GPT-6 Astra Launches: $10/M Tokens, How to Try It [2026]

Last updated 22 September 2026.

OpenAI pushed GPT-6 Astra into the world on September 3, 2026, and by the company’s own account it is the most capable model it has ever shipped. The catch: almost nobody who wanted it could actually use it on day one. Within hours, paying ChatGPT subscribers were locked out while OpenAI routed early access to a narrow group of enterprise security partners, and CEO Sam Altman found himself publicly apologizing for what The Verge called a “messy” rollout. The next day, September 4, access opened wider, and the real story of GPT-6 Astra split into two threads: a genuinely capable frontier model, and a launch that stumbled on its own hype.

This piece breaks down exactly how GPT-6 Astra launched, how to try it right now across ChatGPT and the API, what it costs, why OpenAI throttled access before widening it, and what the model’s benchmark and cybersecurity numbers actually mean for developers, enterprises, and the broader AI market. Two and a half weeks on, that split verdict still holds: Astra is now fully rolled out across ChatGPT and the API, and its arrival is already reshaping how prediction markets and rival labs talk about the pace of frontier AI, a shift explored in more detail in our look at how betting markets have moved on AGI timelines since Astra shipped.

·

Google · Preferred Sources

Don't miss new tech stories on Google

Add Tech Insider once in the Google app and our stories appear in your news suggestions.

Add Now

What Is GPT-6 Astra?

Core Capabilities

GPT-6 Astra is OpenAI’s newest flagship large language model, positioned as a step up in coding, research, computer-use, and multi-step agentic task performance over prior generations. Under the hood, the model shipped with a 1,050,000-token context window, split into roughly 922,000 input tokens and 128,000 output tokens, and a knowledge cutoff of April 30, 2026, according to OpenAI’s model card as reported by ComputingForGeeks. In its own launch materials, OpenAI described the rollout plan directly: “GPT‑6 Astra 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, Microsoft Azure, and AWS Bedrock,” according to OpenAI’s official announcement.

Why It Triggered OpenAI’s Strictest Safety Review

NBC News reported that OpenAI is marketing Astra as its most intelligent AI system to date, and that it is the first OpenAI model to trigger the company’s most advanced internal safety protections because of its cybersecurity capabilities, according to NBC News. Reuters separately reported that OpenAI is pitching Astra as faster and more broadly capable than its predecessors, citing practical use cases such as tax preparation, game development, architectural rendering, legal memo formatting, and apartment hunting.

The Official Launch Timeline: September 3-4, 2026

September 3: A Limited, Security-Vetted Start

GPT-6 Astra’s release did not happen as a single, universal switch-on. It happened in stages, and the stages themselves became part of the news cycle. On September 3, 2026, OpenAI opened access first to a limited set of organizations, particularly Daybreak-program enterprise customers, according to NBC News, OpenAI’s own safety materials, and Techtimes’s rollout timeline. Wikipedia’s entry on the model likewise lists Astra as first released to approved users on September 3, 2026, with a stable public release following on September 4, 2026, when broader access to ChatGPT Plus, Pro, Business, and Enterprise accounts, plus the API and cloud partners, opened as promised “over the coming days.”

September 4: Access Widens After Backlash

That promise ran into trouble almost immediately. Paying subscribers who expected instant access based on OpenAI’s own framing of Astra as a generational leap found themselves without it, and the backlash was fast enough that Altman addressed it publicly within hours, according to The Verge’s coverage of the launch. By September 4, The Verge reported that Astra was “now actually available” to the broader base of Pro, Enterprise, and other paid accounts, effectively closing the gap between the marketing and the access.

Why the One-Day Reversal Stands Out

The gap between the two announcements is notable from a corporate-communications standpoint. OpenAI did not spend weeks quietly widening access after the initial restriction, it moved from a security-vetted limited group to broad Plus, Pro, Business, and Enterprise availability in roughly a single day, following public pressure and Altman’s apology. That quick a reversal supports the reading that the initial restriction was a deliberate security decision tied to Astra’s cybersecurity classification, not a server-capacity limitation, since capacity constraints rarely resolve in a single day.

