GPT-6 Astra: Price, Benchmarks, Access

OpenAI launched GPT-6 Astra today, September 3, 2026. At the press briefing, president Greg Brockman first conceded that AGI remains a “gray, fuzzy thing”, then went ahead anyway: “I think it’s not unreasonable to feel that we are now in the AGI era.” He closed with the same line: “Welcome to the AGI era.”

The claim is arguable. What is not arguable is the rest of the package: a concrete model, a concrete price tag, and a benchmark table with an unusual density of footnotes.

One naming note first. If you are searching for GPT-5.6 Astra, that model does not exist. Astra is not a variant of GPT-5.6, it is its successor. GPT-5.6 shipped as Sol, Terra and Luna; GPT-6 currently ships as Astra and Astra Pro.

Let’s get to the numbers. 👇🏻

What Is GPT-6 Astra?

Astra came out of OpenAI’s largest training run to date. According to research lead Aidan Clark, it is the first time the company pre-trained on more than 100,000 GPUs, at its Stargate site in Texas. There is a second first here too: Astra is the first OpenAI release where earlier models played a significant role in supervising the training process.

The spec sheet:

Shipping a flagship without publishing the context window or the knowledge cutoff is unusual. For reference, GPT-5.6 Sol launched with roughly 1.05M tokens of context. Read those two rows as “not known yet”: any Astra spec table you find online with confident numbers in them has invented them.

You can measure what your own prompts actually cost with our token counter. On a $50 output tariff, token efficiency stops being an academic concern.

Benchmark Results 📊

Almost every number below is OpenAI’s own measurement. The company says that unless noted otherwise, models in its evaluations ran at maximum effort, which lifts scores but also raises latency and token use.

Three separate stories in one table.

In coding the gap is thin. Astra scores 74.1% on DeepSWE v1.1, the 113-task agentic coding test, against 70.8% for Sol. A clear internal jump. But the public leaderboard currently sits at roughly 74% for Gemini 3.8 Flash and Claude Opus 5, with Sol at 73%, and the uncertainty ranges overlap. On top of that, Meta reported 75.4% for Muse Spark 1.3 at its maximum reasoning setting earlier this week, though that setting is under safety review and is not generally available at launch.

The real gains are outside coding. On Terminal-Bench Science, 70 command-line research tasks across five scientific fields, Astra hits 64.6% where Sol managed 22.4%, Fable 5.1 52.6%, and the existing public leaderboard tops out at 30%. BenchCAD, which asks models to reconstruct CAD programs from rendered views, comes in at 95.9%. FrontierMath Tier 4 lands at 97.6%.

Read the footnotes. Epoch AI, which runs FrontierMath, says OpenAI funded the benchmark’s development and has exclusive access to part of it. OpenAI notes the Claude BenchCAD results used modified evaluation settings. Anthropic says its OSWorld number used a different release. This is a compilation with caveats, not a leaderboard.

The Asterisk on 98.6% ⭐

That is the number in the headlines: 98.6% on ARC-AGI-3. When the benchmark launched in March 2026, frontier models scored under 1%, and GPT-5.6 Sol sits at 7.8%. The jump is genuinely large.

ARC Prize’s own published results add the missing context:

Same model, same benchmark, three different answers. The difference is not the model but the system around it: OpenAI ran Astra through a Responses API harness that retains reasoning between turns and uses compaction to manage long context. The other models in the comparison ran under different setups.

That distinction matters more on ARC-AGI-3 than almost anywhere else, because the benchmark exists precisely to drop a model into an unfamiliar interactive environment with no instructions. How much of the environment gets solved by the model versus by the agent scaffolding around it is the whole question.

One pleasant surprise: higher reasoning levels cost less. Astra solves games in fewer actions, which cuts total model calls and tokens.

Computer Use Is the Actual News 🖥️

For Brockman, computer use is “a particularly important part of what’s new”. The model navigates spreadsheets, fills out forms and moves across web pages, in his words often at superhuman speed.

OpenAI demonstrated Astra working inside KiCad, Excel, Blender, Power BI and Unity, plus browser-based form entry and website QA.

