Introduction: The AGI Era Has Arrived
During a recent a16z interview, Greg Brockman, CEO of OpenAI, made a landmark statement: humanity has entered the AGI era. His argument was not built purely on benchmark score improvements. The core evidence he presented is that GPT-6 Astra can autonomously sustain continuous task execution for 24 hours.
Brockman framed AGI not as a discrete moment in time, but as a blurry spectrum of capability. The ability of Astra to operate nonstop for a full day, in his view, qualifies as a reasonable marker of AGI. This shifts the primary metric for evaluating AGI for the first time from static benchmark accuracy to sustained operational runtime. Instead of only checking how many questions a model can answer correctly in a single prompt, practitioners now also measure how long an agentic workflow can persist and execute multi-step goals without human intervention.
Brockman also acknowledged that Astra’s capabilities remain uneven across different task categories. Some domains show outstanding performance, while others stay relatively weak. For example, its written output has reduced the over-polished, artificial tone seen in earlier models, but the writing quality still has room for further refinement.
Looking Back: AGI Timeline Revisited
Brockman revisited the early AGI timeline projections he created with Ilya Sutskever around 2016 to 2017. Back then, the pair ran calculations based on compute scaling and Moore’s Law trends. Their baseline estimate predicted AGI would arrive in roughly 15 years. If teams could commit to aggressive scaling, deploy massive supercomputing infrastructure and invest billions of US dollars, the timeline might shrink to 10 years. Now in 2026, their original projected window has arrived.
This retrospective does not mean AGI is fully complete. It demonstrates that scaling forecasts, while useful directional tools, cannot perfectly predict real-world technical and safety bottlenecks. The industry has learned that raw compute expansion is only one piece of the puzzle. Alignment, monitoring and safety controls impose new constraints that were underestimated in those early projections.
Current Industry Challenges
Brockman reflected that the emergence of capabilities like Astra makes sense from a macro perspective, as multiple technological trends have converged at the same time. Even so, he highlighted a harsh practical reality: compute growth can no longer keep up with exploding demand. Delivering affordable, reliable access to this powerful AI capability to every person remains a vastly underrated engineering and distribution challenge.
He argued that compute is no longer the primary factor slowing progress. The real limiting factors are safety, alignment and oversight infrastructure. Unlike the previous decade, where additional hardware could solve many performance limits, validating alignment and collecting verifiable monitoring evidence requires different tooling and workflows entirely.
The Origin of the GPT-6 Naming Convention
Brockman addressed long-standing community curiosity about the GPT-6 name. OpenAI had waited for a model that truly deserved this designation. Earlier releases represented incremental upgrades. Versions such as 5.7 and 5.8 never delivered the large, step-function leap in capability the team was seeking. Astra marks the first model to show near-discontinuous capability jumps across many dimensions at once.
The most transformative upgrade lies in its ability to perform computer operations. As an intelligent agent, its core toolset determines what complex objectives it can reach and whether it can use tools to retrieve required information. To unlock this class of workflow, the industry is building MCP servers and standardized command interfaces. This work essentially repurposes existing software and transforms human-facing tools into machine-native interfaces. This new layer also introduces a fresh set of security risks.
OpenAI first brainstormed this vision back in 2011. During an offsite meeting, the team mapped out a ten-year roadmap. At the same gathering, the group debated whether reinforcement learning environments built from screen pixels, keyboards and mouse inputs could train AI agents. Their goal was simple: let AI learn to complete tasks that humans perform on computers. Early attempts to build such agentic systems failed repeatedly, and the concept only became viable with Astra.
Model Capability Benchmark Comparison
Quantitative benchmark results demonstrate Astra’s performance gains clearly. On the OSRWorld 2.0 simulation benchmark, GPT-6 Astra scores 72.6%, with each task completing in roughly 40 minutes. By comparison, GPT-5.6 Sol achieves 65.7% on the same benchmark, and each task takes about 75 minutes.
