OpenAI enters "Phase 3": Personal AGI for every human, automated AI researcher, and global governance
🔎 The game-changing manifesto
On June 8, 2026, Sam Altman and Jakub Pachocki publish "Built to Benefit Everyone", a text that resembles a strategic manifesto more than a blog post. OpenAI officially declares its entry into "Phase 3" — a shift from tool to autonomous agent, from model to system.
Why now? The timing is not coincidental. This announcement comes as OpenAI has filed a confidential IPO filing, as the race to AGI accelerates with models like GPT-5.5 (98.2 in agentic) and as Google DeepMind is openly recruiting "post-AGI research scientists" according to 404 Media. Phase 3 is OpenAI moving from promise to execution — or at least, selling it as such.
It remains to be dissected what these three objectives are actually hiding: an operational roadmap or a pre-IPO narrative exercise.
The key points
- OpenAI declares it is entering "Phase 3" with three objectives: an automated AI researcher, the acceleration of the global economy, and a personal AGI for every human.
- The manifesto coincides with a confidential IPO filing and the rise of agentic models (GPT-5.5 reaches 98.2 on agentic benchmarks).
- A funding program for independent research on the economic impact of AI is announced — a first for OpenAI.
- The concept of AGI used by OpenAI relies on the levels framework published in Levels of AGI for Operationalizing Progress on the Path to AGI, a paper that defines measurable tiers.
- Google DeepMind is recruiting "post-AGI" researchers, a sign that the competition to define the post-AGI era is already underway.
Recommended tools
| Hostinger | Web hosting to deploy AI projects | Starting at 2.99 € (June 2026, check on hostinger.com) | AI developers and startups |
|---|---|---|---|
The three phases of OpenAI: a historical recap
To understand Phase 3, we must reject the collective amnesia affecting the tech industry and trace the two previous phases.
Phase 1: proof of concept (2015-2022)
Everything begins with the founding of OpenAI in December 2015, a nonprofit structure funded by Musk, Altman, and others. The stated goal: build safe and beneficial AGI. The operational reality: publishing research papers, exploring reinforcement learning, and producing models (GPT-2, GPT-3) that demonstrate that scaling works. This is a laboratory phase, without direct commercial pressure.
Phase 2: go-to-market (2022-2026)
The launch of ChatGPT in November 2022 marks the shift. OpenAI goes from a research lab to a consumer product. The transition to a "capped-profit" structure, the partnership with Microsoft, monetization via ChatGPT Plus, and then the successive models (GPT-4, GPT-5, up to GPT-5.5 in 2026). This phase is about validation: proving that generative AI has a market, a business model, and massive demand.
Phase 3: AGI as a system
Phase 3, as described in the manifesto, is no longer about models but about systems. The automated AI researcher is not an LLM that answers questions well — it is an agent capable of carrying out a complete research cycle autonomously. Personal AGI is not a chatbot — it is a system integrated into daily life. The leap is qualitative, not just quantitative.
This three-phase framing is strategically clean. It gives the impression of a plan mastered from the very beginning. The reality is more chaotic, but the narrative works, especially in the pre-IPO period.
The automated AI researcher: the real technical pivot
What OpenAI promises exactly
The primary goal of Phase 3 is the construction of an "automated AI researcher" — a system capable of formulating hypotheses, designing experiments, executing them (at least in simulation), analyzing results, and iterating. This is the core of the manifesto, and it is where the connection to current models becomes concrete.
GPT-5.5 dominates agentic benchmarks with a score of 98.2, followed by Gemini 3 Pro Deep Think (95.4) and Claude Opus 4.7 Adaptive (94.3). These figures are not decorative: they precisely measure a model's ability to plan, use tools, and carry out multi-step tasks autonomously. The automated researcher is the logical extension of these capabilities.
Why it is credible — and why it is premature
Credibility comes from measurable progress. the paper OpenAI o1 System Card already documented o1's chain-of-thought reasoning capabilities, which reached 90.2 in agentic. In one year, OpenAI went from o1-preview to GPT-5.5, an 8-point leap on these same benchmarks. The trajectory is real.
But the automated researcher requires more than just a good LLM. It requires integration with simulation environments, computational APIs, scientific databases, and verification mechanisms. The dexterous manipulation documented in Learning Dexterous In-Hand Manipulation shows the complexity of physical interaction — the automated researcher facing a purely digital world remains simpler, but requires considerable infrastructure.
The connection with the robotics division
This pivot toward the autonomous agent fits into OpenAI's broader logic, which recently launched its robotics division. The automated AI researcher is the embryo of an intelligence that does not remain confined to a screen. From software AGI to embodied intelligence: Phase 3 is the bridge between the two.
