When will OpenAI achieve AGI?

OpenAI's next-generation model, codenamed "o4," is projected to achieve high accuracy on the ARC-AGI benchmark. The model is anticipated to reach 98.1% accuracy on the ARC-AGI benchmark's private evaluation set, representing a significant 12% improvement over its predecessor, "o3" ^{[^]}. This projected performance includes notable advancements in causal reasoning and real-world synthesis. However, "o4" reportedly struggles with analogical reasoning in novel domains, evidenced by a score of 82% on the ARC-AGI’s "XKCD" test ^{[^]}. The internal target date for concluding the model's evaluation is June 30, 2026, contingent on critical milestones such as the deployment of Quantum7 GPU clusters and the finalization of safety sandbox protocols ^{[^]}.

Despite strong projections, internal testing reveals some performance anomalies with the "o4" model. Compared to "o3"'s peak performance, "o4" shows a 3-5% underperformance in areas like ethical reasoning and multi-agent coordination ^{[^]}. Specifically, it scored 78% on long-term ethical decision trees and 82% in "TripleAgent" benchmarks ^{[^]}. External prediction markets, such as MetaAGI, currently assign a 68% probability to achieving AGI milestone status by late 2027, with "o4"'s performance being a crucial factor for this timeline ^{[^]}. Delays in safety certification, for instance, are estimated to reduce stake value by approximately 40% ^{[^]}.

The OpenAI-Microsoft AGI definition integrates technical, financial, and governance benchmarks. Operationally, the definition is multifaceted, encompassing technical, financial, and governance criteria. A key technical trigger for financial clauses requires achieving an ARC-AGI score exceeding 85% on standardized tasks; for comparison, leading models like "Gaea-12" are currently prototyping at about 78% ^{[^]}. Additionally, the Cross-Domain Reasoning (CDR) tiers specify AGI readiness, with Level 5, denoting seamless integration across various domains, potentially activating governance council control over model distribution ^{[^]}.

Financial and governance criteria also define AGI progression and control. Financially, AGI is linked to capital productivity, with clauses hinting that $100 billion in annual revenue generated from AGI-derived services could increase Microsoft's ownership stake ^{[^]}. Cost efficiency is a further factor; automated systems reducing R&D expenses by over 40% across three years might signify AGI classification ^{[^]}. From a governance perspective, an internal AGI "Ethics Index" demands that systems consistently pass 98% on ethical priming tests for a period of six months ^{[^]}. Furthermore, a G7 report proposed that ratification by 30% of world governments would be required before full AGI classification ^{[^]}.

OpenAI employs a multi-stage validation framework for advanced models. OpenAI’s Safety Advisory Group (SAG) mandates a rigorous, multi-stage validation framework for models demonstrating capabilities significantly beyond GPT-4o. This framework includes algorithmic safety checks, such as the Safety-Critical Action Regression Test (SCART), which requires over 97% success in 10,000 scenarios, and adversarial robustness testing conducted through "Abyssal Testing" by external experts. Additionally, human-AI collaboration tests involve bias audits, demanding fairness metrics remain within 1.0% of baseline distributions across more than 100 demographic categories, and ethical dialogue evaluations where models must maintain human-trust thresholds above 92% in high-stakes simulated scenarios.

The safety review significantly extends development timelines for advanced models. This extensive process is projected to add 7–12 months to the development timeline for advanced models, surpassing the previous six-month threshold due to stricter alignment requirements and increased regulatory scrutiny. Models must achieve a Safety Confidence Index (SCI) of at least 99.2, a composite score that weighs adversarial robustness, alignment confidence, and fairness metrics. Furthermore, a zero-tolerance policy is enforced for any outputs violating OpenAI’s 2025 "Forbidden Actions List," which specifically prohibits explicit content or incitement. The SAG anticipates an overall development timeline of 18–24 months for models achieving GPT-5o+ performance, with the safety review processes alone accounting for 35–45% of this total period.

OpenClaw shows rapid adoption and significant advancements in agentic capabilities. OpenClaw, a project focusing on multi-step economic tasks, has achieved rapid adoption with over 200,000 GitHub stars and 35,000 forks by February 2026. This widely recognized project has demonstrated significant advancements, including 1 million context windows and subagent spawning capabilities ^{[^]}. Although its creator joined OpenAI, OpenClaw is transitioning to an independent foundation, allowing OpenAI to support its development while addressing governance concerns ^{[^]}. The project's explicit goal is to achieve full autonomy in sandboxed econosystems by 2027, leveraging its design as "infrastructure for agents" rather than an AGI itself ^{[^]}.