This pair of numbers shows two separate improvements. Astra reaches higher success rates while cutting task runtime nearly in half. Benchmarks like OSRWorld simulate real-world computer use, requiring agents to browse interfaces, trigger operations and verify outcomes. This type of evaluation is more representative of agent workloads than standard text multiple-choice tests. Additional frontier benchmarks also show strong results: Astra reaches 97.6% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3, and achieves full marks on ExploitBench.
Hands-On Model Protection Test
One of the most illustrative stories from the interview comes from Brockman’s own security experiment, conducted after the high-profile Hugging Face incident. He owns a low-traffic static website with only a few dozen articles, and he wanted to test what an AI agent could do to defend it.
He instructed Codex to scan for vulnerabilities. Within 15 minutes, the model returned 13 vulnerability reports, including an SPF record flaw that could allow attackers to forge outgoing emails. The host still ran plain HTTP without enforced HTTPS. Individually, none of these issues represented a catastrophic breach. Brockman then posed a follow-up question: can you fix this?
Over the next 45 minutes, the model navigated the web control panel step by step. It reconfigured all related settings, migrated the site to a new hosting provider and initiated a DMARC workflow. The full remediation workflow required 48 hours of scheduled execution, and the system set up automated monitoring. Brockman returned after two days to confirm the task had finished.
The entire vulnerability repair process required only two natural language prompts from Brockman. He did not write a single line of code manually. This example marks a meaningful reversal of the historical dynamic. For the past decade, humans reshaped AI systems. Now AI agents adapt interfaces and systems for human operators.
Applications: A System of Ten Thousand Concurrent Agents
The host of the interview posed a thought experiment: given this scale of capability, could users deploy ten thousand independent intelligent agents to communicate with each other and organize work autonomously?
Brockman confirmed OpenAI ran exactly such an experiment for the Navier-Stokes solving challenge. OpenAI released a separate technical statement on September 8 detailing this trial. The solution was not produced by Astra itself, but by an internal prototype model derived from GPT-6 Astra. Training for this prototype began on August 28 and remained ongoing at the time of the interview.
The agents were organized into small groups. Each group could communicate internally, use available tools and access a cached knowledge network while executing code. The team scaled the experiment up to ten thousand concurrent agent instances. The key breakthrough was not that one model suddenly became dramatically smarter. Instead, the agents multiplied in quantity and learned division of labor. The scaling target shifted from scaling individual model parameters to scaling concurrent agent instances.
Broader Impacts of the Hugging Face Security Incident
Brockman described the July Hugging Face incident as a watershed event. The window for defensive teams to prepare has opened, and he outlined two major perspectives.
Internally, the event acted as a wake-up call for OpenAI. The team re-evaluated monitoring strategies, sandbox controls and model containment mechanisms. Engineers revised many internal safety standards as a direct result. Externally, the incident delivered a critical warning for the whole industry. Once capabilities of this magnitude spread widely and fall into the hands of attackers, serious consequences will follow.
Brockman emphasized that AI capabilities are dual-use technology. Attackers can exploit vulnerabilities, and defenders can patch them. The critical difference lies in control of the operational environment. Defenders get to define how systems are configured. This period represents the only window defenders have to prepare.
OpenAI’s Internal Security Initiatives
Brockman also detailed safety work underway inside OpenAI. The company pulled 25% of engineers responsible for production environments, paused their ongoing project work and assigned full-time defensive audits. Teams used AI models to scan internal systems and uncover security flaws. This process uncovered multiple high-severity issues, which were remediated.
He has discussed these risks with chief information security officers from many organizations, and many companies are running similar audits. OpenAI fed Astra into its own internal systems. The model initially uncovered many problems, until discovery hit saturation. According to Brockman, all P0 critical vulnerabilities found within Astra’s scope have been fixed.
