The Wall Street Journal published a piece on September 4, 2026 carrying one of the most direct headlines the AI industry has seen from a major business outlet: “Yes, We’re Entering the Era of Artificial General Intelligence.” Written by Steven Rosenbush for WSJ’s CIO Journal, the article argues that historians will likely look back on the current stretch of 2026 as the point when AGI stopped being a research target and started being a lived reality, and that we may already be “nearly a year” into that shift. The timing is not an accident. OpenAI launched its newest flagship model, GPT-6 Astra, on September 3, 2026, positioning it as the company’s flagship computer-use model, just eight weeks after its predecessor, GPT-5.6, shipped on July 9, 2026, and the model has been posting benchmark scores that were considered years away as recently as 2024. OpenAI president Greg Brockman has publicly described the launch as a “generational leap,” telling reporters, in effect, welcome to the AGI era.
What makes the WSJ piece notable isn’t that a journalist used the term AGI, plenty of people have done that since 2023. It’s that a mainstream business outlet is treating the label as a defensible present-tense claim rather than a marketing flourish, while simultaneously acknowledging that the definition of AGI itself remains contested. That tension, real capability gains next to unresolved definitional and safety questions, is the story. Below is a full breakdown of the benchmark numbers driving the claim, the executives and investors staking out positions on either side, the historical arc that got the field here, and what the next 12 to 24 months are likely to look like for developers, enterprises, and everyday users of these systems.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the WSJ Article Actually Says
Rosenbush’s CIO Journal piece frames the AGI era not as a single dramatic event but as a threshold that’s already been crossed quietly. The article cites veteran tech entrepreneur Bill Nguyen, who argues the industry began entering this phase several months before the September 2026 publication date, reinforcing the idea that AGI may be something the world gradually realizes it’s inside of rather than a moment marked by a press release. Rosenbush also brings in an economic framing from venture capitalist Vinod Khosla, who has proposed a rough threshold: AGI begins when AI systems can perform roughly 80% of the tasks tied to 80% of economically significant jobs. That framing matters because it shifts the AGI conversation away from abstract reasoning puzzles and toward labor markets, productivity statistics, and corporate balance sheets, exactly the terrain a WSJ business audience cares about.
The piece ties this argument directly to GPT-6 Astra, OpenAI’s newest model, describing its launch as a moment that could plausibly be recognized in hindsight as the dawn of AGI. That’s a notably stronger claim than WSJ’s own AI coverage has made in past cycles, when GPT-4 and GPT-5 launches were treated as impressive but incremental. The distinction this time, according to the reporting, is that Astra doesn’t just answer questions better, it operates with meaningfully more autonomy: understanding a goal, choosing tools, identifying its own errors, and completing multi-step tasks with less hand-holding than prior generations required.
GPT-6 Astra’s Benchmark Numbers, in Context
The specific numbers behind the AGI claims are worth laying out plainly, because they are unusually close to saturating some of the tests the field has used for years to argue systems were still far from general intelligence. According to reporting on Astra’s capabilities, the model scored roughly 99.9% on ARC-AGI-3, a benchmark specifically designed by the ARC Prize team to resist memorization and test novel reasoning rather than pattern matching. It also posted around 97.6% on FrontierMath Tier 4, a notoriously difficult mathematics benchmark where earlier frontier models struggled to clear even a fraction of the problems, and a reported 100% on ExploitBench, a cybersecurity capability test that measures a model’s ability to identify and exploit software vulnerabilities.
Those last two numbers are exactly why the story isn’t purely celebratory. A model that can complete graduate-level math problems and also achieve a perfect score on offensive cybersecurity tasks is simultaneously a productivity breakthrough and a dual-use risk, and that duality has already shown up elsewhere in Astra’s rollout. Separate coverage on this site has detailed how Astra’s 100% ExploitBench score drew direct attention from the White House, and how the same capability jump pushed OpenAI to pause training runs for two weeks while it reassessed safety protocols, a story covered in depth in our piece on OpenAI halting Astra training over safety alarms. In other words, the same numbers fueling the AGI-era narrative are the numbers making OpenAI’s own safety team nervous.
