OpenAI’s AGI Claim Faces Its Hardest Test in the Workplace

That omission does not erase Astra’s reported advances. It changes what the announcement can reasonably prove.

A model that performs well on mathematics, coding, science and computer-use tests may represent a major technical achievement. But general intelligence is a broader claim than excellence across a carefully selected set of evaluations. It implies that a system can transfer its abilities across unfamiliar situations, understand goals, handle ambiguity and produce work that people and organizations can use.

The missing evidence is GDPval, OpenAI’s own benchmark for economically valuable tasks. Its absence from the Astra launch creates a direct tension between the company’s expansive language and the practical proof that would help employers decide whether to delegate real work to the model.

What OpenAI’s launch demonstrates

In OpenAI’s launch announcement for GPT-6 Astra, the company presents the system as its most capable and aligned model. The announcement highlights performance across abstract reasoning, mathematics, coding, science, computer use, long-context work and safety.

That portfolio is designed to show breadth. Mathematics can indicate formal reasoning. Coding tests whether a model can translate instructions into executable systems. Science evaluations can probe knowledge and problem solving in technical domains. Computer-use tests are intended to show whether the model can interact with software rather than merely generate text.

“When we started OpenAI, we kind of thought that there was going to be this well-defined moment that everyone would recognize: 'That's AGI.' It hasn't played out like that. It's a much more gray, fuzzy thing.”

Long-context performance matters because many professional tasks involve large quantities of information. Safety evaluations address whether greater capability is accompanied by controls against harmful behavior. Taken together, these categories support the argument that Astra is not simply a stronger chatbot.

They do not, however, directly answer how well it performs as a junior analyst, contract reviewer, software engineer, financial associate, customer support specialist or healthcare administrator.

That distinction matters because benchmark success is not identical to workplace competence. A model may solve a difficult mathematical problem while failing to identify which part of a business request is important. It may write functioning code while misunderstanding the operational requirements of a project. It may navigate a computer interface while making a costly decision when the instructions are incomplete.

The launch announcement does not report a GDPval result. That is a notable gap because GDPval was created specifically to test the type of transfer that the AGI label implies.

GDPval measures the work companies actually buy

OpenAI’s GDPval research paper describes a benchmark built around 1,320 tasks spanning 44 occupations and nine U.S. economic sectors.

The design is different from a conventional academic test. Rather than asking only whether a model can produce a correct answer, GDPval asks it to create outputs comparable to professional deliverables. Expert graders then compare the model’s work with materials produced by professionals.

That setup shifts the evaluation from capability in isolation to usefulness in context. The central question becomes less “Can the model answer this?” and more “Can the model produce something an expert would accept as adequate?”

For companies, that is a more consequential test. A business does not usually purchase intelligence as an abstract property. It pays for completed work, such as a report, a legal document, a financial analysis, a technical design or a response to a customer.

GDPval also exposes a challenge that standard benchmarks can hide: professional tasks contain unstated expectations. A deliverable may need to be accurate, properly structured, concise, compliant with a process and suitable for a particular audience. It may need to use evidence correctly, preserve important details and avoid introducing risks that are not visible in a simple answer key.

Those conditions make workplace evaluation harder. They also make it more relevant to the commercial meaning of AGI.

The strongest criticism is about evidence, not ambition

The AGI Society’s evaluation is sharply critical of the claim. In its assessment of recent AGI claims, the organization argues that OpenAI President Greg Brockman’s claim does not specify a definition of AGI, a test for reaching it or an evidentiary threshold that the public can use to judge the assertion.

That criticism goes to the foundation of the announcement. If AGI has no stated operational definition, then almost any sufficiently impressive collection of benchmark results can be presented as evidence that the threshold has been crossed. The label becomes a conclusion before it becomes a testable proposition.

The Society’s argument is not that Astra’s results are meaningless. Its concern is that the public evidence does not establish the larger claim. A system can be better than previous systems in many domains without demonstrating the flexible, reliable and autonomous performance associated with general intelligence.

The absence of GDPval strengthens that criticism because the benchmark is closely aligned with the economic consequences of advanced AI. If OpenAI wants to argue that Astra marks a new era, workplace performance would be among the most persuasive evidence. Reporting those results would allow readers to connect abstract capability with measurable value.

Not reporting them leaves several possibilities open. Astra may perform strongly on GDPval but the result may not have been ready for publication. The company may regard the benchmark as unsuitable for the launch. Or its performance may be less striking than the results selected for the announcement. The available sources do not establish which explanation is correct.

That uncertainty is itself important. Selective disclosure can make a model appear more complete than the public evidence supports, even when every reported result is genuine.

An independent workplace comparison could narrow the gap

Artificial Analysis offers its independent GDPval-AA v2.1 benchmark, which is derived from GDPval and provides a separate comparison point for economically relevant agentic work.

The value of an independent benchmark is not that it automatically settles the AGI question. It is that it can reduce dependence on a company’s chosen launch portfolio. When a model maker selects the tests, metrics and framing, the audience sees the capabilities the company wants to emphasize. A separate evaluation can ask whether those strengths persist under a different presentation of workplace tasks.

The supplied information about GDPval-AA does not establish a definitive public workplace score for Astra. It does establish the importance of having a comparison point beyond OpenAI’s own announcement. For businesses, that kind of evaluation may eventually matter more than a record on an isolated reasoning test.

Agentic work introduces another layer of difficulty. A model may need to plan several steps, operate software, recover from errors and decide when it lacks enough information. Its final output can look polished while the process behind it contains hidden mistakes. Workplace benchmarks that assess finished deliverables are better positioned to expose those failures than tests that grade a single response.

The strategic question for the AI industry

OpenAI’s decision to frame Astra as the start of AGI raises the standard for what it should disclose. The more consequential the label, the less sufficient a collection of impressive but disconnected scores becomes.

This is not an argument for replacing mathematics, coding, science or safety evaluations. Those tests measure capabilities that matter. It is an argument for adding evidence that reflects how intelligence creates value outside a laboratory.

The industry has a commercial incentive to favor benchmarks that produce clear headlines. A new record in reasoning or coding is easy to communicate. Professional work is messier. Expert judgments can be slower, results can vary by occupation and failures can be difficult to summarize in one number. Yet those complications are precisely why real-world evaluation is necessary.

For companies considering deployment, the unanswered questions are practical. Can Astra complete a task without constant supervision? Does its work meet professional standards? How often does it make errors that an expert must correct? Does it save time after review, or merely move effort from writing to checking? Can its performance be trusted across different occupations rather than in a narrow demonstration?

The Astra announcement offers evidence of substantial capability. It does not answer all of those questions, and it does not report the GDPval result that could have connected its technical achievements to economic usefulness.

That leaves the AGI claim ahead of the workplace proof. Until the gap is closed, GPT-6 Astra may be understood as a powerful general-purpose model with unusually broad benchmark performance. Calling it the beginning of AGI remains a proposition that depends not only on what the model can do in tests, but on whether it can reliably do valuable work in the world.