“I think it’s not unreasonable to feel that we are now in the AGI era.” – Greg Brockman – OpenAI President
Declaring a new technological era is always a move loaded with strategic intent, not just descriptive bravado. When a leading AI lab claims that we now inhabit an era defined by artificial general intelligence, it is signalling a shift in how capabilities, risks, and responsibilities should be understood, allocated, and governed across the wider digital economy and society at large 1,4,5,15. The underlying issue is whether a class of systems that were once framed as experimental tools have now crossed a threshold where they function as general-purpose, economically central infrastructure that rivals human cognitive labour across a broad spectrum of tasks.
The factual backdrop to this claim revolves around the launch of GPT-6 Astra, described as a frontier system trained on a massively scaled infrastructure and marketed as a generational leap in capability 1,4,12,15. Reports indicate that Astra was built using more than 100 000 GPUs at OpenAI’s Stargate site in Texas, suggesting a training regime significantly beyond earlier flagship models in terms of compute intensity and architectural ambition 12. Benchmarks highlight superhuman or near-ceiling performance on specialist evaluations such as FrontierMath Tier 4 v2, DeepSWE, GPQA Diamond, ExploitBench and ARC-AGI-3, metrics designed to stress-test mathematical reasoning, software engineering, scientific question answering and exploit discovery 4. These numbers are not neutral diagnostic curiosities; they are used rhetorically to argue that a model once considered a text interface has become a general problem-solving engine capable of tackling tasks previously reserved for high-end human experts 2,3,4.
In parallel, Astra is positioned as a step change in what work can be delegated to AI systems, especially through features described as computer use or agentic operation, where the system can operate across a user’s inbox, calendar, enterprise applications and development environments without the user manually driving each interaction 2,4,9,15. The shift from static query-response chat to persistent, autonomous task execution reframes these systems from tools to semi-autonomous agents embedded in workflows, effectively outsourcing large fractions of digital cognitive labour. Brockman has emphasised that the gating factors for such delegation are no longer raw model capability but trust, compute and domain depth, underscoring a view that the technical ceiling is high enough that non-technical constraints now dominate adoption decisions 9,14. In this framing, the era question is less about whether an abstract AGI definition has been formally met and more about whether economic and organisational structures must now assume the continual presence of highly capable, generalist agents as baseline infrastructure.
Defining AGI and the move from moment to continuum
Artificial general intelligence has historically been defined by OpenAI as highly autonomous systems that outperform humans at most economically valuable work, a standard that appears to demand broad, cross-domain, superhuman competence rather than narrow specialist brilliance 1,2,4,15. Yet Brockman and other OpenAI leaders have increasingly described AGI as a gradual, jagged phenomenon rather than a crisp milestone: intelligence that is superhuman in some domains and subhuman in others, with the overall system evolving along a continuum rather than snapping into qualitative existence at a specific date 3,7,11,14. The idea of jagged intelligence captures this: current models can be superhuman at coding or advanced mathematics, while still failing at simple commonsense tasks or brittle edge cases that any human generalist would navigate with ease 7,11,14. In interviews earlier in 2026, Brockman suggested he was 70 to 80 percent of the way to what he would personally label AGI and predicted that AGI in a jagged form would arrive within a few years, making almost any intellectual task performed via computers delegable to AI agents 7,10,11. The more recent assertion that it is not unreasonable to feel we are now in the AGI era continues this logic by treating the era as the onset of pervasive, generalist capability rather than a single formal threshold test 4,5,15.
Strategic tension: capability claims versus evaluative uncertainty
The declaration of an AGI era exposes a deep tension between marketing, internal belief, and external verification. On the one hand, OpenAI and allied commentators highlight striking capabilities: Astra reportedly solved a set of long-standing mathematical problems at low compute cost, can prove formal safety properties of software, and approaches what they describe as cyber-critical capability, where evaluations begin to matter for national security and critical infrastructure 3,4,12. On the other hand, Brockman has repeatedly acknowledged that there is no agreed test for AGI and that the term has shifted from a contractual trigger in earlier arrangements with partners to something closer to a mission concept or spiritual label, a narrative banner rather than a regulatory threshold 5,6,8. This reframing allows strong capability claims without committing to a precise, externally auditable criterion. It also means that when he says future observers might look back and judge AGI as having emerged around Astra, he is offering a speculative historical framing rather than an empirical measurement that could be independently reproduced 2,3,5,8,15.
The enterprise and societal implications of treating Astra-class models as AGI-era systems are considerable. From an enterprise viewpoint, Astra is pitched as the foundation for a new way of computing in which users no longer need to orchestrate their own interactions via classical interfaces; instead, they instruct agents that autonomously navigate applications, compose code, draft presentations, manage documents and coordinate workflows 4,9,14,15. Brockman has described the emerging paradigm as one where individuals become de facto chief executives of personal corporations, orchestrating fleets of agents that perform knowledge work across domains 3,9. If this framing holds, the AGI era implies a reorganisation of white-collar labour and software design around delegation to persistent AI entities rather than discrete tool use. However, this narrative is contested by experts in cognitive science and AI safety who argue that present systems lack robust world models, grounded understanding, and persistent self-regulation, and therefore should not be equated with human-like general intelligence despite impressive benchmark scores 11. For these critics, the AGI era claim risks conflating multi-task automation with genuine general intelligence, potentially distorting policy and public expectations.
