GPT-6 Astra's official API context window is 1,050,000 tokens. That does not mean every Codex session exposes all 1.05M tokens. In the public Codex catalog checked on September 9, 2026, Astra has a default context_window of 272,000 and a max_context_window of 872,000 for configuration overrides. Codex also reserves 5% of the resolved window for system prompts, tool overhead, and model output. That is why a default session may report about 258,400 usable input tokens: 272,000 × 95%.
The 1.05M API specification, Codex's model metadata, and the context available to a particular account are different layers. The numbers are not interchangeable, and 258.4K is not a new API limit.
OpenAI released GPT-6 Astra on September 3, 2026 for difficult end-to-end work across reasoning, coding, computer use, research, and professional documents. Its API model ID is gpt-6-astra; OpenAI does not document gpt-6 as a drop-in API alias. This guide explains the context-window difference first, then covers access, cost, Coding Agent fit, and a controlled way to test the model. Facts were checked against OpenAI's official pages and the official openai/codex repository on September 9, 2026.
GPT-6 Astra at a glance
The official GPT-6 Astra model page is the API contract. OpenAI's launch report covers the release and rollout. For the Codex layer, inspect the official models.json, ModelInfo definition, and runtime context calculation.
Why Codex may show about 258K instead of 1.05M
Three separate numbers answer three separate questions.
1.05M is the API model specification
OpenAI's API model page lists a 1,050,000-token context window and 128,000 maximum output tokens for gpt-6-astra. This describes the model contract on the API route. It does not promise that every product or client will configure each session to use the full window.
272K is the current bundled Codex default
The public Codex model catalog checked on September 9 sets Astra's context_window to 272,000. The same entry sets max_context_window to 872,000 and marks experimental context as supported.
The Codex source describes max_context_window as the maximum allowed for configuration overrides. It is not the bundled default, and it is not a promise that every account, client, provider route, or workspace can activate that value. A remote catalog, cached catalog, client version, account policy, or configuration can change what a real session resolves.
About 258.4K is the default effective input budget
Codex's ModelInfo uses a default effective_context_window_percent of 95. The source says this is the share considered usable for inputs after reserving headroom for system prompts, tool overhead, and model output. At runtime, Codex multiplies the resolved window by that percentage:
272,000 × 95% = 258,400
This calculation explains why a Codex surface may show approximately 258.4K for a default Astra session. Treat it as a result of the public catalog and default runtime rule checked on September 9, not as a universal entitlement or permanent display value. Check the active client and session rather than assuming another person's number applies to your account.
One more distinction matters: the 272K API pricing threshold and the 272K Codex catalog default happen to use the same number, but they have different meanings. Crossing 272K of API input changes the price of the full request; it does not reduce the API context window to 272K. Codex's 272K value is client-side model metadata for its default context configuration.
What changed for coding agents
OpenAI positions Astra for complex work that spans more than a single answer. The model page lists coding-oriented Responses API tools such as hosted shell, apply patch, skills, computer use, MCP, web and file search, and code interpreter. Its large API context and multiple effort levels also give an integration more room to trade speed, cost, and depth.
These are model and API capabilities. A Coding Agent benefits only when its Agent harness exposes them correctly and gives the model useful evidence, permissions, and verification. If the distinction is new, read why an Agent harness and a model are separate choices.
For a real task, Astra's potential value falls into three groups:
- Long, related evidence: a migration, investigation, or review may need code, logs, screenshots, documentation, and prior decisions in one working set.
- Multi-step tool use: an agent can inspect, edit, run tests, check the result, and recover when one step fails.
- High-cost mistakes: deeper reasoning may be worth more when a wrong migration, security conclusion, or production change would create expensive repair work.
None of those is a promise of success. A large context window is capacity, not proof that the model will retrieve the right file or preserve every constraint. A tool listed on an API page is not proof that a particular harness passes the same tool, arguments, and output back to the model. OpenAI's performance statements are reasons to test Astra, not substitutes for your own acceptance checks.
Availability is not one switch
When someone says “GPT-6 Astra is available,” ask which layer they mean:
- Released: OpenAI publicly announced GPT-6 Astra on September 3.
- Documented: the API has an official
gpt-6-astramodel page and contract. - Account access: a specific API organization, subscription, or workspace is eligible to start the model.
- Product surface: ChatGPT, Codex, a cloud marketplace, or another product displays the model.
- Harness support: the selected client can pass the context, tools, effort setting, and events the workflow requires.
OpenAI's current Astra usage guide for Work and Codex says access still depends on plan, workspace permissions, rollout eligibility, and an up-to-date client. It also separates included Work/Codex allowances from usage billed through an API key. Do not infer API access, a particular context size, or a fixed task count from the fact that Astra appears in another OpenAI surface.
Codex has its own version boundary. The current public catalog declares 0.153.0 as Astra's minimum client version; use the current stable release rather than treating that minimum as the recommended version. Client compatibility is still separate from model visibility, account eligibility, provider route, effort settings, and usage allowance. The dedicated guide to using GPT-6 Astra in Codex walks through those checks.