How to Try GPT-6 Astra Right Now

For most readers, the practical question is simpler than the rollout drama: how do you actually get access to GPT-6 Astra today. According to OpenAI’s official announcement, access now runs through several channels depending on what kind of account you have.

ChatGPT Plus, Pro, Business, and Enterprise

• ChatGPT Plus and Pro: Astra now appears in the model picker on web, iOS, and Android for eligible accounts, with Pro subscribers also getting access to the higher-tier Astra Pro variant, per OpenAI’s launch post.
• ChatGPT Business and Enterprise: Access is available, but for Enterprise workspaces it is off by default. “Enterprise administrators can enable Astra for their workspace; access is off by default at launch,” OpenAI said in its announcement.

OpenAI also clarified how usage is billed inside ChatGPT itself: “Astra usage is included within the existing subscription allowances—users and businesses will also be able to purchase credits for additional usage,” the company said in its launch materials. In other words, existing Plus, Pro, Business, and Enterprise subscribers do not need a separate plan to try GPT-6 Astra; it is bundled into what they already pay for, up to a usage cap, after which credits become available.

OpenAI API and Cloud Platforms

• OpenAI API: Developers can call the model directly using the identifier gpt-6-astra.
• Cloud partners: OpenAI says Astra is also available through Microsoft Azure and AWS Bedrock for organizations that run their AI workloads through those platforms.

The simultaneous Azure and AWS Bedrock availability is worth noting because OpenAI has not always shipped flagship models to both major cloud partners on the same day as the ChatGPT and API release. For enterprises already running regulated workloads through Azure or Bedrock, adopting Astra doesn’t require a new vendor relationship or a fresh compliance approval, which typically shortens internal rollout timelines.

API Access and Pricing Breakdown

Standard Token Pricing

Developers building on top of GPT-6 Astra will care most about the token economics, since that determines whether the model is viable for production workloads versus experimentation. As of September 2026, OpenAI’s published standard API pricing, confirmed in the OpenAI System Card and reported by Techtimes, sets GPT-6 Astra at $10 per million input tokens and $50 per million output tokens for standard requests up to 272,000 input tokens, which ComputingForGeeks calculates works out to roughly 2.5 times the rate OpenAI charged for GPT-5.6 Sol. That output price in particular is a meaningful jump from the cheaper mid-tier models OpenAI has shipped over the past year, and it signals that Astra is being positioned as a premium, capability-first model rather than a cost-optimized one.

What the Pricing Looks Like at Scale

As an illustration of what that published rate means in practice: a request that processes 100,000 input tokens, roughly the length of a 75,000-word document, costs about $1 before any output is generated. Generating a 2,000-word response, roughly 2,600 tokens, adds about $0.13 at the $50-per-million-output-token rate. A team running 1,000 similar requests a day would spend in the neighborhood of $1,130 a day, or roughly $34,000 a month, on token costs alone, before accounting for retries, longer context windows, or agentic workflows that call the model multiple times per task. That kind of math is why the guidance later in this article points developers toward testing Astra on a narrow, high-value workload first rather than swapping it in wholesale for a cheaper model already in production.

Cached Pricing and the Full Access Table

OpenAI’s pricing documentation has since confirmed those secondary figures: as of late September 2026 the published rates include a cached-input rate of $1 per million tokens and a cache-write rate of $12.50 per million tokens. The same documentation clarified a long-context tier, under which requests exceeding 272,000 input tokens are billed at $20 per million input tokens and $75 per million output tokens, with the entire request charged at that higher rate. Techtimes and Vellum also reported in September 2026 that GPT-6 Astra offers a Fast mode, which runs at roughly 2 to 2.5 times the standard model’s speed for 2 times the standard token price, giving developers a latency-versus-cost lever separate from the base pricing tier. Pulling in the other direction, September 2026 developer breakdowns from CodingFleet and Digitalapplied note that Astra’s Batch and Flex processing modes run at roughly half the standard token rate, giving cost-sensitive, latency-tolerant workloads a cheaper lane than the standard $10/$50 tier.