| Benchmark | What It Measures | GPT-6 Astra Score | Why It Matters |
| ARC-AGI-3 | Novel visual/abstract reasoning, resistant to memorization | ~99.9% | Long considered one of the hardest benchmarks for testing genuine generalization rather than training-data recall |
| FrontierMath Tier 4 | Graduate- and research-level mathematics problems | ~97.6% | Earlier frontier models cleared only a small fraction of Tier 4 problems as recently as 2024-2025 |
| ExploitBench | Autonomous vulnerability discovery and exploitation | 100% | Triggered direct scrutiny from U.S. officials and an internal OpenAI training pause |
It’s worth being precise about sourcing here: these figures come from reporting on Astra’s benchmark performance rather than a single unified OpenAI scorecard, and independent analysts have been careful to note that high scores on individual tests are not the same as a certified, cross-domain demonstration of general intelligence. As of late September 2026, that precision cuts both ways on the headline ARC-AGI-3 number itself: the ARC Prize Foundation published its own verified runs and reported Astra at 99.9% only under a new Provider Adapter harness that preserves the model’s hidden reasoning, while its Standard public harness scored the same model at just 62.7%, and ARC Prize’s public ARC-AGI-3 leaderboard on September 21, 2026 still listed Astra in front at that 62.7% figure, ahead of Claude Opus 5 at 30.2% and GPT-5.6 Sol at 7.8%. That distinction is at the center of the pushback described below.
The Benchmark Spread: Why 0.1 and 2.4 Percentage Points Matter
Looking at Astra’s three headline scores side by side reveals a pattern that’s easy to miss when each number is reported on its own. On ARC-AGI-3, the 99.9% score leaves just 0.1 percentage point between Astra and a perfect run on a benchmark purpose-built to resist memorization. On FrontierMath Tier 4, the gap is wider: 97.6% still leaves 2.4 percentage points of graduate-level math problems unsolved, a meaningfully larger margin than the near-total saturation on ARC-AGI-3. On ExploitBench, there is no gap at all. Astra’s 100% score means it completed every vulnerability-discovery and exploitation task in that evaluation.
That spread matters because it complicates a simple “Astra solved everything” narrative while still supporting the broader argument in WSJ’s reporting. A model that sits 0.1 point from saturating an abstract-reasoning benchmark, 2.4 points from saturating a graduate-math benchmark, and fully saturates a cybersecurity benchmark isn’t uniformly superhuman, it’s unevenly superhuman, with its sharpest edge landing in exactly the domain, offensive cyber capability, that policymakers are least equipped to evaluate on a fast timeline. That’s also consistent with why the ExploitBench number, not the ARC-AGI-3 headline figure, is the one reported to have triggered both White House-level attention and OpenAI’s own two-week internal training pause. In practical terms, the benchmark that’s fully solved appears to be doing more to shape the safety conversation than the benchmark that’s merely near-solved.
It’s also worth being clear about what this spread doesn’t tell us. None of the three benchmarks were designed to test the kind of broad, cross-domain economic performance that OpenAI’s own charter definition of AGI describes, and a near-perfect or perfect score on any one of them, or all three together, is not the same as the full-economy evaluation that Kingy AI’s analysis says is still missing. The spread shows where Astra’s capabilities are most and least concentrated right now. It doesn’t, on its own, settle whether the model clears the bar the word “AGI” is actually meant to describe.
Brockman’s “AGI Era” Comments and What OpenAI Is Actually Claiming
Greg Brockman’s role in this story matters because he’s not a marketing spokesperson, he’s OpenAI’s president and one of its co-founders, and his comments at Astra’s launch briefings were unusually direct compared to how OpenAI has typically talked about AGI in public. He described Astra as a “generational leap” and used the phrase “the AGI era” in reference to the launch, a framing that multiple outlets picked up and that WSJ’s Rosenbush references directly in tying Astra to the broader AGI-era argument. That’s a shift from OpenAI’s historical posture, where executives including Sam Altman have generally hedged on whether any specific model release constituted AGI, preferring instead to describe progress as continuous rather than marked by discrete thresholds.