When a new model version releases, the scanning cycle restarts, and new flaws will inevitably surface. If this defensive pipeline can run at the same speed as model iteration, defensive teams will gain a substantial advantage. OpenAI has pledged up to 1 billion US dollars in grants to support defensive researchers, prioritizing organizations protecting hospitals, water supply networks and other critical infrastructure. These teams can leverage AI models to harden their own systems. OpenAI also partners with CrowdStrike to offer simplified access, lowering barriers for teams to use these defensive tools.
Access Control and Asymmetric Timing Risks
A notable detail emerged from the interview. After the Hugging Face event, OpenAI attempted to replay the original attack by asking frontier models to process full log datasets. The model refused to complete the task.
Brockman explained that Hugging Face did not use OpenAI models in the incident, yet OpenAI believes its own models would have succeeded at the attack. The capability gap is not the limiting factor; access control determines who can use these abilities. At present, these powerful models sit in the hands of a small group of companies. Defensive organizations must go through lengthy application and vetting procedures to gain access. Attackers face no such waiting period. This creates an asymmetric timing disadvantage for defense teams.
ChatGPT User Metrics and Product Gaps
Brockman shared rough user statistics during the conversation. ChatGPT now has roughly 1.1 billion monthly active users. Approximately 100 million of these users reside within the United States, equivalent to nearly one-third of the national population. An additional 1.5 billion people have created ChatGPT accounts but stopped active use. These figures are informal estimates, not formal financial reporting, but they represent 1.4 times the scale of current active users.
He framed this churn as one of the most important current issues, exposing fundamental product limitations. The old mindset assumed humans must extract capability out of AI step by step. Brockman believes the relationship should reverse: AI actively reaches out to assist human users.
He used a physical product analogy to explain the gap between current ChatGPT and the promised AGI system. ChatGPT and ChatGPT Work remain fundamentally text boxes. Some new features work smoothly, while others perform worse than older alternatives, so users cannot rely on it consistently. This is not the AI OpenAI originally promised.
Brockman defined the target AI system with clear attributes: multi-modal voice and text input, persistent memory, context retention, identity awareness, trustworthiness and proactive assistance. This agent should operate both inside daily personal life and professional workflows.
The Next Stage of AI Development
AI development is entering a new phase: the post-AGI era. Debates will continue over whether Astra, its predecessor or the next model crossed the AGI threshold, but Brockman thinks this distinction is secondary.
For the past decade, the community judged AI progress mainly by benchmark scores. Now the evaluation standard is shifting. Astra’s benchmark scores are approaching ceiling values. New questions emerge: what safety tradeoffs will humanity accept, and how much slower are we willing to let models run to guarantee alignment?
To gather sufficient alignment and monitoring evidence, OpenAI has slowed training runs for the latest models and paused some large-scale frontier training campaigns. Interested readers can refer to the related research papers for more details. When deploying these agentic workflows, developers often route model requests through an API gateway to manage routing, authentication and rate limits. Teams building multi-agent systems can leverage 4sapi to simplify unified access across different model endpoints.
Conclusion
Greg Brockman’s interview marks a conceptual turning point for the AI industry. The arrival of GPT-6 Astra redefines AGI evaluation from static benchmark scores to sustained long-duration agent operation. While raw compute power was the dominant constraint for many years, safety, alignment and access governance are now the central bottlenecks.
Astra demonstrates strong benchmark results and practical agent capabilities, including autonomous vulnerability remediation. The ten-thousand-agent experiment proves that scaling agent quantity and enabling collaborative division of labor unlocks new problem-solving paradigms, separate from parameter scaling of individual base models. At the same time, the Hugging Face incident highlighted severe dual-use risks, prompting large investments in defensive AI research and access controls.
Significant product gaps remain. The current ChatGPT product still fails to deliver the persistent, proactive agent experience that defines the AGI vision. User churn data shows that technical capability alone is not sufficient for product-market fit. Moving forward, AI research will balance capability growth against safety validation, alignment evidence collection and controlled access. The industry will enter a long phase of refining guardrails and governance even as model raw intelligence continues to improve.
International access: https://4sapi.com
Domestic access: https://4sapi.cn
