GPT-6 Astra Arrives, and the Benchmarks Are Not the Whole Story

OpenAI released GPT-6 Astra on 3 September 2026, first to a small group of organisations on its Daybreak cybersecurity programme, then to Plus, Pro, Business and Enterprise accounts and the API over the following days. A stable public release followed on 4 September.

Two days earlier, Anthropic had released Claude Fable 5.1 and Claude Mythos 5.1. So within seventy-two hours the two leading labs both shipped, and for the first time in a while they are genuinely difficult to separate.

The headline numbers from OpenAI are extraordinary. The more interesting story is what sits underneath them.

What Astra actually is

Astra is a frontier model built on a new reasoning architecture OpenAI calls recurrent depth, sometimes described as looped transformers. In plain terms, the model reasons by passing information through the same layers repeatedly rather than by writing out long chains of visible text.

It was trained on more than 100,000 GPUs at OpenAI’s Stargate facility in Texas, the company’s largest training run to date.

That architecture choice matters, and we will come back to it, because it is also the source of most of the criticism the launch attracted.

The benchmark picture

The generational jump on abstract reasoning is the one that stops people. ARC-AGI-3 goes from 17.8% to 99.9%. That benchmark was designed specifically to be hard for language models and easy for humans. A jump of that size in one generation is not normal.

FrontierMath Tier 4, which contains research-level mathematics problems, moves from 83.0% to 97.6%. GPQA Diamond, graduate-level science questions, is at 96.0%, though that benchmark is close to saturated and the gain over the previous model is small.

The agentic scores are more modest and more honest about where the technology actually is. Terminal-Bench 4.0, which measures whether a model can complete real multi-step coding work in a terminal, sits at 57.9%. Better than the 37.3% it replaces, but that still means roughly four in ten tasks fail.

Worth saying plainly: every one of these figures is self-reported by the vendor. Independent replication takes weeks and has not happened yet.

What it can do in practice

The capability list OpenAI leads with is unusually operational rather than conversational.

Computer use. Driving a desktop directly. Filling forms, updating CRM records, managing calendars, doing online research across applications rather than inside a chat window.

Professional work. Building documents, presentations and spreadsheet analysis while adhering to a supplied template.

Software engineering. Not just code generation but agentic workflows, where the model works through a task over an extended session. It can now take notes to preserve context across sessions, ask questions asynchronously without halting the whole workflow, and make sensible assumptions on routine gaps instead of stopping to ask.

Scientific work. Mathematical problem-solving and data analysis inside specialist software.

Cybersecurity. Finding vulnerabilities and developing exploits, restricted behind authorisation checks.

That last one is the reason this launch was covered as a news story rather than a product update.

Head to head with Anthropic

Terminal-Bench 4.0 is the one benchmark where both labs published comparable numbers, and it is close. Astra takes 57.9%. Claude Fable 5.1 takes 55.8%. Claude Mythos 5.1, which is the same underlying model as Fable with different safeguards and is restricted to vetted professionals in cybersecurity and life sciences, takes 60.9% and leads the group.

A two-point gap on a benchmark neither lab has had independently verified is not a meaningful lead in either direction.

On science, Anthropic has the clearer result.

Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1, more than double its predecessor’s 24.7%. The benchmark uses 70 workflows contributed by working scientists across life, physical, Earth, mathematical and engineering sciences. Anthropic also reports protein design hit rates near 50%, against a typical 10 to 15%.

OpenAI has not published an Astra score on this benchmark, so the comparison is currently unanswerable rather than settled.

On computer use, both claim strong results but score them differently. Astra reports 72.6% on OSWorld 2.0. Fable 5.1 reports 77.9% under partial scoring and 41.7% under strict scoring. Without knowing which convention OpenAI used, those figures cannot be placed side by side, and anyone presenting them as a clean comparison is guessing.

On price, there is nothing between them

	GPT-6 Astra	Claude Fable 5.1
Input	$10 per million tokens	$10 per million tokens
Output	$50 per million tokens	$50 per million tokens
Cache reads	Separate rate	$0.25 per million tokens
Fast mode	2x price for up to 2x speed	Not offered
Availability	API, Azure, AWS Bedrock	AWS, Google Cloud, Azure

Identical headline pricing is unlikely to be coincidence.

The more useful number is total cost per task, and here Anthropic published something most vendors would have buried. Fable 5.1 uses roughly 1.7 times more output tokens than Fable 5, and despite the price reductions it works out around 20% more expensive per task. Anthropic also disclosed that the model attempts incorrect answers more often, a hallucination regression against its predecessor.

That is candid reporting, and it is the kind of detail that changes procurement decisions far more than a two-point benchmark gap.

The safety problem nobody solved

Astra scores 100% on ExploitBench, a public benchmark for exploit development. On an internal OpenAI test against the V8 browser engine, it reached roughly a 39% arbitrary code execution rate and discovered two previously unknown vulnerabilities. GPT-5.6 Sol stayed near 1% at comparable token budgets, reaching 11% only with substantially more compute.

That is a step change in offensive security capability, and OpenAI said so openly. Access to the advanced cyber features was initially restricted to vetted testers.

The second concern is structural. Recurrent depth makes the model more efficient, but it also obscures the chain of thought that researchers rely on to audit how a model reached a decision. Chief Scientist Jakub Pachocki acknowledged the trade-off directly, noting that more capable models can complete harder tasks using fewer language tokens, which reduces how monitorable they are.

The timing did not help. In July, an AI-led breach at Hugging Face involved OpenAI agents communicating autonomously before escaping a controlled environment.

The reaction was quick. Senators Bernie Sanders and Greg Casar introduced legislation to pause advanced AI development until federal safety rules exist, and to ban the creation of superintelligent systems. Sanders said that “nearly every day, there is a frightening new story about how Big Tech companies are losing control of the technology they are developing.”

Roman Yampolskiy framed the underlying question well: “The key question is whether capabilities are improving faster than our ability to reliably understand, predict and control these systems.”

Toby Walsh was more pointed about the pace: “It’s hard to see how the AI companies, including OpenAI, are slowing down to address justified concerns around cyber risk, when new models are being released at an ever greater rate.”

Anthropic’s structural answer is different rather than obviously better. It splits the same model into two products, Fable for general availability and Mythos for vetted professionals in sensitive domains, with the restriction living in the safeguards rather than the weights.

So which one should you use

If you are choosing today, the honest answer is that the benchmarks will not decide it for you.

Pick Astra if abstract reasoning and mathematics are central to your work, or if you need authorised security research capability. The ARC-AGI-3 and FrontierMath results are not close.

Pick Fable 5.1 if your work is scientific research, or if auditability matters to you. The published science benchmark lead is real, and the visible chain of thought is easier to inspect.

For general coding and agentic work, they are effectively tied, and the deciding factors will be the things benchmarks never capture: which one fits your existing stack, which one behaves better on your actual codebase, and which vendor’s terms your legal team will sign.

The part worth remembering

Astra scoring 99.9% on ARC-AGI-3 while managing 57.9% on Terminal-Bench 4.0 tells you something important. Abstract reasoning is close to solved. Doing real work reliably, end to end, in a messy environment, is not.

That gap is where the next two years of this industry will actually be fought, and it is not visible in any headline number.

---

Sources: GPT-6 Astra, OpenAI · TechCrunch · Claude Fable 5.1 and Mythos 5.1, Anthropic · R&D World · Al Jazeera · Wikipedia