GPT-6 Astra is a major agentic and cyber step — not a uniform jump in general intelligence

OpenAI's GPT-6 Astra raises Terminal-Bench 4.0 from 37.3% to 57.9% versus GPT-5.6 Sol and reaches OpenAI's Critical cybersecurity threshold. Independent ARC Prize and Artificial Analysis tests confirm large gains in agentic efficiency, but also show harness dependence, an unchanged aggregate Intelligence Index versus Sol, and a higher price.

OpenAI released GPT-6 Astra on 3 September 2026 as a new frontier model for computer use, coding, research and long-horizon agentic work. The launch numbers are large enough to justify a Pulse article, but the most important result is not the slogan attached to them. Astra shows a real step change in several agentic capabilities, especially terminal work, computer use, long-horizon coding and cybersecurity. Independent measurements already confirm part of that step. They also show something more nuanced: the improvement is not uniform across every measure of general reasoning, and some spectacular scores depend strongly on the evaluation harness.

That distinction matters because OpenAI describes Astra as a new generation of intelligence and some commentary has framed the release as an AGI milestone. A benchmark suite cannot establish AGI by itself. The defensible claim is narrower and more useful: compared with GPT-5.6 Sol, Astra can execute a broader class of difficult tool-using tasks more efficiently and has crossed a qualitatively different cybersecurity capability threshold, while independent aggregate intelligence tests remain mixed.

Before → after

Measurement	GPT-5.6 Sol / reference	GPT-6 Astra	What changed
Terminal-Bench 4.0, OpenAI evaluation	37.3%	57.9%	+20.6 points on complex terminal tasks
ARC-AGI-3 Semi-Private, ARC Prize standard harness	—	62.7%	strong independent novel-environment result
ARC-AGI-3, provider-adapter harness	—	99.9%	near-saturation when opaque state is preserved
ARC-AGI-3 action efficiency	human median baseline	fewer actions on 96% of levels	human-level-or-better action efficiency on most tested levels
Artificial Analysis Coding Agent Index	about 65 for Sol max	67	modest score gain with much lower token use
Artificial Analysis Intelligence Index	61	61	no aggregate gain on this independent index
API list price	$4 input /$ 20 output per 1M tokens for Sol	$10 / $50	2.5× higher token price
OpenAI cyber classification	below Critical	Critical	new capability/risk regime under OpenAI's framework

The table shows why a single sentence such as “GPT-6 is much smarter” is technically inadequate. Some dimensions move dramatically; one independent aggregate index does not move at all.

Harness dependence is part of the result

ARC-AGI-3 is designed around novel interactive environments rather than static question answering. ARC Prize reports two very different Astra scores on the same Semi-Private benchmark: 62.7% with its standard harness and 99.9% with OpenAI's provider-adapter harness. The difference is not a contradiction. The provider adapter preserves opaque reasoning state between requests and supports longer-running compaction, while the standard harness exposes a more constrained state interface.

This makes the evaluation scientifically more interesting, not less. It demonstrates that the effective capability of a frontier agent is increasingly a property of the model + memory/state interface + tool harness, not merely a frozen neural network evaluated in isolation. The 99.9% figure is real under the provider-adapter conditions; presenting it without those conditions would be misleading.

ARC Prize also reports that Astra used fewer actions than the median tested human on 96% of levels. Qualitatively, it observed the model converting unfamiliar environments into compact symbolic world models and internal shorthand for planning. That is evidence for improved adaptation and state abstraction inside a bounded benchmark, not proof of open-ended general intelligence.

Coding agents: the clearer independent gain

OpenAI reports 57.9% on Terminal-Bench 4.0, compared with 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1 under its launch evaluation. Terminal-Bench combines software engineering, system configuration and data-analysis tasks performed through a terminal. A 20.6-point absolute gain is not a cosmetic benchmark increment. It changes the fraction of multi-step tasks an agent can complete end-to-end.

