9121 OpenAI launched GPT-6 Astra on September 3, 2026, and called it the best model it has ever shipped for software engineering, computer use, browsing, cybersecurity, and professional work. OpenAI president Greg Brockman went further, describing it as a potential "generational leap" toward artificial general intelligence (openai.com).
If you build with AI coding tools, you have probably already seen Astra show up in a model picker somewhere. This article explains what GPT-6 Astra actually is, what changed for coding specifically, and how the rollout played out across Codex, GitHub Copilot, and Lovable over the following days.
Key Takeaways
- OpenAI launched GPT-6 Astra on September 3, 2026 as its new flagship model, claiming state of the art results in software engineering, computer use, and cybersecurity.
- Reasoning benchmarks jumped sharply, 97.6% on FrontierMath Tier 4, up from 80.5%, but coding specific gains were smaller, with Terminal-Bench 4.0 moving from 37.3% to about 57.9%.
- The real coding upgrade is Codex's new context-notes system, which keeps durable, searchable notes across long sessions instead of compacting everything into one lossy summary.
- The rollout was staged and messy: paying Plus, Pro, and Business subscribers lost access after launch, prompting Sam Altman's public apology and compensation via banked resets.
- Astra reached Codex first, then GitHub Copilot across ten surfaces, then Lovable as a runtime model apps can call, becoming the default across major coding tools within days.
- For product managers, context-notes matters most: it cuts the risk of an agent forgetting an earlier decision during multi-hour agentic work like PRD-to-code passes or discovery synthesis.
Learn this hands-on
Become a 10x PM by learning how to use Claude Code in your daily work as a Product Manager, through 3 highly efficient live sessions of 1h30. Join the Claude Code for PMs live cohort.
What Is GPT-6 Astra?
GPT-6 Astra is OpenAI's new flagship large language model, replacing GPT-5.6 Sol as the company's top-tier model. OpenAI positions it as state-of-the-art across several domains at once: computer use, web browsing, software engineering, cybersecurity, science, and general professional work.
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%. Those are dramatic jumps on reasoning benchmarks. On coding-specific evaluations, the gains were 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.
Pricing is $10 per million input tokens and $50 per million output tokens, with cached input at $1 per million tokens. Usage inside Codex and ChatGPT counts against existing subscription allowances before drawing on paid credits.
What Is Actually New for Coding: the Context-Notes System
The single most consequential change for developers is not a benchmark number, it is how Astra handles long coding sessions inside Codex.
Historically, when a Codex session's context window filled up, the agent compacted everything into one running summary. That process could quietly drop details: why an earlier fix failed, how a specific component behaves, what a teammate said three hours into a session. With Astra, Codex can instead keep durable notes across context windows, preserving accumulated details without collapsing them into a single lossy summary. Earlier context windows also stay searchable, so Astra can go back and retrieve a specific requirement or test result from earlier in the session, even if it was never captured in the running notes.
This matters for anyone who has watched an agent forget a constraint it agreed to an hour earlier. Context-notes is currently an opt-in experimental feature in Codex, and OpenAI has said it plans to make it the default behavior for Astra sessions in the coming weeks. A same-day Codex CLI patch (version 0.153.1) also let API callers configure GPT-6 Astra explicitly, ahead of it becoming broadly available. Codex's rapid patch cadence looks a lot like what we have seen play out with Claude Code's Ultracode updates, where agentic coding tools now ship meaningful capability changes on a near-weekly basis.
The Staged Rollout: Codex, GitHub Copilot, and Lovable
OpenAI shipped Astra as a staged rollout rather than a single flip of a switch, and that staging is worth understanding if you were confused about why the model appeared for some people before others.
Codex
Access started with organizations in OpenAI's application-based "Daybreak" cybersecurity program, then extended to ChatGPT Plus, Pro, Business, and Enterprise plans, and finally to the API and AWS. The caution came from Astra's jump in cyber-capability: at launch it refuses about 91.5% of disallowed cyber-jailbreak prompts, compared with 59% for Sol, and OpenAI applied extra safeguards under its Preparedness Framework following an earlier security incident.
The rollout did not go smoothly. Many paying Plus, Pro, and Business subscribers found themselves without access even after the public launch, which drew visible frustration. Sam Altman posted a public apology for the "messy rollout," comparing it to earlier bumps during the GPT-5 launch. OpenAI compensated affected users with "banked resets": one bonus usage reset for every day a subscriber lacked Astra access. Codex lead Thibault Sottiaux confirmed on September 5 that the rollout had finished ahead of its original weekend target, closing the compensation program with a final reset for every Plus, Pro, and Business user that day.