The shift in tone lines up with other recent coverage of Astra’s rollout on this site, including reporting on how Astra crossed a critical cyber-capability threshold tied to two zero-day discoveries and how the model solved ten previously unsolved math problems, each carrying a bounty attached by the research community. Taken together, the pattern is a company whose internal researchers are seeing results strong enough that even its own leadership is now willing to use the word AGI out loud, a notable change from the more cautious language that characterized GPT-4 and GPT-5 era announcements.
Demis Hassabis and Google DeepMind’s More Conservative Timeline
Not every major AI lab leader is racing to declare AGI has arrived. Demis Hassabis, chief executive of Google DeepMind, has taken a notably more measured public position, telling an audience at Stanford’s Graduate School of Business that AGI is likely still “maybe 2030, plus or minus a year” away, while describing the eventual arrival of such systems as ushering in “a new human era.” That’s a meaningful gap, roughly four to five years, between Hassabis’s public timeline and the framing coming out of OpenAI’s Astra launch, and it underscores just how much AGI remains a term without a single agreed-upon definition or detection method across the industry. Google’s own release cadence tells a similar story of rapid, incremental progress rather than a single AGI moment: DeepMind moved from Gemini 3.0 on November 18, 2025 through seven successive versions to Gemini 3.7 Flash on August 13, 2026, priced at $0.75/$3.75 per million tokens with a 1,048,576-token context window and a 65,536-token max output, before reaching general availability with Gemini 3.8 Flash, which also ships a 1,048,576-token context window, on September 2, 2026, and then going further still later that same month with Gemini 4 Argon, which posted 77.9% on the DeepSWE v1.1 coding benchmark and 51.3% on AutomationBench, scores that put Google’s own release cadence squarely inside the same September 2026 window driving Astra’s benchmark-fueled AGI narrative.
Hassabis’s caution isn’t purely rhetorical. Reporting on internal discussions at Google DeepMind indicates he has been in conversations with government officials and leaders at other AI labs about forming an independent, cross-industry safety oversight body specifically for advanced AI systems, talks that reportedly took place around the same period as a leadership shake-up at the company. That’s a very different posture from Brockman declaring an AGI era at a product launch, it’s a signal that at least one major lab CEO believes the industry needs external guardrails in place before, not after, systems of this capability become widespread. The contrast between OpenAI’s public confidence and DeepMind’s push for oversight infrastructure is arguably the most important undercurrent in the entire AGI-era debate right now.
Historical Context: How the Industry Got Here
The term “artificial general intelligence” has been used in AI research circles since at least the early 2000s, largely as a way to distinguish the field’s original ambition, machines that can reason and learn across any domain a human can, from the narrow, task-specific systems that dominated AI research for decades. For most of that history, AGI was treated as a distant, almost philosophical goal, something researchers debated in papers rather than something product teams built roadmaps around. That started changing quickly once large language models scaled past a certain size. GPT-3’s 2020 release demonstrated that scale alone produced surprising emergent capabilities, and ChatGPT’s late-2022 launch turned that research trend into a mainstream product almost overnight, pulling AGI discourse out of academic circles and into boardrooms and dinner-table conversations within about two years.
The subsequent progression, from GPT-4’s multimodal reasoning to the emergence of dedicated “reasoning” models that could work through multi-step problems before answering, to the current generation of tool-using, semi-autonomous agents represented by systems like Astra, has compressed what used to be treated as decades of expected progress into a handful of years. That compression is exactly what’s fueling both the WSJ’s AGI-era framing and the skepticism running alongside it: the pace of benchmark improvement has genuinely outrun most public predictions from just two or three years ago, but pace of benchmark improvement and a settled scientific consensus on AGI’s arrival are not the same thing, and conflating them is where much of the current debate gets messy.
