OpenAI Releases GPT-6 Astra: First Model Rated Critical for Cybersecurity

OpenAI Releases GPT-6 Astra on 3 September 2026

On 3 September 2026, OpenAI released GPT-6 Astra, describing it as the world's most intelligent and aligned model and its first system to reach the Critical cybersecurity capability threshold under the company's Preparedness Framework. The rollout began with organisations admitted to Daybreak, OpenAI's application-only cybersecurity programme, with access extending to ChatGPT Plus, Pro, Business and Enterprise subscribers over the following days, and then to the API, Amazon Bedrock and Microsoft Azure. On the API, the model is available under the identifier gpt-6-astra.

What "Critical" for Cybersecurity Means

OpenAI's Preparedness Framework assigns models a cybersecurity capability tier from Low through to Critical, based on their ability to assist with offensive security operations. Critical is the highest tier and describes a model capable of autonomously discovering previously unknown security weaknesses in hardened systems and building functional exploits against them, without a person directing each step. No prior OpenAI model had reached this designation. GPT-6 Astra achieves it by scoring 100 per cent on ExploitBench, a benchmark that measures whether a model can develop working exploits from known software vulnerabilities. On the harder subset — finding vulnerabilities not previously disclosed — Astra performs at a level OpenAI describes as surpassing the practical output of a highly skilled human security researcher working alone.

OpenAI President Greg Brockman told reporters at the launch that at this level of capability, safety had become the company's top priority. OpenAI delayed parts of Astra's development by several weeks to complete additional testing of protections against cyber misuse before proceeding with the rollout.

Benchmark Results: Mathematics, Reasoning, and Coding

Beyond cybersecurity, GPT-6 Astra saturates multiple hard evaluations. On FrontierMath, a benchmark testing research-level mathematical reasoning beyond standard graduate curricula, Astra scores 97.6 per cent. On ARC-AGI-3, a test of novel abstract reasoning that prior models have struggled to make progress on, Astra scores 99.9 per cent. The model accepts up to 922,000 input tokens and generates up to 128,000 output tokens, within a 1,050,000-token context window — well suited to analysis of large codebases, extensive document corpora, and multi-session agentic tasks.

Autonomous Coding and Agent Workloads

GPT-6 Astra exceeds all prior OpenAI models on agentic coding evaluations. The model is designed for workloads where it must navigate a real codebase, write code, run tests, interpret failure output, and iterate to a working solution across many sequential steps without a human directing each tool call. OpenAI positions Astra as the model of choice for engineering teams running autonomous development pipelines and security operations teams conducting authorised penetration testing.

Pricing and API Access

GPT-6 Astra is priced at $10 per million input tokens and $50 per million output tokens. Cached input tokens cost $1 per million. The pricing matches the rates of Claude Fable 5.1, Anthropic's frontier model released two days earlier on 1 September 2026, placing the two at direct cost parity for comparable workloads. The context window of 1,050,000 tokens is comparable with Gemini 3.8 Flash's 1,048,576-token window, released on 2 September 2026. Together, the three releases within 72 hours represent the densest concentration of frontier model launches in a single week since the August 2026 wave.

The Daybreak Programme and Gated Access

Daybreak is OpenAI's application-only cybersecurity programme for organisations that need access to the model's full offensive security capabilities for authorised purposes: red teaming, vulnerability research, penetration testing firms, and government security agencies. Daybreak programme members received GPT-6 Astra on the day of launch, ahead of the general rollout. For ChatGPT paid subscribers outside Daybreak, OpenAI is releasing a version with additional cybersecurity guardrails that limits access to the most advanced autonomous exploit capabilities. OpenAI has not disclosed the eligibility criteria for Daybreak membership publicly, but teams can apply through OpenAI's safety programme page.

What This Means for Software Teams in India

For development teams in India integrating the OpenAI API, GPT-6 Astra is the most capable model at the $10 per million input token price point, with a context window sufficient for full-codebase review and long-horizon autonomous workflows. For teams already running agentic coding pipelines on GPT-5.6 or other predecessors, the agentic coding improvements in Astra translate to higher task completion rates on multi-step development tasks without additional orchestration overhead.

For India's growing base of cybersecurity firms, managed security service providers, and in-house security teams at banks and technology companies, the ExploitBench result has direct operational relevance. A model that achieves 100 per cent on a benchmark for developing working exploits changes the economics of authorised penetration testing: tasks that previously required days of skilled researcher time become achievable in hours. Security teams conducting authorised assessments should evaluate GPT-6 Astra through the Daybreak programme. Teams building defensive security tooling should factor autonomous vulnerability discovery capabilities into their threat model, as adversarial access to comparable capabilities is a realistic near-term concern that Indian security architecture needs to plan for.

The Bottom Line

OpenAI released GPT-6 Astra on 3 September 2026, the first model it has ever rated Critical for cybersecurity under its Preparedness Framework. The model scores 100 per cent on ExploitBench, 97.6 per cent on FrontierMath, and 99.9 per cent on ARC-AGI-3. Its context window is 1,050,000 tokens, with 922,000 maximum input and 128,000 maximum output tokens. Pricing is $10 per million input tokens and $50 per million output tokens, with cached input at $1 per million. Advanced cybersecurity capabilities are gated to Daybreak programme members. General ChatGPT subscribers receive a safeguarded version. The API identifier is gpt-6-astra, with availability on Amazon Bedrock and Microsoft Azure following the initial rollout.