Skip to content
NewsModel release

OpenAI Releases GPT-6 Astra, First Model to Trigger Its Critical Cybersecurity Threshold

· by Pondero Newsdesk

The short version

GPT-6 Astra launched September 3, 2026 with the highest capability rating in OpenAI's Preparedness Framework. API access costs $10 per million input tokens with a pricing cliff above 272K tokens that doubles costs on the full request.

OpenAI Releases GPT-6 Astra, First Model to Trigger Its Critical Cybersecurity Threshold

OpenAI shipped GPT-6 Astra on September 3, 2026, and the most consequential detail was not a benchmark score. Astra is the first model OpenAI has released after crossing the Critical tier in its Preparedness Framework for cybersecurity, a rating that gates which users can access the full model and which capabilities are suppressed in the public version.

What

Astra rolled out first to a limited set of Daybreak partner organizations on September 3, then began reaching ChatGPT Plus, Pro, Business, and Enterprise accounts plus the OpenAI API and AWS over the following days, per coverage of the official announcement by 9to5Mac. Enterprise workspace administrators must enable Astra for their org. Access is off by default at launch. Subscribers on Pro, Business, and Enterprise plans also receive a separate "GPT-6 Astra Pro" tier. Usage for all plans is included within existing subscription allowances, with additional credits purchasable.

OpenAI described the model in its announcement as "the world's most intelligent and aligned model." Per 9to5Mac's reporting on the launch, OpenAI claims Astra scored 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench. Computer use runs nearly twice as fast as the prior generation. OpenAI's harness optimizations also brought a roughly 60% speed improvement to existing GPT-5.6 Sol sessions without a model update.

The context window is 1.05 million tokens with a maximum output of 128,000 tokens, per OpenRouter's model listing. Standard API pricing is $10 per million input tokens and $50 per million output tokens. Cached input is $1 per million. The Flex tier halves all rates to $5 per million input and $25 per million output. A long-context pricing cliff applies to any prompt exceeding 272,000 input tokens: the full request is billed at twice the input rate and 1.5x the output rate.

Why it matters

The Critical cybersecurity rating is structurally different from anything in OpenAI's prior model history. The Preparedness Framework defines ascending capability thresholds; when a model hits Critical, access restrictions follow automatically. Astra's public version filters advanced offensive cyber prompts. A version with fewer restrictions exists only for vetted Daybreak organizations. Enterprise administrators who enable Astra for their workspace will see the safety-filtered build. OpenAI president Greg Brockman called the model a generational leap, per CNBC's September 3, 2026 report.

The long-context pricing structure demands attention alongside the headline rate. The 1.05 million-token context window is the marketing figure, but more than 80 percent of it falls into the 2x billing zone. A prompt consuming 700,000 input tokens costs $140 per million inputs rather than $70. For document-heavy pipelines or multi-session retrieval workflows, the difference between the sub-272K and over-272K scenarios is material before a single line of product code ships. Teams evaluating Astra for agentic tasks that accumulate large working contexts should run both scenarios in a cost model before committing to the model.

The default-off enterprise setting is the other deployment decision worth flagging early. Most OpenAI model updates in 2025 and 2026 arrived enabled for Enterprise accounts by default. Astra requires an affirmative administrator action. That is likely driven by the cybersecurity rating: spreading an unconstrained version of a Critical-tier model to unscreened enterprise end users would conflict with the framework that triggered the restricted rollout in the first place.

For teams using Codex in long coding sessions, Astra introduces a context-persistence mechanism that replaces compaction. Rather than compressing prior work into a rolling summary when the context window fills, Astra can keep notes across windows and keep earlier context searchable. Per the OpenAI announcement quoted in 9to5Mac, this addresses the information loss that compaction causes in extended debugging sessions or large refactors. The feature is opt-in at launch and becomes the Codex default for Astra users in the coming weeks.

Context and reactions

GPT-5 launched August 7, 2025 and was followed by a sequence of interim updates: GPT-5.1, 5.2, 5.3 Instant, 5.4 Thinking (which introduced the 1 million-token context), 5.4 mini and nano, 5.5, and the July 2026 GPT-5.6 Sol, Terra, and Luna family, per the version history documented in 9to5Mac's coverage. Astra is the first model carrying the GPT-6 designation and the first released under the revised Preparedness Framework cybersecurity tier structure. CNBC reported the training run used more than 100,000 GPUs at OpenAI's Stargate facility in Texas, per CNBC, September 3, 2026.

The restricted Daybreak rollout follows the pattern OpenAI used for GPT-5.4-Cyber in April 2026, where a capability-gated model reached screened organizations before broader access opened. The structural difference with Astra is that Astra is the primary production flagship rather than a specialized variant. The cybersecurity program it triggers gates the whole model family going forward, not a single-purpose derivative.

The cost difference versus current alternatives is substantial. Anthropic's Claude Sonnet 5 was confirmed at $2 per million input tokens and $10 per million output tokens as a permanent rate in August 2026. Astra's standard API rate is five times the input cost and five times the output cost. Teams that can fit their workloads inside 272,000 tokens pay $10/$50; those that cannot pay up to $20/$75. The performance gain that justifies that premium will be the central question in third-party evaluations this week.

What to watch next

API access began rolling out September 3 and has been expanding by plan tier. The near-term milestone to watch is whether OpenAI reaches full API availability across all account tiers by end of the week. Independent benchmark comparisons will surface quickly given the ExploitBench saturation claim and its direct relevance to the Preparedness Framework rating. Watch whether major coding IDEs add Astra to their model menus and which enterprise buyers are willing to pay the 5x input-cost premium over the current frontier alternatives.

Sources