By GPT-6 Astra landed on September 3, and by September 4, Codex CLI v0.153.4 had made it the default model for sessions without an explicit model configured. That is a 24-hour turnaround for a model OpenAI itself flagged as the first to cross the “Critical” cybersecurity threshold — meaning it can find unknown vulnerabilities and write working exploits against hardened systems. The capability jump is real. So are the sharp edges. Here is what you actually need to configure before Astra quietly takes over your coding sessions.
The 1.05M Context Window (And the 258K Gotcha)
Astra ships with a 1,050,000-token context window — the largest in OpenAI’s family. In practical terms, a mid-sized repository with all its source files, tests, and documentation fits in a single session without a compact cycle. No more losing the thread halfway through a refactor because Codex had to summarize 40,000 lines into a paragraph.
There is a catch. If you spin up Astra in Codex v0.153.4 without touching your config, you will likely see a context limit around 258K, not 1.05M. The default reasoning level in the bundled catalog is low, and at that level Astra runs with a reduced context window. Set reasoning to medium or higher and the full 1.05M unlocks.
One more thing on context: requests over 272K input tokens hit a long-context surcharge — $20 per million input tokens instead of the standard $10, and $75 out instead of $50. Worth knowing before you load an entire monorepo into a single prompt.
How to Configure Astra in Codex
The model ID is gpt-6-astra — exact string, case-sensitive. If you want to try it in a single session without touching your global configuration, this is the command:
codex -m gpt-6-astra --reasoning-effort xhighAstra’s reasoning ladder runs: Low, Medium, High, xhigh, Max, Ultra. The bundled catalog default is low, which is conservative on both context and cost. For actual coding work — debugging, refactoring, code review — medium or high gives you the full context window and noticeably better reasoning without blowing up your token bill. Save xhigh and max for genuinely hard problems.
Context Notes: The Feature Most Coverage Is Missing
The most interesting thing about Astra in Codex is not the benchmark numbers. It is context notes.
Previously, when a Codex session filled its context window, it summarized everything into a single compressed block. Details fell through: why a particular fix failed, how a component behaved during testing, a design constraint you mentioned in message three. The summary kept the gist; the nuance evaporated.
Astra introduces an experimental mechanism that keeps earlier context windows searchable rather than collapsing them. The model can reach back and retrieve a specific test result or requirement from a window that has already scrolled out of scope. It is not perfect — OpenAI explicitly marks it experimental — but it is a qualitative improvement for long-running sessions. Read more in OpenAI’s GPT-6 Astra announcement.
Enable it with one line in your codex.toml:
[experimental]
context_notes = trueImportant limitation: at launch, context notes only works for ChatGPT Plus and Pro accounts. API key logins, Business plans, and Enterprise accounts are not supported yet. OpenAI says this will change in the coming weeks. If you are on an API key, you get the larger context window but not the cross-window retrieval.
Enterprise: Someone Needs to Flip a Switch
If your team is on a Business or Enterprise plan, Astra is disabled by default and will stay that way until a workspace administrator enables it through the OpenAI admin console. This is not a bug — it is a deliberate decision tied to the Critical cybersecurity rating.
Astra is the first model OpenAI has assessed at Critical under its Preparedness Framework. That means, given the right tools and access, it can identify previously unknown vulnerabilities and develop working exploits across hardened systems without step-by-step human guidance. Your security team is going to ask about this model. Get ahead of the conversation rather than explaining it after someone notices the new default.
For individual developers and personal API keys, there is no toggle. Update to Codex v0.153.4 and Astra becomes the default when no model is specified.
What It Actually Costs
Standard pricing: $10 per million input tokens, $50 per million output tokens. Those numbers put Astra at roughly twice the cost of GPT-6 Sol for equivalent work. See the full GPT-6 Astra pricing breakdown for a complete rate comparison.
Cache changes the math significantly. Cached input reads at $1 per million tokens — a 90% reduction. Cache writes cost $12.50 per million tokens, meaning you need 1.25 cache reads to break even on a written block. For any workflow with a large AGENTS.md, long system prompts, or repeated tool definitions, the cache pays back fast. A 100K-token AGENTS.md read 50 times costs $1 cached versus $50 uncached.
The case for staying on Sol: if your work is output-heavy with short system prompts and you are not bumping against context limits, Sol remains more economical. Astra is worth the premium for complex multi-step reasoning, long-context work, and sessions where you actually need the 1.05M window. Check the Codex CLI v0.153.4 release notes for the full change list.
The Verdict
Update to Codex v0.153.4. Set reasoning to medium or higher to get the actual context window. If you run long sessions with large system prompts, check your cache configuration — the economics shift meaningfully. Enterprise admins: review the OpenAI admin console before Astra becomes a surprise in your team’s workflow. And if context notes land for API keys in the next few weeks, that is when Astra becomes the obvious default for serious Codex work.
