Build1 distinct publisher3 min readPublished
Altman says an internal system he would call AGI arrives by the end of 2026. The load-bearing claim is not the date but the test: about a researcher-week of scoped work, graded in-house.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
The definition is doing the work here, not the date. OpenAI's own formulation of AGI is "highly autonomous systems that outperform humans at most economically valuable work" [6], which the Decoder describes as a heavily economics-driven reading [7], and the term has no precise agreed meaning to begin with [8]. The year-end claim is therefore not a claim about minds. It is a claim that one system clears an economic bar the vendor drew and grades.
What clears it, according to chief scientist Jakub Pachocki, is Astra: hand it an experimental idea and it writes the code in OpenAI's codebase, runs the experiment, and reports back with results [3]. Hand it a research paper and it covers what would have taken a human researcher about a week [4]. That is the most useful figure in the whole account, because it is the only one carrying a unit. One researcher-week, measured inside the codebase of the company doing the measuring.
Mark Chen's "80% of the way" [2] has no denominator that anyone can inspect, given that the target term is undefined [8]. Paired with Altman's end-of-2026 timing [1], it also implies the last fifth of the journey closing inside a single calendar year [16], which is a strong claim about the shape of a curve nobody outside the building can plot.
Note also that the milestone is internal. Altman's stated conviction is about a system he would call AGI, and he concedes they are "not quite yet" there [1]. Nothing in a customer contract changes on the day of declaration. The part a buyer can actually touch is the persistence work: Codex folded into ChatGPT in a project called "The Merge," which produced ChatGPT Work [10], with product lead Thibault Sottiaux saying the company is close to shipping something with "persistence and always-on execution" [11]. That is testable. Does the agent hold state across a multi-day task, and what does it do at hour thirty when a dependency moves.
The funding section constrains everything above it. CFO Sarah Friar puts 92 percent of ChatGPT consumer users outside any paid subscription [12], ads have tested well enough that leadership is ramping the volume [13], and Sponsored Agents would drop a user who clicks an ad into a brand-hosted AI experience [14]. The research-intern story and the ad-load story are running on the same balance sheet.
One more thing worth saying plainly: both writeups supplied to this desk are the same Decoder article [17], itself reporting Alex Heath's TIME piece built from more than two weeks of work and over 20 interviews with executives, employees, investors and rivals [9]. The capability claims that matter come from people who work there. The Decoder itself flags Altman's "invents new things in a way that matters" line as heavily debated [5], and reports researchers split on whether recursive self-improvement is near or already visible [15].
So the translation for anyone holding a vendor roadmap with AGI on it: what is being promised, under the seller's own definition, is a persistent agent that finishes about a researcher-week of scoped work in a codebase it knows. That is a real product and worth paying for if it lands. It is not the same purchase as a system that outperforms humans at most economically valuable work, and the gap between those two things is where the contract language should sit.
Ranked by verification strength, evidence, and original report placement.
The Decoder characterises OpenAI's AGI definition as a heavily economics-driven reading, with others arguing more is needed, such as a robust understanding of the world.
The term AGI has no precise definition, so its application is partly a matter of interpretation.
Astra is OpenAI's upcoming model family, announced last year as an automated AI research intern, and it also enables "persistent agents" that work on tasks on their own over longer periods.
OpenAI defines AGI as "highly autonomous systems that outperform humans at most economically valuable work."
The account comes from a TIME report by Alex Heath, who spent more than two weeks interviewing over 20 executives, employees, investors and rivals.
OpenAI is combining its Codex coding agent with ChatGPT in an internal project called "The Merge," which produced ChatGPT Work.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single secondary writeup, self-graded benchmark
Everything traces to one the-decoder.com article (supplied twice) relaying one TIME report by Alex Heath. The core capability claim - Astra clearing an automated research intern bar and absorbing a researcher-week of work - is an internal, in-house-graded benchmark with no task set, pass rate, or independent evaluation. Timeline and invention claims are executive forecasts on a term the same article says has no precise definition. Only the definitional and monetization statements are firm as reported facts.
Adjacent products shipping, core system unreleased
The subject of the claim - Astra - is an upcoming model family with no external availability, so there is no third-party usage to observe. The measurable adoption sits adjacent: ChatGPT Work exists out of The Merge, in-product ads are past early tests with volume being increased, Sponsored Agents are in test, and 92 percent of ChatGPT consumers are disclosed as non-paying. Persistence and always-on execution are 'close' but unshipped, and the puck device is a 2027 expectation.
AGI framing outruns the in-house test behind it
The headline proposition is an AGI arrival date, but the supporting artifact is a vendor-defined internal benchmark of roughly a researcher-week of scoped work, plus an executive expectation that Astra will 'invent new things in a way that matters' which the article itself calls heavily debated. Chen's 80 percent has no stated scale, and the recursive self-improvement premise is explicitly split among researchers. The publisher discounts the claim in its own headline ('if you accept his definition'), which limits the gap somewhat, but the distance between 'AGI by end of 2026' and the evidence offered remains large.
Strong vendor and definitional incentives on all sides
The AGI claim is made by the company that authored the definition it is measured against, in a report where the same executives describe the need to fund training, a 2027 device line and humanoid robots. The monetization block makes the commercial pressure explicit: 92 percent of consumer users pay nothing and ads are being scaled to cover them, with Sponsored Agents in test. The publisher's own incentive is visible too - the article closes with a paid subscription pitch - though it does flag the debate rather than amplify it.
Quotes are clear, verification is absent
Confidence is moderate-low. The attribution chain is transparent and consistent - named executives, a named TIME reporter, a described interview process - and the duplicate sources agree because they are the same text. But the cluster contains no independent corroboration, no primary access to the TIME piece, and no external measurement of any capability claim, so the definitional and monetization facts are more reliable than the timeline and capability assertions.
product
OpenAI's plan to hand everyone a coding agent leaves the hard part to the model1 distinct publisher
product
OpenAI prices its own guardrails: 20% more compute, plus a two-week training pause1 distinct publisher
product
OpenAI's CFO calls an IPO "another fundraise" while the model calendar slips2 distinct publishers
product
ChatGPT Work's real ask is your Slack, and somebody has to say yes on everyone's behalf1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
2 articles · August 26, 2026