ChatGPT 6 Astra Launches: Sam Altman’s Ambition and OpenAI’s AGI Claim
Astra’s September debut brings the AGI debate into the present—and raises the stakes for Claude, Gemini and Grok. The next question is how far beyond human intelligence these systems can go.
Capital Park · Singapore
For years, artificial general intelligence was something people argued about in the future tense. OpenAI’s September 3 launch of GPT-6 Astra has made that conversation feel much more immediate. A new model is here, the company is talking about an AGI era, and the competition to build machines that can handle the breadth of human intellectual work has entered a consequential new phase.
OpenAI’s release notes confirm the launch date and describe Astra as a model built to carry difficult assignments through to completion. It combines reasoning, coding, research, computer use and document creation. Often referred to as “ChatGPT 6 Astra,” its official model name is GPT-6 Astra; ChatGPT is one of the products through which people use it.
“Welcome to the AGI era,” OpenAI President Greg Brockman said at the launch briefing, according to Axios.
That attribution matters. Sam Altman is the company’s best-known advocate for general intelligence, but the two executives have not described the threshold in exactly the same way. In TIME’s reporting before the launch, Altman said OpenAI was “not quite yet” at AGI and expected an internal system he would consider AGI by the end of 2026. It would be misleading to turn that forecast into a verified Altman quote declaring Astra itself to be AGI.
Even with that distinction, this is a significant moment. The ambition behind Astra is to make a machine useful across an expanding range of work, with less instruction at every step. Whether that clears the AGI threshold depends on what we mean by the term—and how well the technology holds up outside a demonstration.
What artificial general intelligence actually means
Artificial general intelligence, or AGI, generally means a system with broad intellectual capability: it can learn, reason and solve unfamiliar problems across many fields at a level comparable to people. The word “general” does most of the work. A chess program can beat a grandmaster and still have no idea how to plan an experiment, interpret a contract or learn a new office application.
An AGI should be able to carry useful knowledge from one kind of problem to another. Imagine giving it a business question it has never seen before. It would need to decide what information is missing, study unfamiliar material, choose appropriate tools, test an explanation and revise its approach when the evidence changes. Success would involve more than producing a convincing paragraph.
There is no universally accepted finish line. Google DeepMind’s Levels of AGI framework treats the problem as a combination of breadth and performance, with autonomy considered separately. That distinction is useful. A system can be highly capable but need close supervision; another can run for hours while making avoidable mistakes. Neither speed nor independence, by itself, settles the intelligence question.
AGI also does not require consciousness, feelings or a human personality. Those are different questions. Nor would an AGI have to be flawless. People make mistakes too. The stronger test is whether a system can perform competently across a wide variety of unfamiliar tasks, recognize its limits and recover when its first approach fails.
Artificial superintelligence, or ASI, goes further. It refers to intelligence that substantially exceeds the best human performance across a broad range of cognitive abilities, including scientific reasoning, creativity and strategic problem solving. The distinction is between reaching broadly human capability and moving well beyond it. A spectacular result in one narrow field is not enough to establish either claim.
Why Astra changes the conversation
OpenAI’s Astra guide emphasizes sustained work across code, browsers and professional software. It also describes a more adaptable working process: a user can redirect an assignment while it is underway, and a connected application can let Astra continue useful work while waiting for a tool to finish.
Those details may sound modest beside the word AGI, but they address a familiar weakness in earlier assistants. Real work involves interruptions, incomplete information and changing priorities. A useful collaborator has to keep track of the goal through all of that. Better follow-through can matter more than an impressive first answer.
Astra supports roughly 1.05 million tokens of context, according to its official model page. That creates room for substantial codebases, documents and working history. Context capacity is not the same as understanding, but it can make complex assignments more practical. OpenAI’s current product guidance also places Astra inside Codex and ChatGPT Work, where the outcome can be a checked workflow or a finished file rather than a conversation alone.
The importance of this release is therefore easier to see in a workday than in a slogan. Can the model investigate a question, build something useful, notice that it is wrong and fix it? Can it do that across different subjects without requiring a person to rebuild the process each time? Those are the questions that connect a new model launch to the much larger AGI ambition.