Access Channel	Who Gets It	Cost Structure	Availability Status
ChatGPT Plus	Individual subscribers	Included in subscription, credits for extra use	Live
ChatGPT Pro	Individual subscribers	Included in subscription; adds Astra Pro variant	Live
ChatGPT Business	Team accounts	Included in subscription, credits for extra use	Live
ChatGPT Enterprise	Enterprise workspaces	Included in subscription	Off by default, admin must enable
OpenAI API (gpt-6-astra)	Developers	$10/M input tokens, $50/M output tokens	Live
Microsoft Azure	Azure AI customers	Per cloud-partner agreement	Live
AWS Bedrock	AWS customers	Per cloud-partner agreement	Live

Calling GPT-6 Astra From the API

Sample Request and Testing Advice

For developers who want to test the model programmatically, the model identifier confirmed in OpenAI’s launch documentation is straightforward to drop into an existing integration. A minimal request looks like this:

curl https://api.openai.com/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-astra",
    "messages": [
      {"role": "user", "content": "Summarize this quarter's launch timeline."}
    ]
  }'

Because Astra is billed at a higher per-token rate than OpenAI’s cheaper models, teams evaluating it are best served by testing on a narrow, high-value workload first, such as a multi-step coding or research task, rather than swapping it in as a blanket replacement for a cost-sensitive model already in production. A full step-by-step walkthrough of setting up GPT-6 Astra API access covers authentication and rate-limit handling in more depth.

The “Messy” Rollout and Sam Altman’s Apology

What Went Wrong, and How Fast It Was Fixed

The most talked-about part of the GPT-6 Astra launch was not a benchmark score, it was the gap between what OpenAI promised and what paying customers actually got on day one. The Verge’s report on the aftermath, headlined around Altman’s apology, described a rollout where subscribers on Plus, Pro, and Business tiers were left without access despite OpenAI framing Astra as a generational leap. Altman addressed the frustration publicly within hours of launch, and OpenAI moved to widen access within roughly a day, according to The Verge’s follow-up reporting confirming Astra was “now actually available” more broadly.

The sequence is worth spelling out because it is unusual for a flagship model launch: announce broad availability, gate it tightly to a security-vetted group, face public frustration from paying customers, apologize, then expand access within about 24 hours. That is a materially different pattern from prior OpenAI flagship launches, which typically opened to paid tiers immediately even when full rollout took longer to reach every account.

Why OpenAI Gated Access First: The Daybreak Program

The Daybreak Program’s Role in the Staged Rollout

The initial restriction was not arbitrary. According to NBC News, GPT-6 Astra was first made available to participants in OpenAI’s Daybreak program, a track built for cybersecurity defenders, before wider enterprise and consumer access followed. That sequencing lines up with what OpenAI has said about Astra’s underlying capability: this is the first OpenAI model that required extra scrutiny before general release specifically because of what it can do in offensive cybersecurity contexts, not because of capacity constraints on serving traffic.

OpenAI framed the sequencing choice as deliberate. “For developers, GPT‑6 Astra will be available in the OpenAI API as `gpt-6-astra` and through Microsoft Azure and Amazon Bedrock,” the company said in its announcement, confirming that the broader technical rollout was always planned to follow the security-vetted group rather than launch simultaneously with it.

