{"id":77966,"date":"2026-08-26T16:54:44","date_gmt":"2026-08-26T20:54:44","guid":{"rendered":"https:\/\/overcentral.com\/en\/?p=77966"},"modified":"2026-08-26T16:54:44","modified_gmt":"2026-08-26T20:54:44","slug":"openai-agi-2026-altman-77966","status":"publish","type":"post","link":"https:\/\/overcentral.com\/en\/openai-agi-2026-altman-77966\/","title":{"rendered":"OpenAI Confirms AGI by End of 2026 Under Altman’s Definition"},"content":{"rendered":"

OpenAI<\/a> has confirmed it expects to reach artificial general intelligence (AGI) by the end of 2026<\/a> under Sam Altman’s definition. The company’s leadership is more publicly confident than ever: Altman says OpenAI is “not quite yet” there, Chief Research Officer Mark Chen puts the distance at “80 percent of the way,” and co-founder Greg Brockman believes that when the history of AI is written, this period will be remembered as the moment AGI emerged. The assessment is based on an investigation that involved more than two weeks of interviews with more than 20 executives, employees, investors, and rivals. But the claim already faces sharp skepticism, because the system at the center of the confidence, the Astra model family, is being asked to do something no AI has reliably done: invent new things that matter.<\/p>\n

OpenAI confirms AGI by end of 2026 under Altman’s definition \u2014 but not everyone shares the confidence<\/h2>\n

The confidence is rooted in a deliberately practical definition. OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” That is an economics-driven benchmark, not a philosophical one. It means an AI system does not need to replicate human consciousness, spark a scientific revolution, or understand the world the way people do. It needs to do economically valuable work better than humans across most domains. By that measure, Altman believes the company’s next major model, Astra, is close enough that the target will fall by the end of 2026.<\/p>\n

What is OpenAI’s definition of AGI?<\/h3>\n

OpenAI defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” In plain terms, the company’s target is a capability threshold based on economic output, not a test of consciousness or general reasoning. This definition makes AGI a practical and measurable goal: if a system can reliably outperform a human worker across most economically meaningful tasks, OpenAI considers it AGI.<\/p>\n

Altman’s confidence is not a soft public relations line. It is tied to concrete internal benchmarks. Astra can already perform tasks that OpenAI treats as evidence of AGI-like behavior. The question of whether that is true AGI turns on which definition you accept. Under Altman’s definition, the company is close enough to publicly name a deadline. Under a more demanding definition, OpenAI is still working through limitations that are not solved by scale alone.<\/p>\n

Altman’s own comments acknowledge that AGI is not yet a finished achievement. Mark Chen’s 80-percent estimate suggests meaningful headroom. And Greg Brockman’s framing that this period will be viewed as the emergence moment indicates that OpenAI sees the transition as incremental rather than a single breakthrough event. That combination of confidence and caution is exactly why the “end of 2026” claim feels simultaneously bold and slippery.<\/p>\n

Astra: The automated research intern that OpenAI says is already working<\/h2>\n

The heart of OpenAI’s AGI argument is Astra, the model family that OpenAI has positioned as an automated AI research<\/a> intern. Chief Scientist Jakub Pachocki says Astra already hits that internal benchmark. Hand it an experimental idea, and it can write the code needed to test the idea inside OpenAI’s own codebase, run the experiment, and report back with results. Give it a research paper, and it can handle follow-up work that would occupy a human researcher for about a week.<\/p>\n

Astra also serves as the engine for “persistent agents” \u2014 virtual coworkers that can be pointed at a task and left to work on it for longer periods. Those agents change the economics of AI as much as the underlying model quality. A system that can plan, execute, and verify a meaningful slice of research work opens the door to a compounding effect: AI that helps researchers build better AI.<\/p>\n

Altman described the expected leap in direct terms during a customer preview. “I expect this will be the first model where the model actually invents new things in a way that matters,” he said. “That’s a very AGI-like thing.” That sentence captures the specific bet: Astra’s value will not be measured by how many lines of code it writes, but by whether it produces a novel molecule, a new algorithm, or an experimental design that a human would not have found on their own.<\/p>\n

The underlying wager is that this kind of capability would enable recursive self-improvement. Once an AI can do research-level work, it can help build the next generation of AI, and that next generation can help build the one after that. The lingering skepticism is substantial. Some researchers believe fully autonomous AI research is still far off, pointing to the gap between running experiments and choosing which experiments are worth running. Others see early evidence that it is already happening, particularly in code generation and scientific hypothesis testing.<\/p>\n

Can language models truly invent new things? The unresolved question at the core of AGI<\/h2>\n

The phrase “invent new things” is a much harder bar than “do economically valuable work.” It is where OpenAI’s definition and the broader research understanding of AGI begin to split. Many researchers agree that language models are excellent at recombining existing knowledge, but they disagree sharply on whether such systems can make true discoveries.<\/p>\n

The case for optimism rests on emergence. Models like Astra can already design and execute experiments within constrained environments, which is a form of scientific behavior that did not exist in earlier generations of AI. The case for skepticism is that a language model has no causal model of the world. It predicts text, which may be enough to simulate research but not enough to verify that a result is genuinely new, broadly true, or worth pursuing.<\/p>\n

Several researchers argue that AGI will require a robust understanding of the world and that language models are just one component of such a system. They point to the difference between statistical pattern matching and genuine discovery. A system can assemble a plausible research report by combining fragments of papers it has seen before, but that is not the same as identifying a hidden structure in nature or proposing an experiment that invalidates a widely held theory.<\/p>\n