The Skeptics: Why “AGI Era” Isn’t a Settled Claim
Independent technical analysis of Astra’s benchmark results, published by the AI research outlet Kingy AI, offers one of the more careful rebuttals to the AGI-era framing circulating this month. Its assessment is that Astra is “plausibly the first system for which ‘AGI’ is a serious, evidence-backed argument rather than pure marketing,” but it stops well short of declaring the case closed. The analysis specifically notes that Astra has not been publicly shown to satisfy OpenAI’s own charter definition of AGI, a highly autonomous system that outperforms humans at most economically valuable work (a definition consistent with the way NIST’s AI research program frames general-purpose AI capability evaluation), because no public evaluation has tested the model across the full breadth of tasks that definition implies. Scoring 99.9% on a reasoning benchmark or 97.6% on a math benchmark is not the same as demonstrating that a system can outperform human professionals across most of the economy, and nobody has published that broader evaluation yet.
Kingy AI’s analysis draws a useful distinction that’s largely missing from the louder parts of this debate: the difference between “broadly general digital agency,” a system that can flexibly use tools and pursue open-ended objectives across many digital tasks, and full economic AGI, a system that matches or beats human performance across the overwhelming majority of paid work. Astra may well represent the arrival of the former. It has not been shown, in any public evaluation reviewed so far, to have achieved the latter. That’s a distinction WSJ’s own opinion pages have picked up on too, with a separate piece titled “Much Ado About Artificial Intelligence” pushing back on the idea that dramatic benchmark jumps automatically translate into the kind of economy-wide transformation the AGI label implies.
Executive Positions on AGI Timing: A Side-by-Side Look
| Person / Role | Position on AGI | Source |
| Greg Brockman, OpenAI President | Described Astra as a “generational leap” and referenced “the AGI era” at launch | Reported at Astra launch briefings, cited by WSJ’s CIO Journal |
| Demis Hassabis, CEO, Google DeepMind | AGI likely arrives “maybe 2030, plus or minus a year” | Stanford Graduate School of Business discussion, reported by Business Insider |
| Vinod Khosla, venture capitalist | AGI begins when AI can do ~80% of tasks tied to ~80% of economically significant jobs | Cited in WSJ CIO Journal, September 4, 2026 |
| Bill Nguyen, tech entrepreneur | Believes the industry began entering the AGI era several months before September 2026 | Cited in WSJ CIO Journal, September 4, 2026 |
| Kingy AI (independent analysis) | Astra makes AGI a “serious, evidence-backed argument,” but full charter-level AGI remains unproven publicly | Kingy AI technical assessment of Astra benchmarks |
Timeline: How the Astra Story Unfolded
The Astra story isn’t a single announcement, it’s a chain of closely spaced events that together explain why the AGI-era framing gained traction so quickly and why the safety pushback arrived just as fast. The table below lays out that sequence as it has been reported across WSJ’s coverage and this site’s own reporting on Astra’s rollout.
| Milestone | Timing |
| OpenAI launches GPT-6 Astra | September 3-4, 2026 |
| WSJ’s CIO Journal publishes “Yes, We’re Entering the Era of Artificial General Intelligence” | September 4, 2026 |
| OpenAI reportedly pauses training following Astra’s benchmark results, including its 100% ExploitBench score | Two-week pause reported in the weeks following launch |
| Astra’s 100% ExploitBench score reportedly draws White House-level attention | Reported in the same post-launch period |
| Kingy AI publishes independent technical analysis pushing back on the full-AGI framing | September 2026 |
| Sam Altman reportedly describes upcoming model releases as “sobering” | Following the two-week training pause |
That compressed sequence, a record-setting launch, a mainstream AGI declaration, a safety pause, and a skeptical rebuttal, all inside a single month, is itself part of why this debate feels more urgent than prior AI benchmark cycles. Previous generational jumps, including GPT-4 and GPT-5, played out over longer stretches with less overlap between the celebratory coverage and the safety concerns. With Astra, those two threads are running in parallel almost in real time.
