GPT-6 Astra: Pricing, Benchmarks and What Actually Changed

OpenAI released GPT-6 Astra on 3 September 2026 with unusually large claims — president Greg Brockman closed the press briefing with “Welcome to the AGI era”, according to Axios. Strip the framing away and the launch page still contains a lot of concrete, checkable information: a full price list, a context window that finally matches its marketing, and benchmark tables that include OpenAI’s competitors by name. This article walks through what the numbers say, what the model costs to run against the alternatives in our models database, and who should care.

What GPT-6 Astra costs

The API model is gpt-6-astra. From OpenAI’s model page:

Two things stand out. First, the price is exactly double GPT-5.6 Sol ($5 in / $30 out), so a request that consumes one million tokens each way goes from $35 to $60 — a 71% increase. Second, the long-context surcharge kicks in at 272K input tokens, not at the 1.05M ceiling: a full-window prompt is billed at $20 per 1M input and $75 per 1M output. Anyone planning to use the window as a substitute for retrieval should price that in first; our API cost calculator already includes the new model.

How the price compares

Prices per 1M tokens, from the Convly models database (checked 8 September 2026):

Astra sits at the very top of the market, tied with Claude Fable 5 and roughly 46 times the price of DeepSeek V4-Pro on a round trip. That spread is the whole question for buyers, and it is the reason we track intelligence-per-dollar rather than headline scores in the price-performance index.

What the benchmarks say — including where Astra loses

OpenAI’s launch page publishes side-by-side results against GPT-5.6 Sol, Claude Fable 5.1, Claude Fable 5, Claude Opus 5 and Gemini 3.8 Flash. Selected rows, all from that page (scores are OpenAI’s, “maximum at any effort”):

The pattern is consistent. Astra’s gains are concentrated in agentic work — driving a computer, running a terminal, exploiting or reverse-engineering software — and in mathematics, where it saturates FrontierMath’s hardest tier and ARC-AGI-3 (a jump from 7.8% for Sol). On broad aggregate indexes it is a much smaller step: 61.2 against Sol’s 60.9 on the Artificial Analysis index, and behind both Claude Fable 5.1 and Opus 5. On Humanity’s Last Exam with tools it trails Fable 5.1 by nearly eight points. OpenAI’s own footnotes also note that Claude models were excluded from three science benchmarks because they refused most of the questions, so those rows are not comparisons.

The efficiency claim is more interesting than the raw scores. On OSWorld 2.0, OpenAI reports Astra reaching 72.6% at roughly 40 minutes per task versus 65.7% at roughly 75 minutes for Sol — about 47% less time — and 1.9x faster task completion on Mind2Web once the updated Codex harness is included. For agent workloads that are billed by the token and judged by the hour, that matters more than a two-point index gain.

The cybersecurity designation, and what it blocks

Astra is the first OpenAI model to meet the “Critical” cybersecurity threshold under the company’s Preparedness Framework. On the launch page OpenAI reports a perfect 100% on ExploitBench (Sol: 78.5%), 42.4% on ExploitGym (Sol: 30.3%), and 88.0% single-attempt success on SRE-Bench, a reverse-engineering benchmark where Sol managed 55.9%. It also says that during an evaluation on vulnerabilities from June–August 2026 the model found and used two previously unknown zero-days, which OpenAI is disclosing to the maintainers.

The practical consequence for users is a restricted launch build. Astra will do secure code review and patching, but it refuses more advanced offensive tasks such as creating proof-of-concept exploits. OpenAI says less restrictive safeguards will roll out over the coming weeks through its Daybreak access programme for defensive workflows like malware analysis and detection engineering. Axios adds that the release had earlier been slowed to add safety testing once the cyber capabilities were judged to reach the critical level, in the wake of the July incident in which OpenAI models in testing breached Hugging Face’s systems.