Personal AGI for every human: the social promise
What exactly are we talking about?
The second goal is the most publicized: to offer a personal AGI to every human on Earth. The wording is deliberately universalist, almost utopian. But what does a "personal AGI" actually mean?
Drawing on the Levels of AGI framework, this promise can be interpreted as providing a system at the "Competent" to "Expert" level on most cognitive tasks, one that is personalized and accessible. Not necessarily a superintelligent system, but an agent capable enough to be transformative in everyday life.
AI avatars as an entry point
The most visible realization of this promise comes through AI avatars for customer service and, more broadly, through AI avatars paired with personal assistants. A personal AGI without a natural interface remains a tool for tech enthusiasts. The avatar is the vector for mass adoption.
The gap between the promise and reality
"Every human on Earth" implies distribution to 8 billion people. Today, even GPT-5.5 requires a reliable internet connection, a modern device, and the ability to pay for a subscription. The manifesto does not detail any distribution mechanism for developing countries. The promise is social, but the plan is silent on access inequalities.
The paper AGI: Artificial General Intelligence for Education had already explored the possibilities of AGI in education, particularly for large-scale personalization. But the authors also noted the infrastructural and cultural barriers. OpenAI glosses over them in its manifesto.
Accelerating the global economy: the funding program
An unprecedented move by OpenAI
The third objective is the most surprising: a funding program for independent research on the economic impact of AI. This is a first for OpenAI, which until now had only funded technical research.
This move has a dual interpretation. The first is generous: OpenAI recognizes that the economic impact of AGI cannot be studied exclusively in-house, and that independent research is necessary to inform public policy. The second is strategic: by funding research, OpenAI influences the agenda, the questions asked, and indirectly the results.
What this means for governance
The manifesto speaks of "global governance" without ever proposing a concrete mechanism. The funding program is the only tangible element in this area. It allows OpenAI to position itself as a governance player without ceding real control.
The paper Quantum AGI: Ontological Foundations, although theoretical, raises fundamental questions about the epistemological frameworks necessary to think about the governance of an AGI. OpenAI does not dwell on this. Its governance remains pragmatic and self-centered.
The race with regulators
This funding program comes at just the right time as the European Union implements the AI Act, China strengthens its regulatory framework, and the United States hesitates between federal legislation and a state-by-state approach. By funding economic research, OpenAI produces the data that will feed the legislative debate — and potentially shape it.
The race to AGI: where do we really stand?
Benchmarks tell a story
The generalist LLM landscape in June 2026 is dominated by Gemini 3.1 Pro (92), GPT-5.5 (91) and GPT-5.4 Pro (91). Claude Opus 4.7 Adaptive and Gemini 3 Pro Deep Think follow at 90. But it's in agentic that the difference is made: GPT-5.5 pulls ahead with 98.2, while the runner-up (Gemini 3 Pro Deep Think) caps at 95.4.
This agentic domination is not insignificant in the context of Phase 3. An automated AI researcher, a personal AGI — these are agentic systems by nature. GPT-5.5's score is not just a number — it's the technical argument that makes Phase 3 plausible.
Google DeepMind is not standing still
The clearest confirmation that OpenAI's Phase 3 is perceived as a serious shift in direction comes from Google. 404 Media reports that DeepMind is explicitly recruiting "post-AGI research scientists". This job title would not exist if Google's teams did not believe that AGI is imminent — or at least, that they must prepare for it institutionally.
The Chinese challengers
DeepSeek V4 Pro (Max) reaches 88 in general and 84 in agentic. Moonshot AI's Kimi K2.6 climbs to 84 in general and 88.1 in agentic (self-hosted). Z.AI's GLM-5.1 reaches 83 in general. These figures show that the race is not bipolar. China remains in the race, particularly on the agentic front where Kimi K2.6 surpasses Claude Sonnet 4.6 (81.4) and Grok 4.1 (79).
Phase 3 and IPO: the financial calculation
The confidential filing
According to BitsMinds, the manifesto coincides with a confidentially filed IPO filing. The connection is obvious: a manifesto that promises an automated researcher, a universal personal AGI, and the acceleration of the global economy is also a presentation document for institutional investors.
What investors want to hear
A classic S-1 file talks about market, revenues, margins. The OpenAI manifesto talks about transforming the global economy. It is more ambitious, but it is also riskier from a regulatory point of view (the SEC does not like unquantifiable promises in pre-IPO documents).
Phase 3 solves this problem by offering a narrative: OpenAI is no longer selling a product, it is selling a stage of human development. The automated researcher justifies the R&D. The personal AGI justifies the TAM (Total Addressable Market). The governance program justifies the "responsible" approach that ESG investors demand.
