GPT-6 Astra
CodexOpenAI launched GPT-6 Astra, its new flagship model, on September 3, 2026, calling it the best model yet for software engineering, computer use, browsing, cybersecurity, and professional work. For Codex specifically, Astra introduces an experimental context-notes system that lets long-running sessions keep durable notes across context windows instead of repeatedly compressing history into a single summary, with earlier context remaining searchable. The model is rolling out in phases, starting with OpenAI's Daybreak cybersecurity-vetted organizations before reaching ChatGPT Plus, Pro, Business, and Enterprise plans, plus the API and AWS in the coming days. OpenAI president Greg Brockman called it a generational leap, while independent coverage was more measured about the size of the coding-benchmark jump.
Featured Video
A video we selected to help illustrate this changelog
Key Takeaways
- Astra is OpenAI's new flagship model, which the company explicitly calls its best model for software engineering to date.
- Codex gains an experimental context-notes system, keeping durable notes across context windows instead of compacting history into one lossy summary, with earlier context still searchable.
- The context-notes behavior is opt-in at launch, and OpenAI says it plans to make it the default for Astra sessions in the coming weeks.
- The benchmark gains are uneven: FrontierMath Tier 4 jumps to ~97.6% from 80.5% and ARC-AGI-3 to 99.9% from 7.8%, but Terminal-Bench 4.0 moves only from 37.3% to 57.9%.
- Rollout is staged for safety, starting with Daybreak cybersecurity-vetted organizations before ChatGPT Plus, Pro, Business and Enterprise, then the API and AWS; Astra refuses about 91.5% of disallowed cyber-jailbreak prompts versus 59% for Sol.
- Pricing is $10 per million input tokens and $50 per million output, with cached input at $1 per million; Codex and ChatGPT usage draws on existing subscription allowances.
Sources & Mentions
5 external resources covering this update
OpenAI's Newest Flagship Model Arrives
OpenAI began rolling out GPT-6 Astra on September 3, 2026, describing it as the product of "years of research and big bets" and, according to OpenAI president Greg Brockman, potentially a "generational leap" toward artificial general intelligence. OpenAI positioned Astra as state-of-the-art across computer use, web browsing, software engineering, cybersecurity, science, and general professional work, with the company explicitly calling it "the best model for software engineering to date."
Independent benchmark trackers backed up parts of that claim while pushing back on others. Astra saturates FrontierMath Tier 4 at roughly 97.6% (up from GPT-5.6 Sol's 80.5%) and posts 99.9% on ARC-AGI-3, versus Sol's 7.8%. On coding-specific evaluations the jump was smaller: Terminal-Bench 4.0 moved from 37.3% to about 57.9%, and on the DeepSWE v1.1 agentic coding benchmark Astra scored competitively with, but not dramatically ahead of, rivals like Opus 5 and Gemini 3.8 Flash. A widely shared Hacker News thread on the coding results was blunt about the gap between OpenAI's "major gains" framing and the underlying chart.
A New Way for Codex to Handle Long Sessions
The most Codex-specific change in this release is a new approach to context management. Historically, when a Codex session's context window filled up, the agent would compact everything into a single running summary, a process that could quietly drop details like why an earlier fix failed or how a particular component behaves. With Astra, Codex can instead keep notes across context windows, preserving accumulated details without collapsing them into one lossy summary. Previous context windows also stay searchable, so Astra can go back and retrieve a specific requirement or test result from earlier in a session even if it wasn't captured in its running notes.
This context-notes behavior currently ships as an opt-in experimental feature for Codex users, with OpenAI saying it plans to make it the default behavior for Astra sessions "in the coming weeks." A same-day Codex CLI patch (0.153.1) also lets API callers configure GPT-6-Astra explicitly without changing the account's default model or its visibility in the model picker, ahead of Astra becoming broadly available.
Safety Posture and Rollout
Because of the model's jump in cyber-capability benchmarks, OpenAI applied extra safeguards under its Preparedness Framework following the earlier Hugging Face security incident. At launch, Astra refuses the most dangerous requests, such as writing proof-of-concept exploits, and refuses about 91.5% of disallowed cyber-jailbreak prompts, compared with 59% for Sol. Reflecting that caution, access is rolling out gradually: companies in OpenAI's application-based "Daybreak" cybersecurity program get it first, followed by ChatGPT Plus, Pro, Business, and Enterprise subscribers, and then the OpenAI API and Amazon Web Services in the days after launch. Pro, Business, and Enterprise members get access to a "GPT-6 Astra Pro" tier, with Enterprise administrators controlling workspace-level enablement.
Pricing
Astra is priced at $10 per million input tokens and $50 per million output tokens, with cached input priced at $1 per million tokens and cache-write at $12.50 per million tokens. Usage inside Codex and ChatGPT counts against existing subscription allowances, with additional credits available for purchase once those are exhausted.