Open AI GPT-6 Astra Release Date

OpenAI launched GPT-6 Astra on September 3, 2026. That is the release date, and it ends months of speculation while confirming that Astra and GPT-6 are the same model. The rollout started that Thursday with enterprise customers in OpenAI's Daybreak cybersecurity program, and access through the OpenAI API, Amazon Web Services, and the ChatGPT Plus, Pro, Business, and Enterprise plans is arriving in the days after. So the release date question now has a real answer, not a prediction-market estimate.

That phasing is the whole story of this launch. A named model, a research post, and a live enterprise preview all landed at once, but general availability did not. OpenAI said the paid ChatGPT tiers and the API would follow "in the coming days," so if you are on a standard plan today, you may still be waiting a short while for access to switch on.

OpenAI also confirmed the naming question directly. The model ships as GPT-6 Astra, so the codename and the version number that people spent months debating turned out to describe one model. The earlier caution in this space, that Astra might arrive as GPT-5.7 or under another label, no longer applies. What follows is the pricing, the benchmark results, and the reasons access is gated at launch.

What GPT-6 Astra costs

GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens on the standard OpenAI API. That is 2.5 times the current promotional rate of GPT-5.6 Sol, the previous flagship, which runs $4 input and $20 output per million tokens, and it lands GPT-6 Astra at the premium end of the current frontier market. Separate rates apply to cache reads and writes, and prompts above 272K input tokens are billed at higher multiples.

A Fast mode is available in the API at twice the standard speed for twice the standard price. For developers, the model is callable as gpt-6-astra, and it is also available through Amazon Bedrock and Microsoft Foundry.

Pricing as of September 2026. Source: OpenAI.

A higher per-token price does not automatically mean a higher bill. OpenAI argues that Astra finishes many tasks in fewer tokens and fewer retries, which can offset the rate on a per-task basis. That claim is plausible, but the launch data is too thin to prove it. As OpenAI's Greg Brockman put it, the price per task is what matters, and OpenAI did not publish enough task-level detail to calculate it.

There is no Luna, Terra, or Sol split for this generation, at least not yet. The lineup is GPT-6 Astra and a stronger GPT-6 Astra Pro, the latter available to Pro, Business, and Enterprise users.

The benchmarks are strong, with caveats worth reading

GPT-6 Astra posts state-of-the-art scores across math, science, computer use, and cybersecurity, but several of the headline numbers need context before you treat them as settled. OpenAI ran its evaluations at maximum effort unless noted, a setting that can lift results while raising latency and token use. All figures below are OpenAI-reported vendor benchmarks, not independently verified, and OpenAI notes the scores reflect the maximum at any effort level.

The standout is ARC-AGI-3, where OpenAI reports a near-perfect score. That result came from a Responses API harness that keeps reasoning between turns, and OpenAI has previously shown those system choices can raise ARC-AGI-3 scores substantially without changing the model. The benchmark here measures Astra plus OpenAI's agent system together, not the model alone.

On math, Astra reports 97.6% on FrontierMath Tier 4. Epoch AI, which runs that benchmark, has disclosed that OpenAI funded its development and holds exclusive access to part of it, so it belongs in the "impressive but not neutral" column. The one cross-vendor figure on OpenAI's own charts is the Artificial Analysis Intelligence Index v4.1.1, where Astra scores 61.2, ahead of GPT-5.6 Sol at 60.9 but behind Claude Fable 5.1 at 65.7.

Selected OpenAI-reported benchmark results. All scores are OpenAI-reported at maximum effort unless noted.

Coding is where the "world's most intelligent model" framing gets softer. Astra's DeepSWE v1.1 score of 74.1% is a real gain over Sol, but Meta reported a slightly higher number for Muse Spark 1.3 at its maximum setting, and the public DeepSWE leaderboard clusters Gemini 3.8 Flash and Claude Opus 5 around the same mark. On a test that size, the gaps are one or two tasks wide, and the uncertainty ranges overlap. Astra leads the pack it was benchmarked against, but it does not clearly lead the broader field on coding.

Why access is gated at launch

GPT-6 Astra is the first OpenAI model to cross the "Critical" cybersecurity threshold in the company's Preparedness Framework, and that classification is the reason the rollout is careful rather than wide. In OpenAI's own tests, the model developed working exploits for hardened browsers and operating systems, and it surfaced two previously unknown vulnerabilities during evaluation, which OpenAI says it is disclosing to the affected maintainers.

The practical effect is real limits on what standard users get. The publicly available version refuses advanced offensive security work such as exploit discovery, and OpenAI is routing less-restricted capabilities to vetted defenders through its Daybreak Blue program. For API users, a cybersecurity safety check can stop a task outright rather than pause for approval, and OpenAI has warned that people outside its trusted-access programs may see slowdowns or blocks, sometimes even on unrelated work.

This is also why the release date and "availability for you" are not the same date. A model flagged Critical for cyber risk follows a gated path by design, closer to a policy rollout than a normal product drop.

The rumors this settles

Most of the specific GPT-6 Astra rumors that circulated before launch can now be marked resolved, and several were wrong. Astra is confirmed as GPT-6, so the naming debate is closed. The release did happen in early September, which the prediction markets had drifted toward after repeatedly mispricing earlier windows. The claim that GPT-6 launched back in April as a model called Spud was always false, since that release shipped as GPT-5.5.

The leaked hardware specs remain unconfirmed. The widely shared 10 trillion parameter count and 1.5 million token context window traced to leak clusters rather than OpenAI, and the official launch materials do not publish those figures. OpenAI did confirm one scale detail through its own team: Astra was its largest training run to date, the first pre-trained on more than 100,000 GPUs at its Texas site.

Stay on top of the latest AI model news

GPT-6 Astra is one launch in a release cycle that now moves week to week, and the release date was a moving target until the day it shipped. The next flagship, from OpenAI or a rival, will follow the same arc: rumors, leaked specs, prediction-market odds, then a launch that confirms or corrects all of it.

For the running story, Emergent News tracks it as it happens, from early rumors and release-date speculation on upcoming models to the confirmed details when a new LLM actually ships. It is the fastest way to know what is real, what is still unconfirmed, and what just launched.