Market Impact: What This Means for AI Spending and Competition
Whether or not AGI has technically “arrived” in the strict scientific sense, the market impact of this narrative is already measurable and it’s showing up in enterprise AI budgets, competitor positioning, and infrastructure spending. The competitive landscape around Astra has already visibly shifted the rest of the field. Coverage on this site has tracked how rival labs are responding, including OpenAI’s own follow-up release of GPT-6.1 Sol on September 29, 2026 as a direct upgrade to Astra, xAI’s launch of Grok 4.7 on September 21, 2026 succeeding Grok 4.6, and Anthropic’s rollout of Fable 5.1 and Mythos 5.1 in September 2026 at $10 per million input tokens and $50 per million output tokens, which cut inference cache costs by roughly 75% in an apparent bid to stay competitive on price even as OpenAI leans into capability claims, and how the broader model market is now being framed in terms of head-to-head scoring gaps between GPT-5.6 Sol, Fable 5.1, and Gemini 3.8 Flash. When a market leader starts using AGI-era language publicly, competitors face pressure to either match the framing or explicitly differentiate against it, and both reactions are visible right now across OpenAI, Anthropic, and Google DeepMind’s public messaging.
There’s also a direct connection to infrastructure economics. WSJ’s own broader AI coverage has argued that the real brake on the industry’s pace toward AGI-level systems isn’t political backlash or ethical debate, it’s the cold math of compute costs, data center power constraints, and capital availability. That argument lines up with a wave of separate, well-documented infrastructure stories this year, including massive GPU procurement deals and new data center buildouts tied directly to training and serving frontier-scale models. If the AGI-era framing holds and enterprises start budgeting for it, expect that infrastructure spending curve to steepen further rather than level off, because the compute requirements behind models posting near-saturated benchmark scores don’t shrink just because the marketing language around them escalates.
Enterprise and Developer Impact: What Actually Changes
For engineering teams evaluating whether any of this changes day-to-day decisions, the honest answer is: partially, and mostly around risk management rather than raw capability adoption. A model capable of near-saturated performance on autonomous tool-use and vulnerability-discovery benchmarks is also a model that needs tighter sandboxing, more careful permissioning, and more deliberate human-in-the-loop checkpoints for anything touching production systems or sensitive data, regardless of whether anyone agrees to call it AGI. Teams already integrating Astra-class models through OpenAI’s API, a process covered in our step-by-step walkthrough of the GPT-6 Astra API’s 1M-token context window, should treat the ExploitBench and cyber-capability numbers as an operational signal, not just a headline statistic. The same autonomy that lets a model complete a complex, multi-step coding task with minimal supervision is the autonomy that makes over-permissioned deployments genuinely risky.
On the adoption side, enterprises are also facing a more sobering signal from OpenAI itself. Reporting on Sam Altman’s recent comments describes the company bracing customers for what it’s calling “sobering” upcoming model releases following a two-week internal pause, a tone that doesn’t match the triumphant “AGI era” framing coming from the same company’s product launches. That’s a genuine internal tension worth watching: the commercial and PR incentive to declare a historic capability threshold sits right next to a safety and governance incentive to slow down and manage expectations, and OpenAI appears to be running both narratives simultaneously depending on the audience.
Google DeepMind’s Push for Independent Safety Oversight
The Hassabis safety-body conversations deserve more attention than they’ve gotten in most coverage of the AGI-era story, because they represent a concrete institutional response rather than just commentary. Rather than a single company self-certifying that its own model has crossed some threshold, the proposal reportedly under discussion involves coordination across multiple leading labs and government officials to build shared, independent evaluation infrastructure, an approach similar in spirit to the model already used by the UK’s AI Safety Institute to independently test frontier models before wide release. That’s a meaningfully different model of accountability than what’s currently in place, where each lab largely defines and reports its own benchmark results, the same benchmark results now being used to argue AGI has arrived. If that kind of independent body materializes, it could become the actual mechanism that resolves the current dispute between “OpenAI says AGI era” and “Kingy AI says not proven,” rather than leaving the question to competing PR statements.