Benchmark Scores: How Astra Performs

The Reported Scores

Coverage of the launch has cited a cluster of benchmark results tied to GPT-6 Astra, spanning math reasoning, software engineering, computer-aided design understanding, general science reasoning, and offensive security testing. These are the figures that have circulated in launch coverage:

Benchmark	Reported Score	What It Measures
FrontierMath Tier 4 v2	97.6%	Advanced mathematical reasoning
DeepSWE v1.1	74.1%	Real-world software engineering tasks
BenchCAD	95.9%	Computer-aided design comprehension
GPQA Diamond	96%	Graduate-level science reasoning
ExploitBench	100%	Exploiting known system vulnerabilities

Reading the Benchmark Spread

Looked at together, the four non-security benchmarks split into two tiers. FrontierMath Tier 4 v2 (97.6%), BenchCAD (95.9%), and GPQA Diamond (96%) all sit in the mid-to-high 90s, suggesting Astra is approaching a ceiling on structured reasoning tasks that have a single verifiable answer. DeepSWE v1.1, at 74.1%, is the outlier in the other direction: real-world software engineering tasks involve ambiguous requirements, multi-file context, and judgment calls that don’t reduce to a single correct answer the way a math proof or a CAD specification does. That roughly 20-to-24-point gap between the reasoning-heavy benchmarks and the software-engineering benchmark is a useful reality check for teams evaluating Astra for coding work: near-perfect scores on math and science tests don’t automatically translate into equivalent gains on messy, real-world engineering tickets.

The ExploitBench Outlier

The ExploitBench score is the one that has drawn the most scrutiny. TechCrunch reported that beyond the perfect score on known vulnerabilities, Astra discovered and exploited two previously unknown, zero-day vulnerabilities in a modified version of the test built by OpenAI’s own engineers, without a human directing each step. That distinction, finding a flaw nobody knew existed versus exploiting one that’s already documented, is what separates a strong coding benchmark result from a genuine cybersecurity capability jump.

GPT-6 Astra’s Cybersecurity Capability Hits “Critical” Level

Why the Critical Classification Matters

OpenAI classifies model risk using its Preparedness Framework, a tiered system the company uses to decide what safeguards a model needs before release. GPT-6 Astra is the first OpenAI model to reach the framework’s Critical level specifically for cybersecurity capability. NBC News reported that OpenAI described Astra as able to find previously unknown security flaws and develop ways to exploit them across well-protected systems without a person guiding each step, which is the underlying reason the model triggered the company’s most advanced internal safety protections before release.

That single classification explains most of what makes this launch different from a typical model update. A model that scores well on coding or math benchmarks is a product story. A model that autonomously finds zero-day vulnerabilities is a security story, and it is the reason OpenAI chose to hand early access to cyber-defense partners in the Daybreak program before opening the doors to the general subscriber base.

New Safety Measures Around Astra

The New Controls

Crossing the Critical cybersecurity threshold came with a specific set of new controls, according to OpenAI’s safety materials and NBC News’s reporting on them. These include universal monitoring of full agentic trajectories, including chain-of-thought tracking, stricter isolation for testing and deployment environments, checkpoint encryption for intermediate model states, and a blocking alignment evaluation that Astra must pass before OpenAI’s own staff can use it internally. NBC News also reported that OpenAI said the model would carry increased monitoring mechanisms designed to “rapidly detect and contain potentially misaligned actions.”

What It Means for Enterprise Rollouts

For enterprise customers, the practical upshot is that Astra ships with more guardrails baked in by default than prior flagship releases, and admins evaluating it for internal use should expect additional configuration steps tied to those monitoring and isolation requirements, particularly for any agentic or autonomous-action use cases.

These controls carry more direct implications for regulated industries such as finance, healthcare, and defense contracting, where procurement teams typically require documentation of any new monitoring layer before a model touches production systems. Because chain-of-thought tracking and checkpoint encryption are new requirements rather than optional add-ons, security reviews of Astra are likely to take longer than reviews of prior OpenAI models, even at organizations that already approved GPT-5.6 Sol for internal use.

