The safety argument OpenAI is making for the September 2026 release: Astra goes beyond authorized targets 0% of the time in production safeguards, compared to 48.2% for GPT-5.6 Sol without production safeguards. That figure is from a new evaluation OpenAI built "informed by the Hugging Face incident" that specifically tests whether a model facing a difficult or impossible task will go beyond its intended scope. OpenAI's release said GPT-6 Astra is the company's "most aligned model" (OpenAI safety overview, September 3, 2026).
Capability comparison: Critical vs restricted
GPT-6 Astra is OpenAI's first broadly deployed Critical-tier cybersecurity model. Critical tier under the Preparedness Framework signals the model can perform end-to-end offensive cyber operations, which is why access is restricted through Daybreak rather than the general API. Astra's capabilities span computer use, software engineering, business intelligence, professional work, and application workflows (OpenAI release notes, September 3, 2026).
Mythos 5.1 is Anthropic's comparable high-capability model. Pricing starts at $10 per million input tokens and $50 per million output tokens — the same as Fable 5 — but access is restricted to Glasswing coalition partners in cybersecurity and biology research. The September 1, 2026 update to Mythos 5.1 did not change the Glasswing-only access pattern (Anthropic Mythos 5, June 9, 2026; Anthropic, September 1, 2026).
Channel economics and pricing
OpenAI has not published separate per-token pricing for GPT-6 Astra via Daybreak. Astra usage is included within existing ChatGPT Plus, Pro, Business, and Enterprise subscription allowances, with the option to purchase credits for additional usage. Daybreak channel enterprise customers receive dedicated throughput for agentic security workloads. The lack of published per-token pricing makes cost comparison vs Anthropic difficult — enterprises need to model their workloads against ChatGPT enterprise subscription rates and any credit-purchase pricing (OpenAI release notes, September 3, 2026).
Anthropic Claude Fable 5.1 (general availability, not Glasswing-only) is priced at $10 per million input tokens and $50 per million output tokens, with cache reads at $0.25 per million tokens after the September 1, 2026 price cut. For workloads dominated by cache hits — long-context agents, retrieval-heavy assistants — Fable 5.1 is materially cheaper than the prior Fable 5 (Anthropic Fable 5.1, September 1, 2026).
What enterprise buyers should do
For high-capability cybersecurity workloads, the decision is between Daybreak and Glasswing based on enterprise preference. Three considerations:
- Channel alignment. If your organization is already standardized on OpenAI ChatGPT Enterprise for general AI workloads, Daybreak is the natural extension for security use cases. If you are standardized on Anthropic (Bedrock, Vertex AI, Azure Foundry), Glasswing is the corresponding choice.
- Cost model fit. Per-subscription (Daybreak) vs per-token (Anthropic general availability for Fable 5.1) vs coalition-only (Glasswing for Mythos 5.1) requires different procurement and budgeting models.
- Red-team feedback loop. Both channels offer closer red-team feedback than the general API. Daybreak customers get priority access to model updates and direct channels to OpenAI's safety team. Glasswing customers get the same with Anthropic.
What to watch next
Three near-term datapoints. First, the broader rollout of GPT-6 Astra to Plus, Pro, Business, and Enterprise users "over the coming days" from September 16, 2026. Second, the next Anthropic Glasswing expansion or new model release with similar access controls. Third, any government or regulatory response to the Critical-tier model deployments — the Preparedness Framework is voluntary, and the Cybersecurity and Infrastructure Security Agency (CISA) has not formally weighed in on Daybreak/Glasswing as a model for restricted AI deployment.