Because AGI has no universally accepted definition, part of the dispute is semantic. OpenAI has selected a definition that is internally consistent and economically testable. Other definitions, such as those that require autonomous discovery, general reasoning, and continuous learning, produce different timelines. The end-of-2026 claim is therefore not a prediction that all researchers would agree with; it is a claim that OpenAI will meet its own threshold. That distinction matters for anyone trying to assess the announcement.<\/p>\n

OpenAI’s hardware roadmap: A puck-shaped device and humanoid robots are on the way<\/h2>\n

AGI is not a software-only story for OpenAI. Altman has outlined a “small handful” of devices the company is building: something that belongs on a table, something for your pocket, and something you wear. The first product in that family will be a small puck-shaped device that senses its surroundings and talks to its owner via voice mode. It is expected to arrive early next year, placing it in 2027 alongside the AGI timeline.<\/p>\n

The device push is significant because it gives agentic models a persistent, physical presence. A puck that can hear, see, and talk is a natural interface for “always-on execution,” the kind of continuous work OpenAI is trying to sell through products like ChatGPT Work. It also turns AGI from a research claim into a consumer product that people can place on a desk and interact with daily.<\/p>\n

Altman also made clear that OpenAI will “definitely” build humanoid robots as well. That extends the company’s ambitions into physical automation, where the same models that perform digital research work could eventually navigate office spaces, manipulate tools, and carry out tasks in the real world. The combination of a research model, persistent agents, a dedicated hardware interface, and humanoid robotics gives OpenAI a broader path to “highly autonomous systems” than a chatbot alone would provide.<\/p>\n

The business layer: The Merge, ChatGPT Work, and advertising<\/h2>\n

The harder problem may be financial. OpenAI has to pay for the enormous computing infrastructure that Astra and similar models require. The company’s answer is a tighter product focus. The most important internal move is a project called “The Merge,” which combines Codex, OpenAI’s coding agent, with ChatGPT. The first public result is ChatGPT Work, a product designed to put agentic AI<\/a> in front of a mass audience. Instead of just answering questions, it handles tasks.<\/p>\n

Thibault Sottiaux, the product lead running the project, says OpenAI is “close” to shipping something with “persistence and always-on execution.” “It’s definitely going to feel like a new thing to people,” he said. That persistent execution layer is essential if OpenAI wants to turn AGI into revenue. A model that can only answer questions is comparatively easy to sell as a subscription; a model that can manage a meaningful slice of someone’s job is a replacement for labor, and therefore a much larger economic opportunity.<\/p>\n

OpenAI is also experimenting with advertising as an alternative revenue stream. Early tests of ads inside ChatGPT have performed well enough that leadership is increasing the volume. CFO Sarah Friar says ad revenue could help cover costs for the 92 percent of ChatGPT consumer users who do not pay for a subscription. The company is also testing “Sponsored Agents,” a format where clicking an ad drops users into a brand-hosted AI experience. That is a direct attempt to monetize agentic workflows, not just chat sessions.<\/p>\n

What are Sponsored Agents?<\/h3>\n

Sponsored Agents are a proposed advertising format that drops users into a brand-hosted AI experience after they click an ad. Instead of sending people to a traditional landing page, the ad opens a conversation with an AI system built around the brand. For OpenAI, this could turn advertising into an interactive product rather than an interruption, and it gives brands a reason to pay for access to the same agentic infrastructure that powers ChatGPT Work.<\/p>\n

This combination of a research-oriented AGI timeline and a commercial agent push is not a contradiction. OpenAI’s definition of AGI is built around economic value, and economic value is built around tasks. ChatGPT Work and Astra’s research-intern capabilities are two ends of the same spectrum. The company is trying to prove AGI by making it useful in the workplace, then charging for that usefulness through subscriptions, advertising, and sponsored agent experiences.<\/p>\n

All of this means the end of 2026 is less a moment of discovery and more a management decision. OpenAI has chosen a definition of AGI that it can hit, built a model family that makes that definition plausible, and arranged a product line that can, in principle, monetize the result. Whether that satisfies the rest of the field is another question. But by Altman’s definition, the announcement is not empty optimism. It is a testable business plan, and enough of the company’s leadership appears to believe the deadline will hold.<\/p>\n","protected":false},"excerpt":{"rendered":"

OpenAI has confirmed it expects to reach artificial general intelligence (AGI) by the end of 2026 under Sam Altman’s definition. The company’s leadership is more publicly confident than ever: Altman says OpenAI is “not quite yet” there, Chief Research Officer Mark Chen puts the distance at “80 percent of the way,” and co-founder Greg Brockman […]<\/p>\n","protected":false},"author":7,"featured_media":77978,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"fifu_image_url":"https:\/\/pub-4d4fc17555de4152be07eaf2a416a31e.r2.dev\/en\/ocie_1787777718582.jpg","fifu_image_alt":"OpenAI Confirms AGI by End of 2026 Under Altman's Definition","footnotes":""},"categories":[31],"tags":[],"class_list":["post-77966","post","type-post","status-publish","format-standard","has-post-thumbnail","category-technology"],"fifu_image_url":"https:\/\/pub-4d4fc17555de4152be07eaf2a416a31e.r2.dev\/en\/ocie_1787777718582.jpg","fifu_image_alt":"OpenAI Confirms AGI by End of 2026 Under Altman's Definition","_links":{"self":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/77966","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/comments?post=77966"}],"version-history":[{"count":0,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/77966\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media\/77978"}],"wp:attachment":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media?parent=77966"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/categories?post=77966"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/tags?post=77966"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}