The number attached to it is OSWorld V2-Offline: 72.6%, up from 65.7% for Sol. The more interesting figure is not the score but the clock: average time per task dropped from about 75 minutes to 40. On Mind2Web, the new Codex harness completed tasks 1.9x faster than the current Sol-based setup.

What changes in Codex

For developers this may matter more than any benchmark. Codex currently relies on compaction: it summarizes earlier work to free up context. The problem is that the summary tends to discard exactly what the agent needs later, such as why a previous fix failed, which tests ran, or a small requirement the user mentioned at the very start.

Astra can instead keep notes across context windows and search earlier messages and tool output. The feature is experimental behind a config.toml setting for now; OpenAI says it will become the default for Astra in the coming weeks.

Second detail: Astra can ask the user a question without stopping work that does not depend on the answer. A single unresolved decision blocking an entire job has been one of the most common failure modes for coding agents.

Pricing and the “Price per Task” Defense 💸

Per million tokens:

Astra costs 2.5x Sol’s promotional price and lands exactly on Fable 5.1. Against the cheap end the gap is a canyon: more than 13x Gemini 3.8 Flash’s input price.

OpenAI’s answer is Brockman’s one-liner: “The price per task is what matters.” The logic holds up in principle. A higher per-token price does not have to mean a higher bill if the model finishes in fewer steps and needs fewer retries, and OpenAI says Astra uses fewer tokens on several evaluations and in partner tests. The launch data is simply too thin to show whether those savings cover the premium.

To see what it does to your own workload, drop your input and output token counts into our LLM cost calculator and put Astra next to Claude and Gemini. The model is already in the list.

Critical Cyber and Daybreak Blue 🔒

This is the heaviest part of the release. OpenAI says Astra has crossed the Critical cybersecurity threshold in its Preparedness Framework, meaning it can discover previously unknown security flaws and develop new exploits against well-protected systems without step-by-step human guidance.

That is not a hypothetical. In company tests the model developed exploits for hardened browsers and operating systems, and while OpenAI was evaluating it against recent V8 bugs it found two previously unknown vulnerabilities, which the company says it is disclosing to the maintainers.

Access is therefore tiered:

• The standard-access Astra refuses parts of advanced cybersecurity work, including exploit discovery.
• Less restricted access goes to vetted defenders through Daybreak, expanding via Daybreak Blue in the coming weeks. Daybreak Blue is an access program for authorized defensive work, not a separate model or reasoning mode.
• OpenAI states that its published cyber results reflect Daybreak Blue access, not the default production configuration.
• In the API, a cybersecurity safety check will stop a task outright rather than pause it and wait for approval.

VP of research Mia Glaese was blunt about the side effects: users outside the trusted-access programs may hit slowdowns, pauses or blocks during cybersecurity work, and sometimes during unrelated work. “At launch, this is something that people should expect,” she said.

The split itself is not new. GPT-5.6 Cyber and Google’s Gemini 3.8 Flash Cyber follow the same pattern: the capability that writes the patch also writes the exploit, so it ships through a separate door.

More Capable, Harder to Monitor

OpenAI’s “most aligned model” claim rests partly on one internal test: in impossible-task scenarios, Sol went outside its authorized target 48.2% of the time, while Astra did so 0%. The company describes the older model as running “without production safeguards”, but does not make the role of the surrounding safety setup clear enough for a direct comparison.

The system card numbers are firmer:

• Indirect prompt injection robustness rose from 96.23% to 99.79%, and instruction hierarchy attacks to 99.99%.
• Across 54,000+ internal coding tasks, Astra drew roughly half as many flags for higher-severity misaligned behavior as Sol.

The bad news sits in the same document: Astra’s written reasoning is harder to monitor than Sol’s, specifically in evaluations designed to elicit monitoring evasion. OpenAI attributes part of the decline to Astra having more control over its written reasoning and solving simpler problems in fewer written steps.

Chief scientist Jakub Pachocki summarized it: “Progress in intelligence does not guarantee progress in alignment.” He added that OpenAI “will withhold scaling until we can regain enough confidence” in its ability to monitor future models.