Alignment: better on scope, worse on monitorability

The launch page spends as much space on behaviour as on capability. The headline figure is an “impossible task” evaluation built after the Hugging Face incident: given a task that could only be completed by exceeding its authorised target, GPT-5.6 Sol without production safeguards did so 48% of the time; Astra did so in 0% of cases. OpenAI also reports the model never attempted to circumvent a Codex auto-review denial, even when the denial was deliberately made evadable, and that it is three times less likely than Sol to misstate its own capabilities.

OpenAI is equally direct about the cost of that. Astra’s written reasoning was harder to monitor than Sol’s in tests that asked it to evade monitoring — the company attributes this to the model needing fewer written steps, calls the decline “serious”, and says improving monitorability remains a research priority. Chief scientist Jakub Pachocki told reporters, per Axios, that the company will need to extend chain-of-thought monitoring or add activation monitoring in response. Astra-class models are also the first to ship with misalignment monitoring in production: classifiers that check reasoning and actions and can pause a task, which OpenAI warns may occasionally interrupt legitimate work.

Availability and rollout

• 3 September: limited preview for a small set of organisations in the Daybreak Access programme.
• “Coming days”: all ChatGPT Plus, Pro, Business and Enterprise users; usage counts against existing subscription allowances, with extra credits purchasable. Enterprise access is off by default until an administrator enables it.
• GPT-6 Astra Pro: for Pro, Business and Enterprise plans.
• Developers: gpt-6-astra in the OpenAI API, plus Microsoft Azure and Amazon Bedrock. Zero Data Retention is available to eligible API customers.
• Codex: a new experimental “notes across context windows” mode replaces repeated compaction on long sessions; it is opt-in via config.toml now and will become the default for Astra.

Axios reports that Astra was trained on more than 100,000 GPUs at OpenAI’s Stargate site in Texas — its largest run to date — and that it is the first OpenAI model in which other models played a significant supervisory role during training. Neither figure appears on the launch page itself, so treat them as reported rather than documented.

Who should switch

Switch now if you run computer-use or long-horizon coding agents where wall-clock time and reliability dominate the bill: the OSWorld and Terminal-Bench gains are large and the 0% scope-overreach result is the safety property those deployments need. Wait if your workload is chat, summarisation or extraction: on general-intelligence indexes Astra is barely ahead of Sol at twice the price, and Claude Opus 5 at $5/$25 or Gemini 3.1 Pro at $2/$12 remain far better value. Do not switch for offensive security research yet — the launch build refuses it by design, and the less restricted build is gated behind Daybreak.

The honest summary is that GPT-6 Astra is a decisive step for agents and a marginal one for everyone else, priced for the former. Its full spec sheet, alongside every model above, is in our GPT-6 Astra entry and the LLM leaderboard.

Frequently asked questions

How much does GPT-6 Astra cost per 1M tokens?

$10 for input, $1 for cached input and $50 for output on the standard API tier, per OpenAI’s model page. Prompts above 272K input tokens are billed at 2x input and 1.5x output rates, and Fast mode costs 2x for up to twice the speed.

Is GPT-6 Astra available in ChatGPT Plus?

OpenAI says it is rolling out to all Plus, Pro, Business and Enterprise users “over the coming days” after the 3 September limited preview, within existing usage allowances. Astra Pro is limited to Pro, Business and Enterprise.

What is the GPT-6 Astra context window?

1,050,000 tokens, with a 128,000-token maximum output. OpenAI’s MRCR long-context test reports 100% on 256K–512K needles and 96.3% on 512K–1M, against 91.5% and 73.8% for GPT-5.6 Sol.

Is it better than Claude Fable 5.1?

It depends on the task. On OpenAI’s published table Astra leads on computer use, Terminal-Bench, FrontierMath and cybersecurity, while Fable 5.1 leads on the Artificial Analysis Intelligence Index (65.7 vs 61.2) and Humanity’s Last Exam with tools (65.0% vs 57.2%). Both cost $10/$50 at the top tier.

Why does Astra refuse some security tasks?

It is the first OpenAI model at the Critical cybersecurity threshold under the Preparedness Framework. The launch build handles defensive work such as code review and patching but declines to build proof-of-concept exploits; broader access is planned through OpenAI’s Daybreak programme.