Access troubleshooting · updated September 22, 2026
GPT-6 Astra Not Showing in ChatGPT?
Use this checklist when GPT-6 Astra is missing from the model picker, unavailable in a managed workspace, or returning an access error from the OpenAI API. The first question is whether the problem is the product surface, account, workspace, rollout, model ID, permissions, billing or service status.
Short answer: open a new ChatGPT conversation, confirm the active account and workspace, check the model picker, then compare the result with OpenAI's current model documentation and status page. For the API, use the exact model ID gpt-6-astra, check the organisation and project attached to the key, and read the HTTP status code before changing anything. A ChatGPT subscription and API billing are separate systems.
OpenAI announced GPT-6 Astra as a new model rolling out to a limited set of organisations before broader availability to ChatGPT Plus, Pro, Business and Enterprise users, as well as the OpenAI API and selected cloud platforms. That rollout language matters. A launch announcement is not a guarantee that every account sees the same picker at the same moment. Your account, workspace policy, region, product surface and current service status all matter.
There is also a naming trap. In ChatGPT, a user-facing picker may present a friendly model label. In an API request, the documented identifier is the value that matters. A wrapper may call a feature “Astra” while sending a different model underneath. If you need to know what served a response, inspect the response metadata and keep the request ID. Never publish a test result that says Astra produced something when the request silently used a fallback model.
What the symptoms usually mean
Missing in picker: account, workspace, rollout or UI state.
Model not found: wrong API ID or project access.
Forbidden: permission, organisation policy or entitlement.
429: rate limit or quota, not necessarily missing access.
Billing error: API billing or project configuration.
All requests fail: status incident, account or network.
Step-by-step fixes
1. Confirm that Astra is the model you mean
Try this when: When a guide, screenshot or SDK says only ‘Astra’
OpenAI uses GPT-6 Astra as the model name and gpt-6-astra as the API model ID. Do not confuse the model with a product label such as ChatGPT, Codex or ChatGPT Work. In ChatGPT, the visible picker can show a simplified label, and the available choices depend on account, workspace and rollout. In the API, use the exact model ID from the current OpenAI model documentation rather than copying a name from a third-party wrapper.
2. Check the ChatGPT model picker
Try this when: When ChatGPT is loading normally but Astra is absent
Start a new chat and open the model picker. Look for GPT-6 Astra or the current Astra-labelled option. If it is absent, do not infer that your account is broken from one old conversation: the picker can differ between personal ChatGPT and a managed workspace. Also check whether a model is hidden behind an expanded or more-models menu. A missing option is an availability signal, not something a prompt can fix.
3. Check your plan and workspace
Try this when: When personal and work accounts show different models
Open your ChatGPT account and workspace switcher, then confirm which workspace is active. OpenAI’s launch material says Astra is rolling out to ChatGPT Plus, Pro, Business and Enterprise users. A Business or Enterprise administrator can control model availability through workspace settings, policy and rollout controls. If you are in the wrong workspace, you can see a different picker even while using the same login. Ask an administrator to confirm the workspace policy rather than creating a second account.
4. Sign out, update, and try a fresh session
Try this when: When the picker is stale, blank or inconsistent
Use a private browser window or the latest official ChatGPT app to rule out a stale session. Sign out only after saving any unsent draft. Reopen ChatGPT, sign in to the intended account, start a new conversation and check the picker again. A hard refresh can update a cached interface, but it cannot grant access that the account or workspace does not have. If Astra appears in a fresh session, the problem was session state rather than model access.
5. Check OpenAI status before changing settings
Try this when: When Astra disappeared suddenly or requests fail
Visit OpenAI’s official status page and look for an incident affecting ChatGPT, model serving, login or the API. If there is an active incident, repeated retries can make diagnosis noisier and can consume quota without proving anything. Record the incident link and the time in Africa/Lagos, then retry once the status page reports recovery. Status pages can lag a user-visible problem, so keep the exact error message and request ID if one appears.
6. For the API, use the exact model ID and read the error
Try this when: When ChatGPT access works but an API call fails
A minimal API request should name gpt-6-astra exactly as documented and use an API key belonging to the intended organisation or project. A 404 or model-not-found response usually means the model ID is wrong or the project is not entitled. A 403 points toward permissions, organisation policy or access. A 429 is a rate or quota issue, not proof that Astra is unavailable. A billing error means the request reached account checks but could not be charged. Keep the full status code and request ID; redact the key before sharing logs.
7. Verify billing and project permissions
Try this when: When the API reports forbidden, quota or billing errors
In the API dashboard, select the same organisation and project used by your key. Check that billing is active, the project is not paused, and the key has permission to make model requests. API access is separate from a ChatGPT subscription: paying for Plus does not automatically fund API usage, and an API balance does not change the ChatGPT picker. If you use a team gateway, confirm that it allows the gpt-6-astra model instead of silently mapping to another model.
8. Check regional and organisation rollout
Try this when: When colleagues have Astra but your account does not
A launch can be announced before every account, region, workspace and surface receives access. Use the official OpenAI announcement and model page as the source of truth, then check your own picker or API response. Do not bypass a regional or organisation control with a VPN, shared key or unofficial proxy: that can violate policy and make billing, privacy and support harder. If the official page says access should exist but your account still lacks it, collect account type, workspace, region and timestamp for support.