On September 4, Codex CLI 0.153.4 made GPT-6 Astra the bundled default model whenever no model is explicitly configured, meaning any new Codex session without a saved preference now lands on Astra automatically instead of Sol. OpenAI's guidance for accounts not yet entitled to Astra is to explicitly pin GPT-5.6 Sol as a stable fallback in the meantime.
GitHub Copilot
GitHub Copilot added GPT-6 Astra to its model picker on September 4, one of the widest-reaching model launches it has shipped: ten surfaces at once, including VS Code, Visual Studio, Copilot CLI, the coding agent, github.com, GitHub Mobile, JetBrains IDEs, Xcode, and Eclipse (GitHub changelog).
GitHub frames Astra's edge as behavioral rather than purely benchmark-driven: the model plans and validates as it goes, batches diagnosis with verification, and independently confirms its results before declaring a task done. On an offline OSWorld 2.0 subset, Astra scored 72.6% in about 40 minutes per task, versus 65.7% and roughly 75 minutes for GPT-5.6 Sol, a jump in both accuracy and speed. Access is limited to Copilot Pro+, Max, Business, and Enterprise subscribers, so Free and base Pro users do not get it in this rollout. Business and Enterprise admins should note that new models are auto-enabled by default under the model policy in Copilot settings, unless an admin has already turned that default off. For a full breakdown of how Copilot's command-line tooling stacks up against Claude Code, see our GitHub Copilot CLI vs. Claude Code comparison.
Lovable
Lovable added GPT-6 Astra a few days later, on September 8, alongside Gemini 3.8 Flash and Gemini 3.5 Transcribe (Lovable changelog). This is a different kind of addition from the Codex and Copilot rollouts: it is not the model Lovable itself uses to generate your app's code, it is a model your app can call at runtime for its own AI features.
Astra is positioned in Lovable as the option for demanding work inside an app you are building: complex reasoning, coding assistance, research, or document creation, at a premium price. Gemini 3.8 Flash became Lovable's new default chat model for app-level AI features, accepting text, image, audio, and video input. Gemini 3.5 Transcribe became the default speech-to-text model, with automatic language detection across more than 85 languages. Builders can either name the specific model they want or describe the desired behavior and let Lovable choose.
GPT-6 Astra for Product Managers
If you are a PM rather than an engineer, the practical question is not "how good is Astra at FrontierMath," it is "what changes in how my team ships." A few things do.
The context-notes system in Codex is the one to watch closest. Long-running agent sessions losing track of an earlier decision is one of the most common ways AI-assisted coding quietly produces the wrong thing, a requirement gets summarized away, a test result gets forgotten, and the fix ships anyway. A model that keeps durable, searchable notes across a session is a meaningful reliability improvement for exactly the kind of multi-hour agentic work PMs now ask engineering (or their own agents) to do: a discovery synthesis, a multi-file refactor, a PRD-to-code pass. If you want this level of control over agentic coding sessions yourself, our Claude Code mastery course walks through exactly this workflow.
The staged, occasionally messy rollout is also a useful data point on its own. Even a company as resourced as OpenAI struggled to align its "available in the coming days" messaging with what paying users actually saw. If you are planning a launch, a migration, or a demo around a new model release, build in slack for exactly this kind of staggered access.
Product manager and want to work like this? This is exactly what we teach in Claude Code for PMs, our live cohort for product teams: 3 live sessions of 90 minutes over 2 weeks. Every PM ships a real feature, builds their own agent, and gets personalized written feedback.
The Bottom Line
GPT-6 Astra is OpenAI's new flagship model, and for coding specifically, the headline change is not the benchmark jump, it is the context-notes system that lets Codex keep durable, searchable notes across long sessions instead of compacting everything into one lossy summary. It reached Codex first (with a rocky, apology-inducing rollout), then GitHub Copilot across ten surfaces, then Lovable as a model apps can call at runtime. For how Astra stacks up against other leading models for AI-assisted development, see our guide to the best AI model for vibe coding in 2026. If you use any of these tools, Astra is either already your default model or will be soon, worth understanding before it quietly starts making decisions in your codebase.