Artificial Analysis independently sees the same qualitative direction. In its Coding Agent Index, Astra reaches 67, around two points above GPT-5.6 Sol at maximum effort. More important than the small index delta is the efficiency change: in the Codex harness Astra uses roughly one third of the tokens of Sol at max effort. At that setting, Artificial Analysis reports approximately the same per-task cost as Sol despite Astra's much higher token list price.

This is a useful engineering distinction: the model is more expensive per token, but can be cheaper or similar per solved coding task if it needs far fewer tokens and retries.

General intelligence: mixed independent evidence

The picture is less dramatic on Artificial Analysis's broader Intelligence Index. Astra max scores 61, exactly the same aggregate score reported for GPT-5.6 Sol. Claude Fable 5.1 is higher on that index. Astra uses around 10% fewer output tokens, but because API pricing is 2.5× higher, Artificial Analysis estimates the max-effort run at roughly 75% more cost per task than Sol on this general benchmark mix.

The component tests are mixed. Astra improves Humanity's Last Exam and long-horizon AA-Briefcase analytical quality, and Artificial Analysis reports a large reduction in its hallucination metric. But it also records regressions on GDPval-AA v2 and several other component evaluations. A new generation can therefore be a major product and agentic advance without being Pareto-superior on every intellectual task.

The cybersecurity threshold is qualitatively different

The most consequential non-benchmark change may be cybersecurity. OpenAI classifies Astra as its first model to reach the Critical cybersecurity capability level under the company's Preparedness Framework. In OpenAI's definition, that means the model can, with appropriate tools and access, identify previously unknown vulnerabilities and develop exploitation paths across strongly protected systems without continuous human guidance.

This is an internal OpenAI classification, not a universal external standard, so it should not be treated as an independently certified property. But it still represents a change in deployment regime: OpenAI has added stronger monitoring, isolation and action controls, and explicitly warns that legitimate defensive work can be slowed or stopped by safeguards. Reuters independently reported the release in the context of heightened scrutiny of agent safety.

The scientific point is that capability and risk are now coupled. A model able to execute longer, more autonomous technical trajectories needs a different containment architecture than a model that mainly answers questions.

What may have changed architecturally

OpenAI attributes Astra to combined advances in pre-training, reinforcement learning and alignment, but it does not publish enough model internals to derive the performance gains mechanistically. One public systems change is clearer: Astra/Codex can preserve notes across context-window boundaries while keeping earlier windows searchable, reducing information loss from repeated compaction.

This matters for long-running software and research tasks. The old pattern is roughly:

long trajectory → context fills → compression → details lost → repeated rediscovery

The new system aims for:

long trajectory → durable notes + searchable prior windows → state recovery → continuation

That is a system-level memory improvement. It should not be confused with evidence that the underlying model has acquired human-like persistent memory.

Monitorability also regressed in one dimension

Astra's system card reports that its written reasoning is more controllable than GPT-5.6 Sol's under tests that ask the model to manipulate the form of its chain-of-thought. OpenAI treats this as a monitoring concern because a model with greater control over how it expresses reasoning may be harder to supervise through reasoning traces alone.

This does not show that Astra routinely hides malicious intent. It shows that a previous monitoring assumption—reasoning traces are difficult for the model to deliberately shape—has weakened. The capability gain therefore comes with a new measurement problem.

New state of the art

Astra changes the frontier most clearly in agentic execution, not in a single scalar notion of intelligence. It can solve substantially more terminal tasks, operate novel interactive environments with much better action efficiency, and has reached a cyber capability level that forces stronger deployment safeguards. Independent evaluation supports these claims, while also showing that Astra is not uniformly better: general aggregate performance can be flat, some tasks regress, and cost depends strongly on workload and harness.

What remains unresolved is equally important: broad independent replication of OpenAI's internal benchmarks, real-world reliability over weeks or months, the extent to which memory/harness improvements rather than the base model drive the gains, and whether the new cyber safeguards remain effective under broad deployment.

Evidence status

Deployed frontier model · official system card and vendor evaluations available · independent ARC Prize and Artificial Analysis evaluations available · broad real-world replication still early.

No benchmark in this article is treated as proof that AGI has been achieved.

Sources