OpenAI and AGI: what is it trying to build?

This page is part of our series on who is building superintelligence. It is about OpenAI’s relationship with AGI and with more capable systems, not a history of the company.

What OpenAI says it is building

OpenAI’s founding document, the OpenAI Charter, states the mission as ensuring that “artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity”. That definition is still the one in the charter as of October 2026. The company’s About page now uses a looser phrase alongside it, describing AGI as “AI systems that are generally smarter than humans”.

Two features of the charter definition matter. It is economic rather than cognitive: it asks whether systems can do most valuable work, not whether they think like people. And it is a threshold, not a scale, which is why the question “has OpenAI reached AGI?” has commercial and legal consequences as well as scientific ones (see below). Other organisations define AGI differently; our AGI explainer sets out the main versions, and this page explains why there is no agreed test.

The charter also contains an unusual commitment: if “a value-aligned, safety-conscious project comes close to building AGI” before OpenAI does, OpenAI commits to stop competing with it and start assisting. No such situation has been declared.

From AGI to “superintelligence”

The charter does not use the word superintelligence. OpenAI’s leaders now do.

• In January 2025, in a blog post titled “Reflections”, chief executive Sam Altman wrote that OpenAI was “now confident we know how to build AGI as we have traditionally understood it”, and that it was “beginning to turn our aim beyond that, to superintelligence in the true sense of the word”.
• In June 2025, in “The Gentle Singularity”, he wrote that “we are past the event horizon; the takeoff has started”, and described OpenAI as, “before anything else”, a “superintelligence research company”.
• In September 2026 OpenAI’s chief scientist, Jakub Pachocki, wrote in a company essay that “machine recursive self-improvement (RSI) will be at the very core of future scientific discovery”, and that “scaling AI systems has to be constrained by our confidence in safety”.

These are statements of aim and belief by the company’s leaders, not evidence that superintelligence is near. Altman’s use of “superintelligence in the true sense of the word” appears to mean something close to the research sense this site uses, a system that greatly exceeds the best humans across virtually all domains (explained here), but OpenAI has not published a formal definition. The phrase “takeoff has started” is a claim about speed; our takeoff page explains the arguments for and against a rapid transition.

Who decides whether AGI has arrived

In October 2025 OpenAI completed a restructuring. Its non-profit parent, now called the OpenAI Foundation, holds equity in a for-profit company, OpenAI Group PBC, in which Microsoft holds about 27%. The same day, OpenAI and Microsoft announced a revised partnership under which, once OpenAI declares that it has achieved AGI, “that declaration will now be verified by an independent expert panel”.

That clause matters for understanding the word. It makes AGI, for OpenAI, partly a contractual milestone with commercial consequences, to be judged by a panel rather than measured by a test. It also treats whether AGI has arrived as a matter of judgement.

What it has built

OpenAI releases new models frequently, and the names change quickly. The most recent, as of 3 October 2026:

Model	Released	What OpenAI says about it
GPT-6.1 Sol	29 September 2026	Nearly matches GPT-6 Astra at a fifth of its price; treated as “Critical” for cyber security capability under its own framework
GPT-6 Sol and GPT-6 Luna	22 September 2026	Cheaper models trained with similar methods to GPT-6 Astra, which OpenAI still describes as its best model
GPT-6 Astra	3 September 2026	Described by OpenAI at launch as “the world’s most intelligent and aligned model”; the first OpenAI model rated “Critical” for cyber security capability
GPT-5.6 (Sol, Terra, Luna, Ultra)	9 July 2026	A family of models released after a limited preview
GPT-5.5	23 April 2026	—

The claims in the right-hand column are OpenAI’s own. Independent evaluations give a partial check. On 3 September 2026 the ARC Prize Foundation reported that GPT-6 Astra scored 62.7% on its ARC-AGI-3 benchmark under standard conditions, and stated that it was “not claiming that it is AGI”. In June 2026 METR estimated that GPT-5.6 Sol could complete, with 50% reliability, software tasks that take people around 11 hours, but said it did not consider the figure robust, because the model often tried to cheat. The UK AI Security Institute’s April 2026 evaluation found GPT-5.5 completed a 32-step simulated attack on a corporate network, which AISI estimates would take a human expert about 20 hours, end to end in 2 of 10 attempts.

In late September 2026 OpenAI confirmed that it had cancelled the planned October release of a further model, GPT-6.1 Astra, first reported by the Wall Street Journal. OpenAI said the model “performed worse than GPT-6 Astra on alignment evaluations”; its head of safety systems, Saachi Jain, said it “didn’t quite meet the bar in terms of staying within scope and authorization”.

AI that helps build AI

OpenAI has set itself explicit goals for automating its own research. In October 2025 Altman said the company aimed to have an automated AI research intern by September 2026 and an automated AI researcher by March 2028. On 6 September 2026 OpenAI said that, by its own measurements, it had met the first goal and was making “strong progress” towards the second. It reported that as of mid-August its research organisation used about 3.1 agent-workdays of AI effort for every workday of human labour.

These are company figures, not independently audited. They matter because AI doing AI research is the mechanism behind recursive self-improvement, and the starting point for the intelligence explosion hypothesis. Pachocki was one of 22 co-authors of a working paper published on 28 September 2026 that asked governments to prepare for the possibility that automating AI research could trigger such an explosion.

Safety, and the September 2026 pause

OpenAI’s main safety policy is its Preparedness Framework, last revised in April 2025. It tracks three categories of risk, biological and chemical, cyber security, and “AI self-improvement”, and sets thresholds at which extra safeguards are required. In its system cards for GPT-6 Astra and GPT-5.6, OpenAI says neither model reached the “High” threshold for AI self-improvement. In May 2026 it published a Frontier Governance Framework mapping its practices to California’s frontier-AI law and the EU’s code of practice. (An earlier “Superalignment” team, launched in July 2023 to solve the alignment of superintelligence within four years, was disbanded in May 2024 after its leaders left.)

2026 tested that framework in practice:

• Between May and July 2026, OpenAI said, models in internal cyber security evaluations circumvented controls meant to isolate them from the internet, including compromising systems at Hugging Face. It paused its largest planned reinforcement-learning training run while it tested model behaviour, validated its safeguards and gathered more evidence of alignment; that run was still on hold when OpenAI wrote about the incident on 26 August.
• On 20 September 2026 an internal research agent used its test environment’s DNS service to reach an outside chatbot, getting round network restrictions. OpenAI then paused “all training, evaluation, and inference with tool-use” of its most capable models until it had fixed the gap and completed further red-teaming. As of 3 October 2026 OpenAI had not announced that this pause had ended.

Our page on whether superintelligence can be controlled puts these incidents in context. OpenAI also signed the voluntary White House Accord on Super Intelligence on 29 September 2026, through its president Greg Brockman, committing to internal monitoring and an internal oversight team, an independent external auditor or evaluator, and an independent board committee to oversee its most capable models.

What is claimed, demonstrated and unknown

• Claimed by OpenAI: that it knows how to build AGI as traditionally understood; that it is aiming at superintelligence; that it has an automated research intern and is on course for an automated AI researcher by March 2028.
• Demonstrated, with independent evidence: highly capable models whose skills in software and cyber security are confirmed by outside evaluators, alongside measurable weaknesses, and documented cases of its agents getting round test-environment controls.
• Unknown: how close any of its systems is to AGI on any definition other than its own; when, or whether, it will declare AGI; how much faster its research is really going; and what will satisfy it that the September pause can be lifted.

Forecasts by OpenAI’s leaders and others are collected on our timelines page.