OpenAI Launches GPT-6 Astra and Says the AGI Era Has Begun

OpenAI has launched GPT-6 Astra, a new frontier AI model that the company says can use computers, reason through complex problems, and carry out multi-step tasks with enough autonomy to mark the beginning of the artificial general intelligence era. President Greg Brockman said on September 3, 2026, that “it’s not unreasonable to feel that we are now in the AGI era,” citing Astra’s advanced capabilities in computer use, reasoning, and autonomy. The model is initially available to select organizations through the Daybreak early access program, with broader access rolling out to ChatGPT Plus, Pro, Business, and Enterprise users, as well as via the OpenAI API, Microsoft Azure, and Amazon Bedrock. OpenAI has not confirmed availability for free ChatGPT users.

The announcement moves the AGI debate from speculation to product claims. If Astra’s capabilities hold up outside controlled benchmarks, enterprises could deploy AI agents that handle software, research, and coding workflows with far less human supervision. But the same autonomy raises urgent questions about security, cost, and whether organizations can monitor systems that reason in ways humans find harder to follow. The staggered rollout through Daybreak and paid tiers suggests OpenAI is initially limiting exposure while it gathers feedback from a narrower set of users.

Benchmarks show a measurable jump over previous models

OpenAI released performance figures that show Astra outperforming GPT-5.6 Sol, its previous frontier model, across several demanding tests. The gains are substantial in areas that matter for autonomous work. On FrontierMath Tier 4, a difficult mathematics benchmark, Astra scored 97.6%, compared with 83% for GPT-5.6 Sol. On GPQA Diamond, a graduate-level science benchmark, Astra reached 96%. In agentic tasks, Astra scored 59.3% on Agents’ Last Exam, ahead of GPT-5.6 Sol at 53.6% and Claude Opus 5 at 55.5%. On OSWorld 2.0, a benchmark for real computer-use tasks, Astra completed tasks in about 40 minutes, roughly half the 75 minutes needed by GPT-5.6 Sol.

These numbers suggest not just incremental improvement but a step change in speed and reliability for autonomous computer use. The OSWorld result is especially significant for enterprise automation, because it measures the ability to complete real tasks in actual software environments rather than answer isolated questions. A model that finishes workflows in 40 minutes instead of 75 could change the economics of AI agents, reducing the time and compute required for real-world computer tasks. The FrontierMath and GPQA results also indicate stronger reasoning in formal and scientific domains, which may translate into better performance on technical work that requires multi-step deduction.

Cybersecurity: OpenAI says Astra reached its “Critical” threshold

OpenAI said Astra has reached the company’s “Critical” threshold in cybersecurity, a designation that signals the model could meaningfully assist or accelerate offensive cyber operations. In testing, Astra achieved 100% on ExploitBench, compared with 78.5% for GPT-5.6 Sol, and discovered two previously unknown zero-day vulnerabilities. Because of those dual-use risks, the released version of Astra refuses requests to create exploit proofs, while defensive tasks such as secure code review remain enabled.

OpenAI is committing $1 billion in subsidized access for cybersecurity defenders, an attempt to tilt the model’s deployment toward defense. The company did not specify the timeframe or exact mechanism for that subsidy, but the figure underscores how seriously it treats the risk. Security researchers are likely to scrutinize whether the refusal mechanism can be bypassed and whether the model’s offensive knowledge leaks through indirect prompts. The combination of a perfect ExploitBench score and the discovery of two zero-days means Astra’s release will be watched closely by both defenders and attackers, even with the restrictions in place.

Safety and alignment: fewer failures, but harder to monitor

OpenAI’s safety evaluations show a sharp reduction in misaligned outcomes for computer-use tasks. Astra had a 2.4% misaligned outcome rate, down from 22% for GPT-5.6 Sol. That is a substantial improvement, but OpenAI’s chief scientist Jakub Pachocki cautioned that monitoring Astra’s internal reasoning is more challenging than for previous models, raising concerns about long-term alignment. A model that fails less often but is harder to interpret may still pose risks if its reasoning drifts in ways that are not immediately visible to human overseers.

“It’s not unreasonable to feel that we are now in the AGI era.” — Greg Brockman, OpenAI President

Critics, including independent researchers, point to Astra’s tendency to over-engineer solutions and consume excessive tokens during research. That behavior could increase operational costs even if the model completes tasks correctly. The tension between lower misalignment rates and harder interpretability means safety teams may need new tools to audit long-horizon autonomous behavior, especially when Astra is given access to sensitive systems. For enterprises, the practical question is whether a 2.4% misaligned outcome rate is acceptable when the cost of monitoring and token consumption is high.

What it means for professionals and enterprises

For professionals, Astra represents a leap in AI autonomy. It can execute multi-step workflows across software, research, and coding environments, which could transform productivity in fields ranging from software development to financial analysis. Its integration into enterprise tools through Azure and Bedrock means organizations can deploy it within existing cloud infrastructure, but they will need to weigh cost, safety, and monitorability. The model’s ability to complete OSWorld tasks in about 40 minutes suggests that routine computer-based workflows could be automated more aggressively, but the over-engineering tendency noted by researchers may erode some of those efficiency gains.

Access is not yet universal. The Daybreak early access program suggests OpenAI is initially limiting deployment to select organizations, likely to gather feedback and control risk before a wider release. The lack of confirmation for free ChatGPT users also means the model’s most advanced capabilities will first reach paying customers and enterprise partners. For global businesses, the key question is not whether Astra can perform tasks, but whether the cost of running an autonomous model that over-engineers solutions will be justified by the productivity gains. The $1 billion subsidy for cybersecurity defenders is a notable exception, but it does not address the broader cost concerns for commercial deployments.