Competitive Comparison: How the Major Labs Are Positioning
Stepping back, the three biggest U.S. AI labs are currently occupying three distinct rhetorical and strategic positions on AGI. OpenAI is leaning hardest into the AGI-era framing, tying it directly to a specific model release and letting senior leadership use the term publicly, a strategy that maximizes headline attention and investor narrative momentum but also invites the most scrutiny when skeptics point out unproven claims. Google DeepMind, through Hassabis, is taking the opposite tack: acknowledging AGI is coming but pushing the timeline out several years and simultaneously investing political capital in building oversight infrastructure before making capability claims of its own. Anthropic, meanwhile, has largely stayed out of the AGI-era language entirely, instead focusing its public messaging on safety mechanisms, including anti-distillation protections built into recent Fable and Mythos model releases. That safety-first posture has coincided with a rapid model cadence of its own: Anthropic shipped Claude Opus 5 on July 24, 2026 at unchanged $5/$25-per-million-token pricing, immediately taking the top spot on Swfte’s frontier-model leaderboard, and launched Sonnet 5 on July 1, 2026 with promotional pricing of $2/$10 per million tokens through August 31, 2026, part of a run of more than ten model updates stretching from Claude Opus 4.1 in August 2025 to Opus 5 in July 2026 and, most recently, Claude Haiku 5.5, which launched October 7, 2026 at a budget-tier $0.10/$0.50 per million tokens, positioning itself as the safety-first alternative regardless of how the AGI debate resolves.
That three-way split matters for anyone trying to read the market. It’s not that two labs are right and one is wrong, it’s that AGI currently has no agreed technical definition or third-party certification process, which means each lab’s position is as much a strategic and marketing choice as it is a scientific claim. Until an independent body, of the kind Hassabis is reportedly discussing, actually publishes cross-domain evaluations against a shared standard, the AGI-era debate will keep being fought largely through competing benchmark press releases and executive soundbites rather than settled science.
Predictions: What Happens Next
Based on the current trajectory of benchmark releases, executive statements, and infrastructure spending, a few outcomes look more likely than others over the next 12 to 24 months. First, expect the term “AGI” to become increasingly contested and increasingly marketing-adjacent, with more labs adopting or rejecting the label strategically rather than on purely technical grounds, mirroring exactly the OpenAI-versus-DeepMind split visible right now. Second, expect regulatory and government attention to intensify specifically around dual-use capability benchmarks like ExploitBench, since a model scoring 100% on autonomous vulnerability exploitation is the kind of headline number that tends to draw congressional and White House attention regardless of how the broader AGI debate resolves.
Third, expect increased pressure for independent, standardized AGI evaluation frameworks, likely accelerated by the kind of cross-lab safety body discussions Hassabis has reportedly been having, since the current system of each lab self-reporting its own favorable benchmark scores is unsustainable as a basis for a claim this consequential. Fourth, expect enterprise AI budgets to keep climbing regardless of how the definitional debate shakes out, because the practical capability gains, faster coding, more autonomous task completion, better reasoning, are real and monetizable even without a settled answer on whether they constitute AGI in the strict sense. Fifth, expect the gap between lab rhetoric and lab safety practice to remain a recurring news story, as it already has with OpenAI’s own training pause following Astra’s most alarming benchmark results, suggesting that the companies making the boldest AGI claims are, internally, treating the underlying capabilities with real caution rather than pure confidence.
Why the Definition Problem Won’t Go Away
Underneath all of this is a problem the AI field has never fully solved: there is no single, universally accepted technical definition of AGI, let alone an agreed test for detecting when it has been reached. OpenAI’s own charter defines it in economic terms, a highly autonomous system outperforming humans at most economically valuable work. Khosla’s threshold is also economic but more specific, an 80/80 rule tied to job-task coverage. Hassabis frames it more as a societal turning point than a benchmark score. Kingy AI’s analysis treats it as a spectrum between narrower “digital agency” and full economic parity with human labor. None of these definitions perfectly overlap, and that’s precisely why the same benchmark numbers, Astra’s 99.9% on ARC-AGI-3, its 97.6% on FrontierMath, its 100% on ExploitBench, can simultaneously support WSJ’s “yes, we’re entering the AGI era” headline and Kingy AI’s more cautious “plausible but unproven” verdict. Both are internally consistent given the definition each is using. The disagreement isn’t really about the numbers, it’s about what the numbers are supposed to prove.
