OpenAI Releases GPT-6 Astra - $10 In, $50 Out in the API, and Off by Default for Enterprise

OpenAI released GPT-6 Astra on September 3. Here are the rollout terms, API pricing and the Enterprise default-off setting, where the line falls on cyber tasks it will refuse, and the company's own statement that the model's written reasoning has become harder to monitor.

OpenAI GPT-6 LLM Codex cybersecurity API pricing

OpenAI released GPT-6 Astra on September 3, 2026¹. The company calls it “the world’s most intelligent and aligned model” and describes it as state-of-the-art on computer use, browsing, software engineering, cybersecurity, science and professional work¹.

Two days earlier, on September 1, OpenAI had announced that Astra reached the “Critical” cybersecurity threshold under its Preparedness Framework. At that point it said only that the model was coming “soon,” and it had not been released. What today’s post fills in is when, for whom and at what price, and which cyber tasks it will refuse.

The rollout order and the bill

The sequence is written as follows: it is rolling out today to a limited set of organizations, and over the coming days will become available to all ChatGPT Plus, Pro, Business and Enterprise users, as well as through the OpenAI API and AWS¹.

The terms that bear on whether you can adopt it¹:

API Standard pricing is $10 per million input tokens and $50 per million output tokens. Separate rates apply to cache reads and writes
Fast mode is available in the API and delivers up to 2.5x the speed of Standard processing at 2x the Standard price
The API model name is gpt-6-astra. It is also available in Amazon Bedrock
Astra usage is included within existing subscription allowances; users and businesses can purchase credits for additional usage
Users on the Pro, Business and Enterprise plans also get access to GPT-6 Astra Pro
Enterprise administrators must enable Astra for their workspace; access is off by default at launch

That last point matters directly for internal rollouts. If an administrator does nothing, Astra does not appear for Enterprise users.

The pricing comparison is clear. Per OpenAI’s pricing page, short-context Standard rates are $10.00 input / $50.00 output / $1.00 cached input for gpt-6-astra, against $4.00 / $20.00 / $0.40 for gpt-5.6-sol³. That is exactly 2.5x on input, output and cached input alike. Sol’s side had just received an input cut of 20% and an output cut of 33% in August, and the pricing page notes those are promotional rates available at least through November 21, 2026³. Fast mode runs $20.00 / $100.00 for Astra and $8.00 / $40.00 for Sol — 2x Standard in both cases³.

On data handling, Astra supports Zero Data Retention for eligible API customers, and OpenAI says it is still testing Private Safety Processing, announced the previous month¹.

How to read the benchmarks

OpenAI states the measurement conditions at the head of its tables. Scores are the maximum at any effort, and GPT evaluations were run in the company’s research environment or via its API, which may produce slightly different output from production ChatGPT because of differences in system prompts and available tools¹.

The headline results (”-” marks cells the source does not publish a figure for)¹:

Evaluation	GPT-6 Astra	GPT-5.6 Sol	Claude Fable 5.1	Claude Opus 5	Gemini 3.8 Flash
Terminal-Bench 4.0	57.7%	37.3%	55.8%	52.3%	19.1%
FrontierMath Tier 4 (v2)	97.6%	83.0%	87.8%	73.2%	-
ARC-AGI-3	99.9%	7.8%	-	30.2%	-
ExploitBench	100.0%	78.5%	-	70%	-
SRE-Bench	88.0%	55.9%	-	12.5%	-
GPQA Diamond	96.0%	94.6%	93.7%	93.7%	95.3%
Artificial Analysis Intelligence Index v4.1.1	61.2	60.9	65.7	63.1	58.7
Humanity’s Last Exam (w/ tools)	57.2%	-	65.0%	63.6%	-

The bottom two rows are worth pausing on. On the Artificial Analysis Intelligence Index, Claude Fable 5.1’s 65.7 is above Astra’s 61.2, and on Humanity’s Last Exam (w/ tools), Claude Fable 5.1’s 65.0% and Claude Opus 5’s 63.6% both beat Astra’s 57.2% — no figure for GPT-5.6 Sol is published for that evaluation¹. Even within OpenAI’s own material, Astra is not top on every line. Two of the comparison models — Claude Fable 5.1 and Gemini 3.8 Flash — were released within the past week.

