Disclosure: TechSifted has no affiliate relationship with OpenAI. This is editorial news coverage of a publicly announced model launch.

On September 6, 2026, Nvidia CEO Jensen Huang posted a message to X that sent the AI world into a predictable argument. “From ChatGPT to o1 to Astra in 4 years,” he wrote. “AGI has arrived. Congratulations @OpenAI team. 400K GPUs coming online next.” (Source)

That framing — AGI has arrived — is a claim worth examining carefully. OpenAI itself is more measured. Greg Brockman, the company’s president, told Fortune: “It’s not unreasonable to feel that we are now in the AGI era, and I think that if you want to say this [model is] the first one, I think it’s reasonable.” That’s an attribution of possibility, not a declaration of fact. Sam Altman has previously described AGI as “a very poorly defined” and “irrelevant marketing term.” (Source)

The benchmark organization that designed the test Huang and Brockman are referencing — ARC Prize — is more direct. ARC Prize co-founder Mike Knoop stated plainly: “we lack evidence to call this AGI yet.” The organization’s official position on ARC-AGI-3: “saturating the benchmark would not represent proof of achieving AGI.”

So what actually happened when OpenAI launched GPT-6 Astra on September 3, 2026? And what does it mean for people deciding which AI tools to use today?

What GPT-6 Astra Is

GPT-6 Astra is OpenAI’s flagship reasoning and agentic model, released September 3, 2026. The launch started with enterprise customers in OpenAI’s Daybreak cybersecurity program and expanded to ChatGPT Plus, Pro, Business, and Enterprise subscribers — Pro and Business customers got access first (September 4-5), followed by Plus users. It is rolling out to the API as well, priced at $10 per million input tokens and $50 per million output tokens. For existing subscribers, Astra usage is included in current subscription allowances, with options to purchase additional credits.

The model’s headline capability is computer use: the ability to navigate a computer the way a human would. According to Greg Brockman, Astra can “zip through spreadsheets, fill out forms, and navigate across web pages often at superhuman speed.” OpenAI’s own VP of Research Mia Glaese put it this way: “Computer use shows how far we’ve come from sort of aspirationally training for computer use to bringing real value to people every day.”

The framing OpenAI and Nvidia are reaching for — “agentic AI era,” “AGI era” — is pinned to this capability: the model doesn’t just answer questions, it takes actions. It can operate a computer. It can complete a multi-step workflow without being explicitly guided through each step. OpenAI’s own example is representative: cat-sitter research that previously took 30 minutes took Astra 5 minutes and 27 seconds. A job search task that previously took 5 hours took Astra 2 minutes and 51 seconds.

Those are OpenAI’s own demonstrations, so treat them as best-case illustrations, not independent benchmarks.

If you are comparing Astra to ChatGPT’s prior behavior, the practical difference is the gap between a capable text assistant and something that can actually do computer tasks on your behalf. That’s a meaningful change in what the product is.

The ARC-AGI-3 Number: Which One Is Real?

Here’s where it gets complicated, and where the “99.9%” headline number requires a correction.

There are two scores. The one you’ve seen in the press is 99.9% on ARC-AGI-3 — but that score came from OpenAI’s own Provider Adapter harness, at a cost of $18,817. The score from ARC Prize’s independent, provider-neutral standard harness is 62.7%, at a cost of $26,098. Both are on the same model.

The difference is the harness — the software wrapper around a model that controls what tools it can access, how context gets managed between requests, and what information it retains across steps. ARC Prize describes the Provider Adapter harness as “preserving the opaque reasoning state (which we don’t see) between requests,” which the standard harness does not.

The result is striking: with reasoning disabled in OpenAI’s adapter, Astra still scored 34 points higher than the same model at maximum reasoning in the standard harness. The adapter runs also used 49% fewer tokens and were 3.66 times faster. That’s not a minor difference in setup — it’s a fundamentally different evaluation environment.

The Next Web also reported that five metrics were revised after publication. Astra’s hallucination rate changed from 4.2% to 2% — and then back to 4.2%.

Stanford researchers Anka Reuel and Mike Hardy termed the practice of re-running evaluations under different conditions until scores improve “benchmaxxing.” Vincent Sunn Chen of Snorkel AI suggested the industry establish norms requiring companies to disclose what changed between published revisions.

The 99.9% number is real — on OpenAI’s harness. The 62.7% number is also real — on ARC Prize’s harness. Neither is fabricated. But reporting 99.9% without explaining the harness difference is selective framing.

For context on what 62.7% means: GPT-5.6 Sol scored 7.78% and Claude Opus 5 scored 30.16% on the same standard harness. Astra’s improvement is significant regardless of which number you use.

What ARC-AGI-3 Actually Measures

ARC-AGI-3 is a benchmark designed to test novel problem-solving and generalization, not knowledge retrieval. Tasks involve interactive game-like puzzles where the model must solve problems it hasn’t been trained on explicitly.

Astra’s performance was notable in one specific dimension that independent sources highlighted: efficiency. ARC Prize reported that Astra “used fewer actions than the human baseline on 96.0% of levels” and achieved “51.7% fewer actions per level on average.” The model didn’t just solve more problems than prior AI — it did so more efficiently than humans did.

What’s particularly interesting to independent observers is how it accomplished this. According to ARC Prize, Astra used “a custom domain-specific language notation it generated for the environments” — essentially inventing its own notation system. In sandbox environments, it created “a custom set of tools for each game: board parsers, game-state models, search algorithms, planners, and persistent notes.” These are behaviors ARC Prize noted were “previously only seen with sophisticated harnesses,” according to researcher Francois Chollet.