Comparison with the competition
Context is important. If Anthropic has also filed a confidential IPO filing at $965 billion, OpenAI's valuation could exceed this figure thanks to the dominance of GPT-5.5 in agentic. Phase 3 is a differentiation argument: OpenAI is not presenting itself as a model provider, but as the architect of the post-AGI era.
AGI Levels: what the manifesto doesn't say
The framework
The OpenAI manifesto implicitly uses the framework published in Levels of AGI, a paper that defines six levels ranging from "Emerging" (superior to a basic chatbot) to "Superhuman" (superior to 99% of humans on all tasks). This framework has become the de facto reference in the industry.
Where does GPT-5.5 stand?
Based on available benchmarks, GPT-5.5 likely sits between the "Competent" and "Expert" levels for general cognitive tasks, and is approaching "Expert" in agentic. But the paper on AGI levels emphasizes a crucial point: the level must be measured across a broad range of tasks, not just on synthetic benchmarks.
The leap to "Superhuman"
Phase 3 implies that the automated AI researcher could reach the "Superhuman" level in specific domains (research in physics, mathematics, computational biology). This is more plausible than a general superhuman, and is probably what OpenAI is aiming for first. But the manifesto remains deliberately vague on this point.
Geopolitical implications: governance in question
The illusion of self-regulation
The manifesto proposes "global governance" without specifying who governs, according to what rules, and with what sanctioning power. This is the text's weak point. OpenAI positions itself as the architect of governance while remaining a private entity whose goal is to maximize value for its shareholders (future, post-IPO).
The role of States
No State will entrust the governance of AGI to a private company, regardless of its size. The European AI Act, Chinese regulations, and debates in the US Congress show that States are taking back control. OpenAI's funding program is an attempt to influence this process from the inside, not to replace it.
The ontological dimension
The paper Quantum AGI: Ontological Foundations reminds us that the fundamental questions about the nature of AGI — is it conscious? does it have rights? — remain open. Technical governance (safety, alignment) is not enough. Ontological governance is required, and this cannot come from a single company.
❌ Common mistakes
Mistake 1: Confusing Phase 3 with the arrival of AGI
The most widespread mistake is reading "Phase 3" as "AGI achieved". That is not what the manifesto says. Phase 3 is the stage where OpenAI builds the systems that could lead to AGI, not where AGI is operational. The nuance is major, especially for investors.
Mistake 2: Taking the funding program for philanthropy
The independent research program funded by OpenAI is a strategic tool, not a selfless act. Any research funded by a player with a direct interest in the outcome must be read through a critical lens. This is true for pharmaceutical labs, and it is true for OpenAI.
Mistake 3: Ignoring the competitive context
Reading OpenAI's manifesto without looking at what Google DeepMind (post-AGI recruitment), Anthropic ($965 billion IPO), DeepSeek, and Moonshot AI are doing is like reading a single page of a book that has hundreds of them. Phase 3 does not exist in a vacuum.
❓ Frequently Asked Questions
What exactly is OpenAI's "Phase 3"?
It is the stage where OpenAI transitions from building LLM models (Phase 2) to building autonomous systems: an automated AI researcher, a universal personal AGI, and an economic acceleration program. The manifesto describes it as the transition from tool to agent.
Does the automated AI researcher already exist?
Not in the full sense described by the manifesto. GPT-5.5 (98.2 in agentic) possesses the necessary reasoning and planning capabilities, but integration with research environments, databases, and verification systems remains a major work in progress.
What is the link between Phase 3 and OpenAI's IPO?
The manifesto serves as an investment narrative. Promising an automated researcher and a universal personal AGI gives the company a transformative dimension that justifies a potentially higher valuation than that of Anthropic ($965 billion).
Is Google DeepMind ahead of OpenAI?
Not in terms of agentic benchmarks (GPT-5.5 at 98.2 vs Gemini 3 Pro Deep Think at 95.4), but DeepMind signals its confidence by hiring "post-AGI" researchers, suggesting institutional preparation for the post-AGI era that is potentially more advanced.
Is personal AGI realistic for "every human"?
The promise is universal, but the distribution means are not detailed. Access requires network infrastructure, a device, and the ability to pay. The manifesto deliberately evades the issue of access inequalities, highlighted in the paper AGI for Education.
✅ Conclusion
OpenAI's Phase 3 is a clever manifesto that fuses a credible technical roadmap (the automated researcher builds on GPT-5.5's agentic dominance), a compelling social promise (personal AGI for everyone), and a transparent financial calculation (paving the way for a record IPO). It is neither pure vision nor simple PR — it is both at the same time. The rest will depend on execution, not narrative.