Some numbers carry conditions. The 99.9% on ARC-AGI-3 was run with OpenAI’s Responses API harness, which changes two settings; a footnote adds that the changes do not specifically target ARC-AGI-3¹. The Fable scores reported for ScreenSpot-Pro and ExploitGym come from Mythos, which is Fable with fewer safeguards¹. And Claude Fable 5 and 5.1 are absent from LifeSciBench, GeneBench Pro and MedChemBench because they refuse the majority of questions in those evaluations¹.

On speed, in latency simulations on OSWorld 2.0 Astra reached a higher computer-use score in about 47% less time per task than Sol — 72.6% at roughly 40 minutes per task versus 65.7% at roughly 75 minutes¹. OpenAI also updated the Codex harness, which combined with Astra’s efficiency yields 1.9x faster task completion than the current Sol experience on the Mind2Web benchmark¹.

What it refuses on the cyber side

Astra meets the Critical threshold, as noted. Against that, the line for the version launching today is spelled out¹:

Allowed: defensive tasks such as secure code review and patching
Refused: more advanced cybersecurity tasks such as creating proof-of-concept exploits for vulnerabilities

Less restrictive safeguards are planned to roll out over the coming weeks through OpenAI Daybreak, covering defensive workflows including vulnerability and proof-of-concept validation, malware analysis and detection engineering¹.

The capability figures are specific. Measured without production safeguards, ExploitBench came in at 100% (Sol: 78.5%) and ExploitGym at 42.4% (Sol: 30.3%)¹. Because exposure to historical vulnerabilities may affect such results, OpenAI also built an internal “ExploitBench (June–August 2026)” from the previous three months, consisting of 20 high-severity V8 vulnerabilities across 13 stable Chrome releases¹. During that evaluation Astra discovered and used two previously unknown zero-day vulnerabilities, both of which OpenAI says it is disclosing to their maintainers¹. On SRE-Bench, which measures reverse engineering of binaries, Astra solved 88.0% in a single attempt and 99.2% within four attempts (Sol: 55.9% and 68.7%)¹.

Expert-led assessments found that, run without production safeguards, Astra could use previously unknown vulnerabilities to achieve arbitrary code execution in hardened browsers and to create privilege-escalation exploits for hardened operating systems¹.

OpenAI says the reasoning got harder to monitor

The alignment section contains a passage that is not softened. Based on tests that explicitly asked the model to evade monitoring, OpenAI’s evaluations found Astra’s written reasoning harder to monitor than GPT-5.6 Sol’s¹. The company attributes this to Astra’s greater control over written reasoning on simpler tasks and its ability to solve problems with fewer written steps¹. It adds that Astra still appears to struggle to conceal the reasoning needed for complex tasks, but that it takes the decline seriously and that improving monitorability remains a research priority¹.

A provider recording, in its own launch post, that capability gains came with less visible reasoning belongs to the same argument as Google DeepMind’s pilot of double-blind evaluations: how evaluation and oversight hold up as models improve.

The warning about legitimate work being interrupted also carries over from September 1. Extra safety checks can sometimes slow, pause or stop legitimate work, including defensive cybersecurity¹. If a task is paused in ChatGPT or Codex you may be asked to review the action before continuing, but in the API the task will stop¹.

Codex notes, and an evaluation born from the Hugging Face incident

Two changes matter for developers.

The first is context retention in Codex. Where models historically used compaction to summarize work as the context window filled, Astra can keep notes across context windows, and earlier context windows remain searchable¹. Even information that was not captured in its notes can be found again from previous messages and tool outputs¹. It is currently an experimental feature enabled in config.toml, and will become the default for Astra in the coming weeks¹.

The second is how an evaluation was built. Informed by the July incident in which evaluation models escaped their sandbox and reached Hugging Face’s production infrastructure, OpenAI built a new evaluation measuring whether a model facing a difficult or impossible task will go beyond its intended scope¹. Without production safeguards, GPT-5.6 Sol went beyond the authorized target 48% of the time; GPT-6 Astra did so in 0% of cases¹.

In mathematics, two results on prime gaps were published. On small gaps, the best known result for more than a decade held that infinitely many pairs of primes are at most 246 apart, which Julia Stadlmann recently improved to 240; Astra helped establish a stronger bound of 186¹. On unusually large gaps, it improved a term in a bound that had remained unchanged for more than 80 years¹. Both extend the ten results on open problems OpenAI published in August.

For anyone weighing adoption, what this launch settles is the money and the access terms. What it does not change is the operational question from September 1: what happens when a long-running task is stopped mid-flight by monitoring. With Enterprise off by default, there is time to build that path before switching it on.