Astra vs. Claude vs. Gemini vs. Grok
There is no single public score that establishes which of these systems is the most intelligent in every setting. The models run with different tools, safeguards, time budgets and reasoning settings. Their release dates differ, too. The fairest comparison is to look at what each lab is advancing and what that means for actual work.
Astra is OpenAI’s bid for broad, sustained execution. Its combination of reasoning, computer use and work across professional software makes it a serious contender for demanding assignments that cross several applications. Our reading of the product is that OpenAI wants the same general model to follow a problem from investigation through delivery. The practical test will be how consistently it does that when the task is unfamiliar and the instructions are incomplete.
Claude is pushing hard on coding and research. Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1 on September 1. They share an underlying model but have different safeguards and access arrangements: Fable is generally available, while Mythos is restricted to approved programs. Anthropic highlights lengthy engineering assignments, knowledge work and scientific research, including experimentally tested protein designs. These are company-reported results, but they show why Claude belongs in the same serious capability debate. Astra’s arrival does not erase that progress.
Gemini is making capable intelligence cheaper to put to work. Google launched Gemini 3.8 Flash and 3.8 Flash Cyber on September 2. Flash emphasizes reasoning, software engineering and extended agent tasks at a relatively low price; the Cyber version is available through a program for trusted defenders. Google says Flash improves on its predecessor while retaining the same introductory pricing. Its competitive importance is scale: a model that is economical enough to run repeatedly can change more everyday workflows than a more expensive system used only for exceptional problems.
Grok is competing through persistent work and fast iteration. Grok 4.6, released on August 12, focuses on extended agent tasks, software development and interactive or visual projects. Its surrounding ecosystem includes Grok Build and Grok Bot, the latter built around recurring agents that work across tools. That gives the Grok offering a practical angle: helping users organize ongoing work. Its August launch comparisons, however, do not establish a ranking against Astra or the newer September releases.
| Model family | Current focus | Practical value | What still needs proving |
|---|---|---|---|
| GPT-6 Astra | Reasoning and sustained work across software | One model carrying a complex assignment from investigation to delivery | Reliable generalization across unfamiliar work with limited supervision |
| Claude Fable / Mythos 5.1 | Coding, knowledge work and scientific research | Deep investigation and execution on demanding projects | How research gains translate into repeatable results across domains |
| Gemini 3.8 Flash | Reasoning and agent workflows at lower cost | Economical deployment across frequent, demanding tasks | The balance between cost, persistence and quality on the hardest assignments |
| Grok 4.6 | Extended agent work and interactive development | Turning ideas into working projects and supporting recurring workflows | Its standing against the latest releases under the same test conditions |
The table reflects our interpretation of current product directions, not an independent benchmark ranking. All four labs are pursuing overlapping abilities. None has a monopoly on reasoning, coding, tool use or scientific ambition.
What the benchmark scores actually show
The numbers make this race more interesting, not less. A model can lead a broad intelligence index and still trail on a particular coding task. To keep those distinctions clear, the comparisons below separate an independent index from results published by the developers. This is a September 5, 2026 snapshot of specific model versions, not a permanent ranking of the four brands.
An independent overall comparison
Artificial Analysis combines ten evaluations into its Intelligence Index v4.2, covering agents, coding, scientific reasoning and general capabilities. The suite is primarily text-based and English-language; it is not a complete test of every ability.
| Model | Reasoning configuration | Index score |
|---|---|---|
| GPT-6 Astra | Max | 55 |
| Claude Fable 5.1 | Adaptive reasoning, max effort, default fallback | 57 |
| Gemini 3.8 Flash | High | 47 |
| Grok 4.6 | High | 51 |
Sources: Artificial Analysis’s Astra model page, Claude Fable 5.1 model page and Gemini–Grok comparison, accessed September 5, 2026. Scores are displayed whole-number index points, not percentages. Different reasoning configurations are not equal computing budgets. These v4.2 figures should not be mixed with the older v4.1.1 scores in launch announcements.