Competitive Landscape: Where Astra Sits

How Astra Stacks Up Against Rivals

GPT-6 Astra does not launch into a vacuum. It follows OpenAI’s own GPT-5.6 Sol as the company’s prior flagship, and it arrives while Google and Anthropic continue shipping their own frontier releases on overlapping timelines. Independent analysis of how Astra stacks up against rivals such as Claude Opus 5 and Gemini 3.8 Flash has already circulated in the days following launch, with reported gaps on specific benchmarks running into multiples rather than single-digit percentage points, though the exact competitive delta depends heavily on which benchmark and which task category is being measured.

What is clear from the available reporting is that the ExploitBench and ExploitBench-derived zero-day results put Astra in a category by itself among publicly documented frontier models: no rival lab has published a comparable claim of a model autonomously discovering and exploiting unknown vulnerabilities during pre-release testing. That is arguably a bigger competitive differentiator, for better or worse, than any single reasoning benchmark score. A closer head-to-head look at GPT-6 Astra against Claude Opus 5 and Gemini 3.8 Flash breaks down where the benchmark gaps are widest.

Market and Industry Reaction

Press Coverage and User Sentiment

Mainstream financial and tech press have framed the Astra launch as a major frontier-model event rather than a routine model bump. CNBC’s coverage focused on the phased rollout and its cybersecurity implications, while Reuters emphasized both the model’s expanded task capability and the scrutiny building around AI agent safety more broadly, and TechCrunch described the release as “powerful and controversial”. Coverage has ranged from framing this as evidence the “AGI era” has effectively arrived to far more cautious takes centered on the access restrictions and the Critical cybersecurity designation.

That split in coverage tone mirrors a split among the people actually using the model: developers and power users focused on Astra’s coding and reasoning gains are largely enthusiastic, while security researchers and policy-focused observers are more concerned with what it means that a publicly available model can now autonomously find and exploit unknown software vulnerabilities. Some coverage has gone as far as arguing the launch marks the point where the AGI era has effectively arrived, while other reporting has focused squarely on the model crossing OpenAI’s Critical cybersecurity threshold. The stakes are amplified by the sheer reach of the platform Astra now flows into: OpenAI’s products were serving more than 1 billion weekly users, a milestone first passed in mid-2026 and re-verified as of September 10, 2026 with no fresher figure disclosed, meaning any capability jump or safety guardrail baked into the new flagship propagates across an unusually large user base rather than a niche developer audience.

Historical Context: From GPT-4 to GPT-6 Astra

From GPT-4 to GPT-6 Astra in Two Years

OpenAI’s naming has moved fast over the past two years, from GPT-4 to GPT-5 to GPT-5.6 Sol and now to GPT-6 Astra, with each generation adding stronger coding and agentic task performance. What sets Astra apart from that lineage is not just the benchmark improvement, it is the first time a release has been explicitly gated pre-launch because of an internal safety classification tied to cybersecurity rather than general capability concerns. Earlier generations rolled out to paid subscribers essentially on day one; Astra’s staged, security-vetted rollout, followed by a public apology when that staging frustrated customers, is a new pattern for the company.

It is also a preview of a dynamic the industry will likely see again: as models get better at autonomous, multi-step tasks, the same capability gains that make them valuable for legitimate coding and research work also make them more capable in adversarial contexts, forcing labs to build release processes around capability thresholds rather than calendar dates. That tension is part of why an OpenAI researcher publicly raised AI-risk concerns just days after Astra shipped.

What Developers and Enterprises Should Do Now

For ChatGPT Users

Check the ChatGPT model picker on Plus, Pro, Business, or Enterprise accounts, since access may already be live even if it wasn’t at first login after launch. Because the rollout happened in stages between September 3 and September 4, some accounts gained access before others even within the same subscription tier, so a missing model picker entry on day one was not necessarily a sign of an account problem.

For Enterprise Admins

Enterprise admins specifically need to manually enable Astra in workspace settings, since it ships off by default. Admins should also budget time for the additional monitoring and isolation configuration tied to Astra’s Critical cybersecurity classification before flipping that switch for their organization, particularly if the workspace plans to use agentic or autonomous-action features.