A client that recognizes gpt-6-astra can still receive an access error. A model visible in one OpenAI surface may still be absent from another. A model available directly from OpenAI may still be unsupported in an independent third-party product.
Cost, context, and the 272K API boundary
Astra's Standard API rates are $10 per million uncached input tokens, $1 per million cached-input tokens, $12.50 per million cache-write tokens, and $50 per million output tokens. Batch and Flex are listed at half of Standard; Fast is twice the applicable rate. Tool fees can be additional.
The load-bearing detail is the 272,000-input-token pricing boundary. Above it, OpenAI applies 2× input and cache rates and 1.5× output rates to the entire request. A 300K request is not billed at the standard rate for the first 272K and a premium only for the remainder.
This makes context selection an economic decision even when a route technically supports a larger window. A repository may fit inside the maximum, but a harness that retrieves only the relevant working set is usually cheaper and easier to verify. Stable cached prefixes can help, but cache writes, misses, changing context, output, retries, and tools still count.
The GPT-6 Astra API pricing guide separates API token billing from ChatGPT plans and Codex usage, then works through short, long-context, cached, Batch/Flex, and Fast examples. Do not convert a subscription allowance into an invented API token price.
How to read the benchmark claims
OpenAI's launch report lists strong Astra results across terminal, coding, computer-use, science, and safety evaluations. For example, it reports 57.9% on Terminal-Bench 4.0 and 64.6% on Terminal-Bench Science 0.1, ahead of the Claude Fable 5.1 results shown in the same OpenAI table.
Those are vendor-reported evaluations. OpenAI says scores are the maximum at any effort and that research or API conditions can differ from production ChatGPT because of prompts and tools. A score without the task version, harness, tools, effort, retries, safeguards, and cost definition is not a routing policy.
Use the GPT-6 Astra benchmark evidence audit to inspect those conditions. Then use the comparison that matches the decision you face:
- GPT-6 Astra vs GPT-5.6 Sol for the OpenAI family upgrade question; or
- GPT-6 Astra vs Claude Fable 5.1 for the cross-provider model, cache, and harness decision.
Neither comparison should declare a universal winner without the same starting task and a result you can verify.
A practical coding-agent value map
The decision is not “most capable model or no AI.” A practical routing policy usually keeps a lower-cost lane for repeatable tasks, an advanced lane for ambiguity, and a protected escalation lane for high-consequence work.
The published maximum should also not become permission to send an unfiltered repository. Search first, load the evidence needed for the task, preserve acceptance criteria explicitly, and split independent work when one giant request becomes harder to audit.
A controlled next step
Before changing a team's default, run one bounded task and record the route precisely:
- Confirm the harness, client version, provider, account access, exact model ID, effort setting, and context configured for the active session.
- Use a task with a known answer or deterministic tests, not a subjective demo.
- Record uncached input, cache writes and reads, output, tools, retries, elapsed time, and reviewer corrections.
- Check whether the result stays within file, permission, network, and destructive-action boundaries.
- Compare total cost per accepted result, then repeat on a normal multi-file task and a failure-prone investigation.
Fluency is not the acceptance test. A strong result should leave the requested artifact, pass the relevant checks, respect scope, and expose enough evidence for another person to review it.
Agent.Space is independent from OpenAI and currently includes GPT-6 Astra in every Share plan. The GPT-6 Astra Workspace guide explains the task fit, persistent project workflow, and current product boundary. Use the live model selector as the source of truth for the compatible Agent combinations available to your account, then apply the same controlled test to a combination that is actually displayed.
Start an Agent.Space session and begin with one task whose result you can verify.
FAQ
What is the GPT-6 Astra context window?
OpenAI's API model page lists a 1,050,000-token context window and 128,000 maximum output tokens for gpt-6-astra. That is the API contract. A Codex client can expose a smaller default or effective window for a session.
Why does Codex show about 258K for GPT-6 Astra?
The public Codex catalog checked on September 9, 2026 sets Astra's default context_window to 272,000. Codex's default 95% effective ratio reserves headroom for system prompts, tools, and model output, leaving about 258,400 tokens for inputs. This value may differ with the client, catalog, account, route, or configuration, so it is not a universal promise.
Can Codex use more than 272K with GPT-6 Astra?
The current public catalog marks experimental context as supported and lists 872,000 as the maximum allowed for configuration overrides. That is a catalog ceiling, not the default and not guaranteed access for every client or account. Verify the configuration and effective context of the session you are actually using.
Does the 272K API pricing threshold cap Astra at 272K?
No. The API model contract lists a 1.05M context window. More than 272K input tokens triggers higher rates for the full request; it does not turn 272K into the API context limit. The same number in Codex's bundled default describes a different layer.
When was GPT-6 Astra released?
OpenAI released GPT-6 Astra on September 3, 2026. Access remains route- and account-specific, so check the current model picker, workspace settings, API organization, and official usage guidance rather than relying only on the release date.
Is GPT-6 Astra available in Agent.Space?
Yes. Agent.Space currently includes GPT-6 Astra in every Share plan. Check the live Agent.Space model selector for the compatible Agent and model combinations available to your account. Visibility establishes availability for that route, not a performance endorsement.