GPT-6 Astra Is “Here.” So Who Can Actually Use It?

Last checked: September 3, 2026 at 4:42 p.m. Pacific (23:42 UTC). Astra is in a rolling launch window, so this access matrix separates a documented release from access that has reached a customer account.

Update: Microsoft now confirms that GPT-6 Astra has begun rolling out through the Microsoft Foundry Limited Access Program. Microsoft also published the supported deployment types and Azure-specific token prices; the access matrix and pricing section below have been updated.

GPT-6 Astra is officially here. OpenAI now has a GPT-6 Astra system card, a developer model page, an API changelog entry, and ChatGPT release notes. That still does not mean Astra has reached every eligible account.

The official rollout language is narrower than “available to everyone.” OpenAI’s ChatGPT release notes say access is rolling out to a limited set of organizations, Astra is not generally available, and broader availability is planned over the coming days. The developer page says API access and access through Plus, Pro, Business, and Enterprise are coming in the coming days.

That gap matters because there are at least three materially different Astra experiences: the model announced to the press, the safeguarded production model intended for ordinary ChatGPT and API use, and the Daybreak Blue configuration behind Astra’s headline cybersecurity results. Those should not be treated as interchangeable.

The live GPT-6 Astra access matrix

Evidence labels: “Confirmed” means an official provider page supports the claim. “Reported” means credible launch coverage says it is happening, but the provider’s product documentation has not caught up. “Still unknown” means neither public documentation nor a verified account-level result settles the question.

The three Astra experiences

1. The model announced to the press

This is the GPT-6 Astra in launch headlines: the new flagship model, its benchmark claims, and the promise of access across paid ChatGPT plans, the OpenAI API, AWS Bedrock, and Microsoft Azure. OpenAI now officially uses the GPT-6 Astra name and calls it its most capable model, but the distribution story still describes a rollout rather than a universal entitlement.

Astra’s official product documentation is now live, while general availability is not. The practical reading is that announcement day, documentation, account entitlement, and a successful production request can occur at different times.

2. The safeguarded production Astra

This is the version ordinary ChatGPT, Codex, and API customers should expect. OpenAI says its misalignment monitoring reviews an agent’s reasoning, actions, inputs, and outputs across ChatGPT, Codex, and supported Responses API use. A flagged conversation can be paused or ended. Some product surfaces let the user review and resume; conversations stopped through the API may not be resumable.

That difference is operational, not cosmetic. A long autonomous workflow needs to handle a stopped API response without silently retrying the same action forever. OpenAI also says this misalignment monitor does not apply to Chat Completions, because that interface cannot combine reasoning and tool use; separate model refusals and misuse safeguards still apply.

3. Astra evaluated with Daybreak Blue

OpenAI’s most important disclosure is easy to lose in the benchmark coverage: the published Astra cyber results reflect Daybreak Blue access, not the default production configuration. Daybreak Blue reduces some cyber-safety refusals for verified defenders working on authorized tasks.

The new system card quantifies the gap. On Astra, Daybreak Blue raises proof-of-concept exploit completion from 2.4% to 92% and cyber red-teaming completion from 7.4% to 76.9%. Even with Blue, Astra fully completes only 3.5% of the arbitrary cyber requests in OpenAI’s Advanced Cybersecurity Completion Rate evaluation. Those are safety-evaluation completion rates, not a general benchmark of model intelligence.

That does not mean every Daybreak Blue customer can switch on the benchmark configuration. OpenAI’s current FAQ says reduced refusals are not available on Astra for most Daybreak customers. The same page maps the API alias gpt-daybreak-blue-latest to gpt-5.6-sol, while AWS maps Daybreak Blue to openai.gpt-daybreak-blue-5.6-sol. Neither public mapping points to Astra.

So a Daybreak badge, a Daybreak alias, and Daybreak-configured Astra are three different things. If you need the last one, ask OpenAI to confirm the underlying model, access level, and safeguard profile in writing.

API ID, pricing, limits, and data controls

At the official Standard rates, a request that consumes 100,000 uncached input tokens and 20,000 output tokens would cost about $2: $1 for input and $1 for output. Fast would make the same token mix about $4, before tool or service-specific charges. The long-context surcharge begins above 272,000 input tokens.