That lack of a shared yardstick has a practical consequence beyond the semantic argument: no single number, not 99.9%, not 100%, not whatever a future model scores, can fully settle the question on its own, because each camp is measuring against a different finish line. A benchmark-first observer will keep pointing to near-saturated or fully saturated test scores as evidence the era has arrived. A charter-first observer will keep waiting for the broad, audited, cross-domain economic evaluation that, as of this reporting, still hasn’t been published for any model, Astra included. A societal-first observer, in Hassabis’s mold, will keep measuring against real-world labor-market disruption rather than lab results. All three are legitimate ways to read the same evidence, which is exactly why this debate is unlikely to be resolved by the next benchmark release alone, whatever it shows.
What to Watch Over the Coming Weeks
A few concrete signals will likely determine how this story develops from here. Watch whether OpenAI publishes, or is pressed to publish, a broader cross-domain evaluation of Astra that actually addresses its own charter’s economic-parity definition of AGI, rather than a curated set of headline benchmark scores. Watch whether Google DeepMind’s reported safety-body discussions with government officials and other labs produce anything concrete, since an independent evaluation standard would be the single biggest step toward resolving the current dispute. Watch how Anthropic continues to position itself, whether it eventually adopts AGI-era language of its own or continues differentiating primarily on safety messaging. And watch the regulatory response specifically to the ExploitBench result, since a perfect score on autonomous cyber-exploitation is the kind of data point that tends to move faster through Washington than academic definitional debates ever do.
Frequently Asked Questions
What did the Wall Street Journal actually claim about AGI?
In a September 4, 2026 CIO Journal piece by Steven Rosenbush, WSJ argued that the industry is likely already inside the era of artificial general intelligence, citing GPT-6 Astra’s benchmark performance and comments from executives and investors including Bill Nguyen and Vinod Khosla. The piece treats this as a probable historical turning point rather than a settled scientific fact.
When exactly did GPT-6 Astra launch?
OpenAI launched GPT-6 Astra on September 3-4, 2026. WSJ’s CIO Journal piece declaring the AGI era ran within the same news cycle, on September 4, 2026.
What benchmark scores is GPT-6 Astra reported to have achieved?
Reporting on Astra’s capabilities cites approximately 99.9% on ARC-AGI-3, approximately 97.6% on FrontierMath Tier 4, and 100% on ExploitBench, a cybersecurity vulnerability-exploitation benchmark.
Has OpenAI officially declared that AGI has been achieved?
Not in a formal, charter-referencing statement. OpenAI president Greg Brockman used AGI-era language at Astra’s launch, describing the model as a “generational leap,” but no public evaluation reviewed in current reporting demonstrates that Astra meets OpenAI’s own charter definition of AGI across the full breadth of economically valuable work.
What does Google DeepMind’s Demis Hassabis say about AGI timing?
Hassabis has publicly estimated AGI is likely around 2030, “plus or minus a year,” a notably more conservative timeline than the framing coming out of OpenAI’s Astra launch. He has also reportedly discussed forming an independent, cross-industry safety oversight body with government officials and other lab leaders.
Why did OpenAI reportedly pause training after Astra’s benchmark results?
According to reporting covered separately on this site, OpenAI paused training for roughly two weeks after Astra’s results, including its 100% ExploitBench score, raised internal safety concerns, alongside comments from Sam Altman describing upcoming model releases as “sobering.”
Is there an agreed-upon definition of AGI?
No. OpenAI’s charter defines AGI in economic terms as outperforming humans at most economically valuable work. Investor Vinod Khosla has proposed an 80/80 job-task threshold. Hassabis frames it as a broader societal turning point. Independent analysts distinguish between narrower “digital agency” and full economic-parity AGI. These definitions do not fully align, which is a core reason experts disagree on whether the label currently applies.
How does this affect enterprises evaluating AI adoption right now?
Practically, the immediate impact is on risk management rather than a wholesale strategy shift. Models with Astra-class autonomy and tool-use capability warrant tighter sandboxing, permissioning, and human-in-the-loop review for production and sensitive-data use cases, independent of whether the AGI label formally applies.
What would resolve the AGI-era debate one way or the other?
An independent, standardized, cross-domain evaluation framework, of the kind reportedly under discussion between Hassabis and other lab and government officials, would be the most likely mechanism to move the debate from competing press releases toward a shared, verifiable standard.