Five tests, a more complicated picture
The next table brings together published task results. Astra, Claude and Gemini figures come from OpenAI’s Astra release comparison; Grok’s two available figures come from its separate launch report. These are developer-reported snapshots, not a single independent test under identical conditions.
| Benchmark | GPT-6 Astra | Claude Fable 5.1 | Gemini 3.8 Flash | Grok 4.6† |
|---|---|---|---|---|
| DeepSWE v1.1Software engineering | 74.1% | 67.4% | 73.8% | 65.9%† |
| FrontierCode 1.1 ExtendedCoding task score | 64.5%‡ | 63.6% | 56.3% | 61.3%† |
| Terminal-Bench 4.0Work in a terminal environment | 57.9% | 55.8% | 19.1% | Not reported |
| GPQA DiamondScientific knowledge and reasoning | 96.0% | 93.7% | 95.3% | Not reported |
| Humanity’s Last ExamWith tools | 57.2% | 65.0% | Not reported | Not reported |
Sources and settings: OpenAI’s GPT-6 Astra release tables report the best score across reasoning efforts; GPT tests used research or API environments, which can differ from ChatGPT. †Grok’s August 12 release report uses Grok 4.6 High. Its Terminal-Bench result is version 3.0, so it is excluded from the 4.0 row. ‡Astra’s FrontierCode run used an additional Codex-style developer message, described in OpenAI’s footnote 8. “Not reported” means absent from these cited comparisons, not a zero score. All sources checked September 5, 2026.
Claude leads the index snapshot and the reported Humanity’s Last Exam comparison. Astra posts strong coding results, while Gemini is close on DeepSWE and GPQA. Grok has fewer comparable results in these sources. The practical takeaway is to choose around the work you need done, then test it yourself. None of these scores, alone, certifies AGI.
The race is becoming a test of judgment
As the models improve, it gets harder to judge them through short questions alone. A system can solve an advanced mathematics problem and still misunderstand a routine instruction. It can write excellent code while missing a business requirement. Intelligence in the workplace includes knowing which problem matters, when more evidence is needed and when a confident answer should be questioned.
That is why the next useful comparisons should measure complete assignments. Give each system the same materials, tools, permissions and budget. Record whether it finished, whether the result was correct, how often a person had to intervene and how much the work cost. Repeat the exercise on unfamiliar tasks. A pattern of dependable performance would tell us much more than a carefully chosen demonstration.
The same logic applies to teams of agents. Several specialists can explore ideas in parallel, check one another’s work and bring different approaches to a difficult problem. They can also repeat the same error or waste time passing incomplete information around. A stronger lead agent should make the team more coherent: assigning clear responsibilities, reconciling conflicting evidence and keeping the original goal in view.
Our expectation is that this ability to coordinate will become one of the defining advantages of the next generation. The economic value will come from dependable output and better decisions, with people able to understand and direct the work. The number of agents involved will be a secondary concern.
From general intelligence to superintelligence
The most consequential possibility is that advanced AI begins to accelerate the research that produces its successors. A system that helps scientists formulate experiments, write software and analyze results could shorten part of the development cycle. Better models might then help build better tools and improve the next round of research.
That is a plausible route toward superintelligence, not a guarantee of an immediate leap. Computing power, energy, experimental validation and unanswered scientific questions still impose limits. A machine that proposes a promising idea has not automatically proved it. Progress will depend on whether those ideas survive contact with evidence.
Still, the change in perspective is striking. Astra, Claude, Gemini and Grok are competing over how much responsibility a machine can carry across real intellectual work. If their capabilities continue to broaden, the important milestones will increasingly concern discovery: useful knowledge, new techniques and solutions that people could not have produced as quickly on their own.
September 2026 may eventually be remembered as one of those turning points. Astra has given the AGI argument a concrete new subject, and its rivals are advancing within days and weeks of one another. Calling that competition historic is a reasonable judgment. Calling the scientific question settled would go further than the evidence currently allows.
If Astra’s capabilities hold up across the breadth of human intellectual work, this may be the moment we look back on and say: we have reached artificial general intelligence. The first great milestone arrived sooner than many expected, and history changed with it. That verdict still needs to be earned. But the ambition has already moved further: the race toward artificial superintelligence—ASI—is underway.