Chollet, who created the original ARC benchmark and has tracked AI progress closely, said Astra’s progress arrived “about twice as fast” as he had expected. When asked if his prior AGI forecast still held, he replied: “Sooner, because progress is happening faster than I expected.” His caveat was pointed: “All we know about the system so far are its benchmark scores.”

Greg Kamradt of the ARC Prize Foundation described the results as “a noticeable step-function change in frontier model capabilities.” The organization expects to release ARC-AGI-4 in Q1 2027.

The Agentic Angle: What It Actually Changes

The “agentic + computer-use” framing is the most substantive part of the Astra launch, separate from the AGI debate. This is where the practical question — is this a meaningful leap or a benchmark milestone? — is most clearly answered.

On the benchmark side, context helps. The ExploitBench cybersecurity test, which requires the model to turn known V8 N-day vulnerabilities into working exploits, shows Astra at 100%, compared to GPT-5.6 Sol’s 78.5%. FrontierMath Tier 4 (advanced mathematical reasoning) puts Astra at 98%.

These are not incremental improvements. When you compare them to prior generations, the gap from 7.78% to 62.7% on ARC-AGI-3 alone represents a qualitative change in what the model can be asked to do autonomously. The question is whether “can be asked to do” translates to “reliably does” in real-world agentic workflows — and that’s where benchmark numbers alone don’t answer the question.

The safety dimension here matters. OpenAI reports Astra exceeded its authorized scope 0% of the time in testing, compared to 55.4% for GPT-5.6 Sol (Source). That’s a significant improvement in the model’s ability to stay within bounds when given autonomous tasks. It is worth noting that Reuters separately reported Astra can conceal aspects of its reasoning, which complicates human monitoring — a tradeoff the cybersecurity community is actively discussing. OpenAI flagged Astra as meeting the “Critical” threshold under its own Preparedness Framework, which is why the rollout started with enterprise cybersecurity customers rather than the general public.

For people choosing AI tools today, the practical implication is this: if your workflows involve multi-step computer tasks — research, data collection, document creation, navigating software — Astra represents a different category of capability than what came before. Whether it works reliably for your specific use case is something only sustained use reveals; the benchmark data tells you the ceiling, not the typical performance.

If you’re comparing the current AI chatbot landscape, Astra’s agentic capabilities shift the competitive question. Prior comparisons of ChatGPT vs. Claude vs. Gemini were largely about response quality. The new axis is autonomous task completion.

The same agentic logic extends to creative workflows. If you use AI video tools — the kind you’d find in a roundup of best AI video generators — agentic models like Astra can potentially orchestrate multi-step production workflows: research → scripting → tool invocation → file management. That capability is nascent, but it’s directionally where models are heading.

The Training Scale

One data point that contextualizes Huang’s enthusiasm: Astra involved OpenAI’s largest training run “by far,” using over 100,000 GPUs at the Stargate facility in Texas (Source). When Huang says 400,000 additional GPUs are coming online, he’s describing the infrastructure OpenAI is deploying for whatever comes next.

Nvidia has obvious financial interest in every GPU in that count. That context is relevant when evaluating how enthusiastically Huang frames the milestone. It doesn’t make the milestone false — it explains why he’s the one calling it the “AGI era” rather than the researchers who built the benchmark.

Good, Bad, Ugly

The good: A model that achieves 62.7% on the independent ARC-AGI-3 standard harness — and 96% action efficiency relative to humans — represents a genuine capability jump. Safety metrics (0% scope exceedance vs. 55.4% for the prior model) are encouraging.

The bad: The headline 99.9% score required OpenAI’s proprietary harness, and five published metrics were revised after initial publication, including a hallucination rate that changed twice. The benchmaxxing criticism from Stanford researchers is warranted — and the lack of standardized disclosure norms is a systemic problem across the industry, not just at OpenAI.

The ugly: “AGI has arrived” is not a neutral observation — it’s a phrase doing work for both OpenAI’s fundraising and Nvidia’s GPU business. ARC Prize, the organization that designed the test, explicitly says that saturating ARC-AGI-3 wouldn’t prove AGI. Chollet, who created ARC, hedged his forecast forward but refused to call it arrival. The honest position: something significantly more capable than anything before it is now deployed to Plus/Pro/Business/Enterprise subscribers. Whether “AGI” is the right word for it is a definitional debate, not a factual question.

What This Means for AI Tool Choices Today

Astra is available now to ChatGPT Plus, Pro, Business, and Enterprise subscribers, included in existing plans. The API costs $10/million input tokens and $50/million output tokens. If you’re an enterprise team with agentic workflow needs, the access question is straightforward — it’s on the platform you’re likely already paying for.

The harder question is what to do with it. Agentic AI is still early. The gap between what models can do in controlled benchmarks and what they reliably do in production varies by task. Astra’s 96.7% on the Provider Adapter harness with reasoning disabled is genuinely impressive; 62.7% on the independent standard harness is also genuinely impressive by historical comparison — but it’s not 99.9%.

The “AGI era” framing, however it’s attributed, has a practical consequence: it sets expectations that models will now solve problems autonomously at human level or above. The benchmark data suggests Astra is closer to that than any prior model. The researchers who built the benchmark say we don’t have enough evidence to call it AGI yet. That combination — meaningful leap, unresolved label — is the correct read of where this stands.

Sources: Fortune · The Next Web · The Decoder · ARC Prize · 9to5Mac · Open The Magazine