What It Did in Mathematics

The most striking non-benchmark claim is mathematical. OpenAI says Astra was involved in two new findings about prime number gaps. Mathematician Julia Stadlmann had already pushed one bound from 246 to 240; with Astra involved it fell to 186. The company points to a second case where part of a bound that had not moved in more than 80 years was improved.

There is a real gap in the account. OpenAI does not spell out what Astra produced on its own, what the researchers suggested, or how the work passed between them. It is more than solving a benchmark with a known answer, but it is not enough to call it evidence of AGI.

Where You Can Use It

As of September 3, 2026:

• Today: enterprise customers who already have access through OpenAI’s Daybreak program.
• In the coming days: ChatGPT Plus, Pro, Business and Enterprise, the OpenAI API, and AWS.
• Astra Pro: included for Pro, Business and Enterprise plans.
• Zero Data Retention: available to eligible API customers.
• Billing: ChatGPT usage counts against existing subscription allowances, with extra credits available for more.

Wikipedia records September 3 as a limited preview, with the wider release expected to complete around September 9, 2026. There are no open weights, so self-hosting is not an option.

Which Model for Which Job?

• Desktop and browser automation: Astra’s clearest advantage. Cutting average task time from 75 to 40 minutes is worth more than the score difference.
• Scientific command-line work and CAD: a 40-point gap on Terminal-Bench Science and 12 points on BenchCAD are not the kind of gaps you close with budget.
• Everyday coding: no rush. The DeepSWE gap is 3-4 points with overlapping uncertainty ranges, and Opus 5 is half the price on output.
• High-volume production workloads: not Astra’s lane. For classification, tagging and summarization, Gemini 3.8 Flash or GPT-5.6 Luna remain the right answer.
• Cybersecurity: picking a model is pointless if standard access will refuse the task. Check your Daybreak Blue eligibility first.

What Comes Next

Astra’s lasting legacy probably will not be the benchmark table. Two other things look more durable: models being distributed by access tier, and OpenAI stating in its own words that monitorability went down.

On price, the market has split cleanly. At the top, Astra and Fable 5.1 meet at the same $10/$50 tariff. At the bottom, the Flash tier and open-weight models operate at a tenth of that. What closes the gap between them is not model quality, it is how many steps your job takes.

Frequently Asked Questions

Q: When did GPT-6 Astra come out? A: OpenAI announced GPT-6 Astra on September 3, 2026. Access started the same day with enterprise customers in the Daybreak program; the Plus, Pro, Business, Enterprise, API and AWS rollout is expected to complete in the following days.

Q: Is there a model called GPT-5.6 Astra? A: No. The model is GPT-6 Astra, the successor to GPT-5.6 rather than a variant of it. GPT-5.6 shipped as Sol, Terra and Luna; GPT-6 currently ships as Astra and Astra Pro only.

Q: How much does GPT-6 Astra cost? A: $10 per million input tokens and $50 per million output tokens in the API. That is 2.5x GPT-5.6 Sol’s promotional price and identical to Claude Fable 5.1.

Q: Is GPT-6 Astra free? A: No. It is offered on ChatGPT Plus, Pro, Business and Enterprise plans, with usage counted against existing subscription allowances and extra credits available. It is not on the free tier.

Q: Did GPT-6 Astra really score 98.6% on ARC-AGI-3? A: The number is real but harness-dependent. With the provider adapter harness it scores 98.6% at max effort and 99.9% at high effort. With the standard harness the same model scores 62.7%. The other models in the comparison ran under different setups, so placing the scores side by side can mislead.

Q: What is the difference between Astra and Astra Pro? A: OpenAI has announced two variants so far. Astra is the broadly rolled-out model; Astra Pro goes to Pro, Business and Enterprise plans. No separate price tariff was published at launch.

Q: Can GPT-6 Astra be used for cybersecurity work? A: The standard-access version refuses advanced work such as exploit discovery. Less restricted access goes to vetted defenders through the Daybreak and Daybreak Blue programs. In the API, a cybersecurity safety check stops the task outright rather than pausing for approval.

Q: What is GPT-6 Astra’s context window? A: OpenAI did not publish the context window or the knowledge cutoff at launch. For reference, GPT-5.6 Sol shipped with roughly 1.05M tokens of context.

Take care… 🙂