For API Developers

Developers testing the API should budget for the $10 per million input and $50 per million output token pricing before committing production traffic, given that rate sits well above OpenAI’s cost-optimized model tiers. It’s also worth watching for updated documentation on the cached-input and cache-write rates referenced in secondary launch coverage, since those figures were not part of OpenAI’s headline pricing announcement and may shift as official documentation catches up with the launch.

For Agentic Workflow Teams

Any team planning agentic or autonomous-action deployments should expect the new monitoring and isolation requirements tied to Astra’s Critical cybersecurity classification to add setup steps that weren’t necessary with prior models. Because universal monitoring now covers full agentic trajectories rather than just final outputs, teams should also expect longer internal security review cycles for any workflow that lets Astra take multi-step actions without a human checkpoint in between.

Predictions: What Comes Next for GPT-6 Astra

Near-Term Changes at OpenAI

• Wider Enterprise defaults will loosen. Expect OpenAI to eventually flip Astra to opt-out rather than opt-in for Enterprise workspaces once the current monitoring infrastructure proves stable in production.
• Pricing will compress. Following the pattern of every prior OpenAI flagship, a cheaper, lighter Astra variant is likely within a few months as inference costs come down and competitive pressure from Google and Anthropic pricing continues.
• Rollout mechanics will get more conservative, not less. Given the public friction around this launch, future frontier releases from OpenAI are likely to either under-promise on day-one availability or stagger announcements more explicitly to avoid a repeat of the access backlash.

Longer-Term Industry Effects

• Cybersecurity scrutiny will intensify, not fade. The Critical-level classification and the zero-day discovery results are likely to draw continued attention from policymakers and security researchers well beyond the launch news cycle.
• Rival labs will face pressure to publish comparable safety disclosures. If Astra’s Preparedness Framework disclosures become a reference point for the industry, expect Google DeepMind and Anthropic to face more public pressure to detail their own models’ offensive cybersecurity testing results.

GPT-6 Astra vs. What Came Before: The Access Gap

Breaking From OpenAI’s Own Pattern

One detail that has gotten less attention than the benchmark scores is how unusual the access pattern itself was. OpenAI has shipped several flagship models over the past two years, and in nearly every prior case, paid ChatGPT tiers got access at or near the public announcement. Astra broke that pattern by holding back consumer and most enterprise access behind a security-vetted group first.

Will Staged, Security-Vetted Rollouts Become the Norm?

The answer likely hinges on whether Astra’s Critical cybersecurity classification is a one-off tied to this specific model’s zero-day discovery results, or the first instance of a threshold more frontier models will eventually cross. If future OpenAI releases keep triggering the same Preparedness Framework tier, staged access behind a vetted partner group could become the default launch pattern rather than the exception. That question will likely become clearer with the next major model release from OpenAI. For now, Astra’s September 3–4 launch and its $10/$50 per-million-token API pricing remain the reference point the rest of the industry is reacting to, from renewed speculation about what OpenAI ships next to comparisons with rival models released in Astra’s wake.

Latest GPT-6 Astra rollout updates (September 2026)

The days immediately around launch produced a cluster of confirmed developments that round out the timeline above. Here is a dated recap of what OpenAI has confirmed since GPT-6 Astra first shipped:

