people use—screens, browsers, forms, menus, spreadsheets, and other graphical tools—rather than depending entirely on purpose-built integrations.
From Answering Questions to Operating Software
OpenAI describes Astra as its strongest computer-use model so far. According to the company, it can fill out forms, update customer records, organize calendars, conduct web research, draft results into documents or email, analyze scientific data, create websites, and install or troubleshoot software.
Demonstrations accompanying the release show the model working in tools ranging from spreadsheets and Power BI to computer-aided design applications. Other examples include preparing legal documents, finding apartments, booking appointments, constructing websites, laying out circuit boards, and turning a basic visual concept into a playable 3D game.
The important change is not that Astra can use one particular application. It is that OpenAI is presenting general computer operation as an interface layer for AI.
For years, businesses have connected language models to internal systems through APIs, plugins, retrieval pipelines, and custom integrations. A sufficiently capable computer-use agent could perform some workflows through interfaces that already exist for human employees. That would not eliminate the need for secure APIs or enterprise integrations, but it could lower the engineering barrier for automating fragmented, legacy, or poorly connected software.
As VentureBeat reported, the practical transition is from employees prompting a chatbot at every step to people directing work at a higher level and reviewing the outcome.
Computer-Use Performance
On an offline subset of OSWorld 2.0, a benchmark that evaluates agents performing tasks in computer environments, Astra scored 72.6%. GPT-5.6 Sol scored 65.7% under the comparison reported by OpenAI.
OpenAI also says the Astra demonstration runs took approximately 40 minutes per task, compared with roughly 75 minutes for Sol—a reduction of about 47% in elapsed time. The company cautions that these times refer to specific demonstration runs rather than guaranteed production performance.
Astra reached 59.3% on Agents’ Last Exam, which covers professional tasks carried out in real software, and 92.7% on ScreenSpot-Pro without tools. OpenAI says the model and an updated Codex environment together completed Mind2Web tasks approximately 1.9 times faster than the previous GPT-5.6 Sol experience.
These results suggest that latency and task completion are becoming as important as raw reasoning scores. An agent that reaches the correct answer but takes an hour, repeatedly loses its place, or requires constant intervention may have limited operational value. Astra is designed to improve both capability and continuity across longer workflows.
Benchmark Results Show Broad Gains
OpenAI reported strong performance across professional work, coding, academic reasoning, and cybersecurity. Selected results include:
| Benchmark | GPT-6 Astra | Capability Evaluated |
|---|---|---|
| ARC-AGI-3 | 99.9% | Interactive reasoning and adaptation in unfamiliar environments |
| FrontierMath Tier 4 v2 | 97.6% | Advanced mathematical reasoning |
| GPQA Diamond | 96.0% | Graduate-level science questions |
| BenchCAD | 95.9% | Reconstructing 3D objects by generating CAD code |
| DeepSWE v1.1 | 74.1% | End-to-end software-engineering work |
| Terminal-Bench 4.0 | 57.9% | Complex terminal, configuration, coding, and data tasks |
| ExploitBench | 100% | Developing working exploits for known vulnerabilities |
The ARC-AGI-3 result requires particular care. OpenAI’s current launch page lists a score of 99.9%, while early briefing material reported by VentureBeat cited 98.6%. OpenAI also notes that Astra was evaluated with its Responses API harness and that published scores represent the maximum result at any tested reasoning effort.
That distinction matters because modern agent evaluations measure more than an isolated neural model. Memory, tools, system instructions, retry strategies, and the surrounding software harness can all materially affect performance.
For customers, the complete system may ultimately matter more than benchmark purity. For researchers trying to determine whether a model has achieved broadly general intelligence, however, separating the contribution of the underlying model from its tools and orchestration remains essential.
Does Astra Mark the Beginning of AGI?
OpenAI executives are using unusually consequential language around the release. Company president Greg Brockman told reporters that it was “not unreasonable” to regard the present period as the beginning of the artificial general intelligence era, according to WIRED.
That is an executive’s interpretation, not a settled scientific conclusion.
There is no universally accepted test for AGI. Strong results on ARC-AGI-3, coding evaluations, computer-use tasks, and graduate-level science questions demonstrate breadth, but they do not automatically establish that a system can outperform humans across most economically valuable work.
VentureBeat also identified a notable omission from the launch: OpenAI did not report an Astra result for GDPval, its benchmark for work products across a broad selection of occupations. GDPval itself has limitations because its current design focuses on one-shot assignments rather than extended interactive workflows. Nevertheless, the missing result leaves the broadest claims about Astra’s economic generality dependent on a collection of narrower benchmarks and demonstrations.
ITD Insight
Whether Astra meets a philosophical definition of AGI is less immediately important to enterprise buyers than whether it can finish real work reliably. The meaningful threshold will be sustained performance across messy production environments—including permissions, interruptions, incomplete information, and consequential actions—not a single near-perfect benchmark score.