Fast changes routing for speed; OpenAI does not describe it as a higher-intelligence configuration. Kingy’s Fast mode pricing guide explains the speed-versus-cost tradeoff on the previous generation.

Microsoft publishes separate Foundry pricing. Standard Global short-context deployments match OpenAI’s $10 input and $50 output rates, with $1 cached input and $12.50 cache writes per million tokens. Long-context Global is $20 input, $2 cached input, $25 cache writes, and $75 output. U.S. Data Zone rates are $11/$1.10/$13.75/$55 for short context and $22/$2.20/$27.50/$82.50 for long context. Those are Azure deployment prices, not OpenAI’s Fast tier.

How to check whether you can use Astra

In ChatGPT

1. Open the model picker in the specific workspace you intend to use.
2. Look for Astra by name. Do not assume that Plus, Pro, Business, or Enterprise access has reached every account at once.
3. If you are in a managed workspace, ask an admin to check model controls and regional restrictions.
4. For cyber work, note whether the workspace is in Daybreak and whether Astra itself has the requested access level. A general Daybreak approval does not settle that question.

In Codex

1. Check the model picker in the current Codex app or environment.
2. If Astra is absent, do not substitute gpt-daybreak-blue-latest and call the result Astra. OpenAI’s current mapping points that alias to GPT-5.6 Sol.
3. Plan for monitored cyber tasks to pause for review. That is a stated product behavior, not proof that your account already has Astra.

If Astra has not reached your account, use Kingy’s OpenAI model selection guide to choose the best documented model available now.

In the OpenAI API

1. Confirm that gpt-6-astra appears in the model list available to your project.
2. Use the Responses API for tool calling. Although the model also supports Chat Completions, OpenAI’s changelog says Astra tool calling requires Responses.
3. Make a minimal request, then record the UTC time, project, endpoint, returned model value, and request ID.
4. Test the failure path. A monitored task can stop rather than opening an interactive review step.
5. Check your organization’s limits page before moving production traffic.

Use OpenAI’s official API quickstart and API changelog for the request shape and migration constraints. A copied launch-day code sample cannot grant project entitlement.

For Daybreak Blue

Daybreak access requires an application, identity and trust verification, and approval. Enterprise onboarding also requires an administrator to enable Daybreak for the relevant project and provision fresh credentials. OpenAI’s system card adds that individual members must enable Advanced Account Security to use Daybreak Blue; members who do not will receive standard access. Daybreak does not itself grant Zero Data Retention. The two controls have separate approval paths.

The official enterprise Daybreak onboarding guide is the right checklist. The important Astra-specific question for your account team is: “Does this project receive Daybreak Blue on Astra, or the current GPT-5.6 Sol mapping?”

What is confirmed, and what is still moving

OpenAI confirms that GPT-6 Astra is released, that it crossed the Critical cyber-capability threshold, and that the headline cyber evaluation used Daybreak Blue rather than the default production setup. It also confirms that external monitoring covers ChatGPT, Codex, and supported Responses API use, while a stopped API conversation may not be resumable.

OpenAI now supplies the near-term distribution and commercial details: rollout to Plus, Pro, Business, Enterprise, and API customers over several days; the gpt-6-astra model ID; $10 input and $50 output per million tokens for Standard; and twice those rates for Fast. Microsoft now separately confirms a limited Foundry rollout and Azure-specific pricing. Third-party reporting still supplies the unconfirmed AWS Bedrock timing.

Official product pages still need to settle the remaining access questions: plan-by-plan dates, individual account eligibility, Codex rollout timing, the AWS model ID and regions, Azure’s exact model ID and regional availability, and the first independently verified successful requests outside the limited early-access group.

For anyone tracking the launch from rumor to documentation, Kingy’s earlier Astra evidence tracker remains the paper trail. The model should appear in your product or project, the provider should document the exact configuration, and a request should succeed under the limits you intend to run. Treat your project as unprovisioned until all three are true.

The Kingy Brief

Follow The Kingy Brief.

One consequential launch, one pricing, limit, or shutdown change, one hands-on test, one exact prompt or Test Pack, and one try / watch / skip verdict.

Free · Choose your subjects · Double opt-in · Unsubscribe anytime

Load subscription form

Form not showing? Subscribe on Beehiiv.