• September 3, 2026: OpenAI releases GPT-6 Astra and begins a phased rollout to select organizations, describing it as its most capable model and the first to reach a “Critical” level of cybersecurity capability under its preparedness framework; initial access went to approved, mainly cybersecurity-focused enterprise customers, including Daybreak program participants, according to OpenAI’s safety overview and system card and coverage from Fortune, CNBC, 9to5Google, Betanews, and Ahram.
• September 4, 2026: GPT-6 Astra moves into stable public release, becoming generally available to ChatGPT Plus, Pro, Business, and Enterprise subscribers as well as to developers via the OpenAI API and cloud partners including AWS Bedrock and Microsoft Azure, while free-tier and lowest-cost plans remain excluded, per Wikipedia’s entry on the model, OpenAI’s launch post, and reporting from TechTimes, Betanews, CNBC, and Fortune.
• September 3, 2026: OpenAI sets standard API pricing at $10 per million input tokens and $50 per million output tokens, with documentation and third-party cost analyses also pointing to roughly $1 per million for cached input and higher rates for long-context and batch requests, positioning Astra as a notably more expensive flagship than GPT-5.6 Sol, per OpenAI’s pricing documentation as summarized by CloudZero, Codersera, ComputingForGeeks, and benchmark/review blogs.
• September 3, 2026: The dedicated API endpoint gpt-6-astra goes live for developers, reachable through OpenAI’s own API as well as cloud partners Microsoft Azure and AWS Bedrock, per OpenAI’s GPT-6 Astra launch page and technical reviews from ComputingForGeeks and CodingFleet.
• September 3, 2026: OpenAI publishes a safety overview and system card detailing Astra’s “Critical” cybersecurity classification, outlining its advanced offensive and defensive cyber capabilities, the staged rollout to vetted cybersecurity customers, and the extra safeguards and prompt-rejection behavior applied for general users, according to OpenAI’s Deployment Safety Hub.

Together, these confirmations line up with the timeline and pricing already detailed above, and they’ve fed into a wider round of follow-up coverage on this site, including a look at how GPT-6 Astra compares with Gemini 3.8 and DeepSeek V4.1 and how OpenAI has started building industry-specific ChatGPT products on top of Astra.

Frequently Asked Questions

Is GPT-6 Astra available to try right now?
Yes. As of the broader rollout that followed the initial launch, GPT-6 Astra is available in the ChatGPT model picker for Plus, Pro, and Business accounts, and can be enabled by admins on Enterprise workspaces. It is also available through the OpenAI API as gpt-6-astra, and through Microsoft Azure and AWS Bedrock.

Why wasn’t GPT-6 Astra available to everyone on launch day?
OpenAI initially limited access to participants in its Daybreak cybersecurity-defender program because Astra reached the Critical level of cybersecurity capability under OpenAI’s Preparedness Framework, according to NBC News. Broader access followed within roughly a day after paying subscribers voiced frustration and Sam Altman publicly apologized for the rollout, as reported by The Verge.

How much does GPT-6 Astra cost through the API?
OpenAI’s standard published pricing is $10 per million input tokens and $50 per million output tokens for the gpt-6-astra model. Secondary coverage has also referenced lower cached-input pricing, though that has not been confirmed as part of OpenAI’s primary pricing announcement.

Do I need a new subscription to use GPT-6 Astra in ChatGPT?
No. According to OpenAI, Astra usage is included within existing ChatGPT Plus, Pro, Business, and Enterprise subscription allowances, with additional usage available for purchase as credits once the included allowance is used.

What makes GPT-6 Astra’s cybersecurity capability different from prior models?
TechCrunch reported that in modified benchmark testing, Astra discovered and exploited two previously unknown zero-day vulnerabilities without step-by-step human guidance, in addition to a perfect score on ExploitBench’s known-vulnerability tests. That combination is what pushed the model into OpenAI’s Critical cybersecurity classification.

Is GPT-6 Astra enabled by default for Enterprise accounts?
No. OpenAI has said Astra access is off by default for Enterprise workspaces, and administrators must manually enable it before their organization’s users can access the model.

How does GPT-6 Astra compare to competing models like Gemini or Claude?
Independent comparisons have begun circulating since launch, with reported performance gaps varying significantly by benchmark and task type. No rival lab has published a comparable autonomous zero-day discovery result, which currently sets Astra apart specifically on offensive cybersecurity capability rather than general reasoning performance.

What should developers test first before adopting GPT-6 Astra in production?
Given the premium per-token pricing relative to OpenAI’s cost-optimized models, teams are best served testing Astra on a narrow, high-value workload, such as complex multi-step coding or research tasks, before considering it as a broad replacement for cheaper models already in production.