Scientific Work and Software Engineering
Astra’s release also strengthens the case for AI as a research collaborator. OpenAI says versions of the model helped generate new results on long-standing problems in mathematics and theoretical computer science. The company subsequently prepared manuscripts and formalized the mathematical arguments as Lean certificates, providing a machine-checkable layer of verification.
In software engineering, Astra’s 74.1% DeepSWE score represents a more modest improvement over GPT-5.6 Sol’s 72.7% than some of its other benchmark jumps. Its 57.9% Terminal-Bench 4.0 result is more pronounced, compared with 37.3% for Sol in OpenAI’s evaluation.
The company is also introducing a Codex capability that allows Astra to retain notes and search earlier context after its active context window fills. If it works reliably, that could help address a persistent weakness in long-running coding agents: losing important requirements, failed approaches, and test results as a task grows.
Cybersecurity Is the Largest Deployment Risk
Astra is the first OpenAI model classified at the “Critical” cybersecurity level under the company’s Preparedness Framework. OpenAI says that, when given suitable tools and access, the model can discover previously unknown vulnerabilities and develop exploits against well-protected systems without continuous human direction.
The headline 100% ExploitBench result applies to known vulnerabilities and may be affected by exposure to historical information. On a newer internal benchmark built from recent V8 vulnerabilities, Astra achieved arbitrary code execution in 39% of cases, compared with 11.5% for GPT-5.6 Sol. OpenAI also says Astra discovered and used two previously unknown zero-day vulnerabilities during evaluation and is disclosing them to the affected maintainers.
These capabilities are valuable for defensive security teams, but they also raise the consequences of misuse, excessive permissions, or an agent operating beyond its authorized scope.
OpenAI says Astra is protected by multiple layers of safeguards, including alignment training, classifiers that inspect actions and reasoning for unauthorized behavior, and automated review systems capable of stopping a task. Legitimate cybersecurity work may occasionally be paused or blocked by those controls.
There is an unresolved technical concern as well. OpenAI found Astra’s written reasoning harder to monitor than that of GPT-5.6 Sol in tests designed to examine whether the model could conceal its intentions. Chief scientist Jakub Pachocki told WIRED that declining confidence in monitorability could eventually constrain further scaling.
Stargate and the Scale Behind Astra
OpenAI has not publicly disclosed Astra’s parameter count or a detailed model architecture. Claims that it uses a particular mixture-of-experts configuration or contains internally embedded sub-agents should therefore be treated as speculation unless the company publishes further technical documentation.
What has been disclosed is the scale of the training effort. VentureBeat reported that OpenAI researcher Aidan Clark described Astra as the company’s largest training run and the first OpenAI model pretrained using more than 100,000 data-center compute units at Stargate infrastructure. Clark also said earlier models played a substantial role in supervising the training of Astra.
This points to an increasingly important feedback loop: existing frontier systems are helping train, evaluate, and improve the systems that follow them. It also reinforces the strategic importance of Stargate as OpenAI attempts to secure enough computing capacity for larger training programs and widespread agent deployment.
Pricing and Availability
Developers will be able to access the model through the API under the name gpt-6-astra. Standard API pricing is listed at $10 per million input tokens and $50 per million output tokens, with separate prices for cache operations.
A Fast processing mode is expected to offer up to twice the Standard speed at twice the price. OpenAI also says Astra supports Zero Data Retention for eligible API customers and that it is testing a Private Safety Processing system intended to combine security monitoring with stronger privacy protections.
ChatGPT Pro, Business, and Enterprise customers will receive access to GPT-6 Astra Pro in addition to the standard model. Enterprise administrators must enable Astra for their workspaces because access is disabled by default at launch.
The Bottom Line
GPT-6 Astra is a consequential release even if the AGI label remains disputed. Its strongest claim is not that it performs well on another set of question-and-answer tests, but that it can navigate software, maintain direction across longer assignments, and turn high-level instructions into completed digital work.
The launch also exposes the central tension of increasingly autonomous AI. The capabilities that make Astra useful—computer access, extended execution, coding ability, and cybersecurity expertise—are the same capabilities that demand stronger permissions, monitoring, auditing, and human oversight.
For organizations evaluating Astra, the decisive questions will be operational: how often it completes real workflows correctly, how much supervision it requires, how safely it handles access to sensitive systems, and whether its higher per-token price produces a lower total cost per successful task.
If Astra performs outside controlled demonstrations as OpenAI says it does, the release could mark a genuine shift in enterprise computing. The user’s role would move further away from operating every piece of software directly and closer to defining objectives, setting boundaries, and reviewing the work of an AI agent.
Sources: OpenAI, VentureBeat, and WIRED.