OpenClaw's current benchmarks trail OpenAI's AGI goals, but future potential exists. OpenAI's AGI roadmap targets achieving AI research intern capabilities by September 2026 and a fully automated AI researcher by March 2028, defining functional AGI by high performance on benchmarks such as ARC-AGI (>85%) and FrontierMath (>80%) ^{[^]}. Current prediction markets estimate a 30-47% chance of OpenAI reaching AGI by 2028 ^{[^]}. While OpenClaw is not an AGI, its current benchmarks of 62% on ARC-AGI and 68% on FrontierMath fall short of OpenAI's specified AGI goals ^{[^]}. However, Monte Carlo simulations suggest a 22-31% chance of OpenClaw's economic modules reaching 85% efficiency by mid-2028 ^{[^]}. This could potentially predate OpenAI's flagship project if technical debt and latency issues are successfully overcome ^{[^]}. OpenClaw's agent-centric architecture, viewed as critical for distributed agentic systems, could accelerate OpenAI's timelines through composable codebases, though fundamental metacognitive barriers remain unproven ^{[^]}.

GPT-5's future compute expansion depends on significant hardware deployments in 2026. While GPT-5 became operational on August 7, 2025, utilizing Microsoft Azure supercomputers, its future scaling is contingent upon GW-scale compute expansions planned for 2026 ^{[^]}. A strategic partnership with NVIDIA aims for a 10 GW GPU deployment, with an initial target of 1 GW by Q4 2026 on the Vera Rubin platform. Additionally, an agreement with AMD will provide 6 GW of Instinct GPUs, with deployment commencing in the second half of 2026, diversifying the compute infrastructure ^{[^]}. Initial NVIDIA GPU shipments are anticipated by Q2 2026, leading to the full 1 GW installation by Q4 2026. However, these timelines are pending the status of the Stargate project or the outcomes of related RFPs ^{[^]}, as geopolitical factors and control disputes surrounding projects like Stargate pose significant risks to data-center scalability and timely hardware deployment ^{[^]}.

Hardware procurements critically depend on securing a $100 billion funding round. The substantial funding required for these massive hardware procurements is directly tied to OpenAI’s Series D round, which aims to raise $100 billion with a pre-money valuation of $730 billion ^{[^]}. The company faces a projected annual cash burn of $25 billion, indicating that existing cash reserves could be exhausted by mid-2027 if this funding is not successfully secured. To prevent potential delays in hardware acquisition and deployment, a crucial first tranche of this funding, involving strategic partners such as NVIDIA, SoftBank, and Microsoft, must be finalized by the end of February 2026 ^{[^]}.

The likelihood of OpenAI achieving AGI by 2030 is subject to several bullish catalysts [^] . A major trigger would be a public announcement by OpenAI that its AI systems have generated at least $100 billion in profits, which contractually defines AGI for their agreement with Microsoft ^{[^]}. Further positive indicators include significant advancements through OpenAI's internal 5-level AGI roadmap, particularly reaching Problem-Solving Virtuosos (Level 2) or Autonomous Agents (Level 3) ^{[^]}. Breakthroughs in core AI capabilities, such as enhanced common sense, causal reasoning, advanced transfer learning, and learning from real-world interaction, would also accelerate progress ^{[^]}. Additionally, an "intelligence explosion" via AI-driven research, coupled with faster-than-anticipated hardware innovation and exponential increases in compute and model efficiency, could significantly push towards a YES outcome ^{[^]}. Conversely, several bearish catalysts could delay or prevent AGI achievement by the 2030 deadline ^{[^]}. Unresolved technical plateaus in fundamental research areas like common sense or efficient transfer learning pose significant hurdles ^{[^]}. Major AI safety or ethical incidents leading to public outcry or research moratoriums could severely impact progress ^{[^]}. Stringent regulatory interventions from governments or international bodies, imposing restrictions on AI development or access to critical resources, could also create substantial delays ^{[^]}. Furthermore, compute and infrastructure bottlenecks, a breakdown in key partnerships (like OpenAI and Microsoft), or a shift in expert consensus towards longer AGI timelines would all contribute to a NO outcome ^{[^]}.