9. Re-test with a small, low-risk request
Try this when: After Astra becomes visible again
Once the model appears, send a small request that does not require private files, external actions or a long context. Ask for a short structured response and check the model label in the response metadata when using the API. This is an access check, not a benchmark. Do not call a successful hello-world response proof that Astra is faster, smarter or cheaper for your work; measure those claims only with a defined evaluation and retained evidence.
ChatGPT access versus API access
These paths are related but not interchangeable. ChatGPT is a hosted application with plans, a model picker, workspace controls, and product-level limits. The API is a developer service with projects, keys, organisation permissions, billing, rate limits and model IDs. If GPT-6 Astra is visible in ChatGPT but your API request fails, start with API project and billing checks. If your API request works but the ChatGPT picker is missing Astra, start with the active workspace and product rollout. Buying one does not automatically repair the other.
Business and Enterprise users have another layer: an administrator can decide which features and models are available. Ask for the policy to be checked in the workspace where the problem occurs. A screenshot from a personal account is not evidence about a managed workspace. Likewise, a colleague's successful request proves only that their project and permissions work; it does not prove that yours has the same entitlement.
A decision path that prevents wasted retries
Use a simple split. If the ChatGPT website itself will not load, begin with network, login and status checks. If the website loads but the model is absent, begin with account, workspace, plan and rollout checks. If the model appears but generation fails, capture the exact message and test a short request in a new chat. If only the API fails, leave the ChatGPT product alone and inspect the API project, key, model ID, permissions, billing and quota. This keeps unrelated fixes from obscuring the actual fault.
Change one variable at a time. For example, test the same account in a private browser window before switching networks; test the same API request with a newly created project key only after checking the original project; and check status before changing a prompt. If you change the account, browser, network, model ID and prompt at once, a successful result tells you almost nothing about the original cause.
Keep a small diagnostic record. Write down the local time, product surface, account or workspace, model label, exact request shape, response status, request ID and what changed between attempts. This is useful for your own runbook and for support. It also prevents a common mistake: remembering that “it worked yesterday” while forgetting that yesterday used a different project, model alias or workspace.
Availability is not a quality benchmark
Once access is restored, keep access verification separate from quality evaluation. A short structured response can show that a request reached a model, but it cannot establish factual accuracy, reasoning quality, latency, cost or suitability for your workload. If you need to compare Astra with another model, define the task set first, preserve the prompts and inputs, record the model ID and date, and score against criteria that matter. Do not turn one successful request into a claim that the model always follows instructions or produces a particular result.
For production, add a fallback policy that is explicit to the user and visible in logs. If Astra is unavailable, the application should either stop and ask for approval or use a named, approved fallback. It should not silently swap models while your dashboard still labels the output “GPT-6 Astra.” A transparent fallback makes cost, privacy, quality and incident review possible.
For teams, put this policy in the runbook beside the owner and escalation path. Define who can approve a fallback, which data classes may be sent to it, how users are told about the change, and how the event is closed. For a personal workflow, the equivalent is a note in your prompt library: name the intended model, name an acceptable alternative, and ask the interface to disclose when the alternative was used. This small habit prevents a temporary access issue from becoming an undocumented change in output or data handling.
How to report a real access problem
A useful support report is reproducible and minimal. Include the product surface, account or workspace type, country or region, browser or app version, timestamp with timezone, exact model label or ID, exact error text, HTTP status and request ID if applicable, and the result of one fresh-session test. Do not include API keys, personal documents, private conversation content or unnecessary customer data. Say what you expected and what actually happened. “Astra is missing” is less actionable than “Plus account, personal workspace, web app, new private window, picker has GPT-5.6 but no GPT-6 Astra at 08:35 WAT, status page green.”
Do not repeatedly rotate keys, create duplicate accounts, or use an unofficial proxy to force access. Those actions can create a second problem involving security, billing, data residency or terms. If access is rolling out, the safest action is to monitor the official source and retry after a reasonable interval while preserving evidence of the original failure.
Prompts for a safe access check
These prompts help you label assumptions and avoid mistaking a fallback for Astra. They cannot grant access or override a workspace policy.
What not to conclude
Astra appearing in a picker does not mean every feature is enabled, every region has identical limits, or every account has identical context and output allowances. A successful short request does not prove performance. A missing picker option does not prove a permanent exclusion. A third-party wrapper's “Astra” label does not prove which API model ran. Keep those claims separate until the product or response metadata supports them.
Likewise, avoid promising a specific price from a static article. OpenAI publishes model pricing and can change plan limits or API rates. Check the current model page and pricing page on the day you deploy a product or commit to a budget. If you are writing an internal runbook, store the verification date and an owner who will review it after a model or plan announcement.
FAQ
Why can my colleague see GPT-6 Astra but I cannot?
You may be using a different account, workspace, plan, region, rollout cohort or app session. Compare those variables before concluding that one account is broken.
Does ChatGPT Plus include API access?
No. ChatGPT plans and API billing are separate. A Plus subscription can affect the ChatGPT product, while an API project needs its own billing and permissions.
What is the API model name?
OpenAI's model page documents gpt-6-astra. Confirm the live page before shipping because model IDs and aliases can change.
Should I use a VPN if Astra is missing?
No. A VPN does not establish entitlement and can complicate security, region and support diagnosis. Use official rollout and account channels.
How do I know whether a fallback model answered?
In the API, inspect response metadata and logs. In a hosted product, use the model label the interface exposes and avoid claiming more precision than the product provides.