Debates, objections, and the problem of jagged performance
Objections centre on two main axes: conceptual clarity and practical reliability. Conceptually, some researchers argue that defining AGI purely in economic terms, as outperforming humans on most economically valuable tasks, sidelines key dimensions such as consciousness, intentionality, and robust causal understanding, which many consider essential to any serious use of the term general intelligence 11,14. Practically, jagged performance undermines confidence that these systems can be treated as general agents: being superhuman on formal mathematics yet brittle on everyday reasoning creates complex risk profiles, particularly when such systems are tasked with safety-critical operations or given broad autonomy over digital environments 3,11,14. Brockman acknowledges this jaggedness but maintains that trajectory and line of sight are the important factors: he regards it as extremely clear that AGI in a jagged but economically transformative form will be achieved within a short horizon and increasingly treats current models as occupying the early phase of this takeoff 7,10,11,14. Critics respond that line of sight from current benchmarks to robust general intelligence is far from secure and could be derailed by limits in data, compute, interpretability, or alignment, making era declarations premature or strategically motivated.
Why the AGI era claim matters for governance and public trust
The assertion that we now live in an AGI era matters not just for branding but for how regulators, enterprises and citizens calibrate their expectations of AI systems and the organisations that deploy them. If Astra-class models are framed as AGI-era systems, regulators may feel pressure to sharpen frameworks for frontier model oversight, safety evaluation, and incident reporting, particularly as capabilities inch towards cyber-critical thresholds where misuse could have systemic consequences 3,4,12,14. At the same time, Brockman emphasises that capability is only part of the story; connecting models to the real world to deliver value, maintaining human control over goals, and ensuring AI expands rather than constrains human capabilities are presented as central challenges 14. Trust emerges as the defining feature of what some describe as the agentic era: users must be confident that persistent agents acting on their behalf will behave reliably, respect constraints, and avoid harmful actions, despite operating with capabilities that may surpass human experts on narrow tasks 9,14. Declaring an AGI era without corresponding, demonstrably effective governance mechanisms could therefore erode public trust, particularly if systems fail in unexpected ways.
Reading the statement as a marker of trajectory rather than destination
Interpreting the claim that it is not unreasonable to feel we are in the AGI era as a trajectory marker rather than a definitive announcement helps reconcile internal optimism with external scepticism. Brockman himself has shifted from earlier comments that we are 70 to 80 percent of the way to AGI to a view that present models occupy the early phase of an AGI-era takeoff, with further models arriving in the same year expected to surpass Astra in capability 7,10,12. He stresses that AGI should be understood as a spectrum of capability and impact, and that what matters more than labels is the actual work these systems can perform reliably in practice 3,14,15. From this perspective, the statement functions as a strategic signal that frontier models have crossed a threshold of versatility and autonomy substantial enough to require a new organising concept, even as technical, philosophical and regulatory debates over AGI remain unresolved. Whether wider expert communities accept this framing will depend on how Astra and its successors perform in real-world deployment, how jaggedness is reduced, and whether governance regimes evolve to handle systems that increasingly resemble general-purpose cognitive infrastructure.
References
1. OpenAI unveils GPT-6 Astra with major advances in AI … – 2026-09-03 – https://www.foxbusiness.com/technology/openai-unveils-gpt-6-astra-major-advances-ai-capabilities
2. OpenAI does a victory lap for its new AI model: ‘Welcome to the AGI era’ – 2026-09-03 – https://africa.businessinsider.com/news/openai-does-a-victory-lap-for-its-new-ai-model-welcome-to-the-agi-era/zcqpxbz
3. How Close Are We to True Artificial General Intelligence … – 2026-09-03 – https://thetranscriptdesk.substack.com/p/how-close-are-we-to-true-artificial
4. ‘Welcome to the AGI era’: OpenAI launches GPT-6 Astra – 2026-09-03 – https://venturebeat.com/technology/welcome-to-the-agi-era-openai-launches-gpt-6-astra
5. OpenAI’s AGI Eras Tour – Spyglass – 2026-09-03 – https://spyglass.org/agi-2026/
6. OpenAI spuckt grosse Töne: «Willkommen in der AGI-Ära» – 2026-09-03 – https://www.blick.ch/digital/openai-stellt-neues-ki-modell-vor-und-spuckt-grosse-toene-willkommen-in-der-agi-aera-id22232406.html
7. OpenAI President Greg Brockman: Doubling Down on Text Models, The Superapp Plan, Codex’s Potential – 2026-04-07 – https://www.bigtechnology.com/p/openai-president-greg-brockman-doubling
8. OpenAI President Declares The “AGI Era” Has Officially Begun – 2026-09-03 – https://www.zerohedge.com/ai/openai-declares-agi-era-has-officially-begun
9. OpenAI is building a personal AGI you talk to like a coworker and … – 2026-07-27 – https://x.com/gokulr/status/2081747011431473629
10. OPENAI PRESIDENT GREG BROCKMAN ON WHEN WE … – 2026-04-02 – https://x.com/ChrissGPT/status/2039568947909963874
11. OpenAI says 70% to AGI. A prominent cognitive scientist says we’re nowhere close – 2026-04-02 – https://www.rdworldonline.com/openai-says-70-to-agi-a-prominent-cognitive-scientist-says-were-nowhere-close/
12. Techmeme – 2026-09-04 – https://www.techmeme.com/
13. OpenAI Stock Price, Valuation & News – https://stockanalysis.com/private/openai/
14. “Now the most important challenge is to connect AI models with the world.”Greg Brockman, president o.. – 2026-06-16 – https://www.mk.co.kr/en/it/12075925
15. OpenAI launches GPT-6 Astra, its most powerful model yet … – 2026-09-03 – https://fortune.com/2026/09/03/openai-debuts-gpt-6-astra-computer-use-greg-brockman-says-start-of-agi/
