GPT-6 Astra Usage Limits: Plans, Resets and Real Task Records

GPT-6 Astra does not have one universal prompt limit. In Work and Codex, usage depends on your plan and the work performed, with shared five-hour and applicable weekly allowances. GPT-6 Pro in Chat uses a separate message allowance. API usage is billed separately.

We ran two bounded Astra coding tasks and recorded real account readings. Both passed their acceptance checks. The account’s weekly used meter moved from 56% to 57% during the observation period, but other Codex work was active. That movement cannot be attributed solely to our test.

Official rules and observations checked September 27, 2026 UTC (September 26 Pacific). The experiment lasted about 2 minutes 18 seconds between its first and last meter readings. It is a short session, not a week-long usage diary.

First identify where you are using Astra

On a narrow screen, scroll tables sideways to see every column.

Sources: Astra usage in Work and Codex, GPT-6 Pro in Chat and OpenAI pricing and billing.

Plus users can access Astra in Work and Codex; that does not grant GPT-6 Pro in Chat. Also check the execution surface: OpenAI’s model table lists Astra for local Codex surfaces but not Codex cloud. CLI access requires a sufficiently recent client; OpenAI specifies 0.153.0 or later. Model availability; Access and CLI version guidance.

Astra limits by individual ChatGPT plan

Source: OpenAI’s current usage estimates. These ranges are not an entitlement to their upper endpoint. Context, reasoning, tools and workload affect consumption. Weekly limits can also apply. Currency, tax and regional checkout differences matter.

The $200 tier continues for existing subscribers, but most new purchases and upgrades are paused at this check. The $100 tier remains available. Review eligibility before planning an upgrade or downgrade. Pro availability and return rules. For the buying decision, see ChatGPT Pro vs Claude Max.

Five-hour windows, weekly resets and shared usage

The five-hour window begins with your first Work or Codex message after the previous window ends. A weekly window can apply as well. When both apply, remaining capacity in one does not override an exhausted other window. Read the reset times shown for your account. Source: OpenAI’s Astra usage guidance.

Work and Codex draw on shared capacity, including local and cloud tasks. Other eligible features, including Excel use described in OpenAI’s pricing guidance, can also draw on this pool. Opening another task, changing interface or switching models does not replenish it. A less demanding model may stretch the remaining allowance; it does not start a fresh balance. Shared usage rules.

Check the signed-in Codex usage dashboard or /status in the CLI. An unavailable meter is missing information, not proof of unlimited use. Read whether a display shows used or remaining before recording it.

Our actual task-and-allowance records

We used Codex CLI 0.155.0-alpha.16.4, selected gpt-6-astra, requested low effort and left speed at its default; Fast was not requested. Authentication used the ChatGPT subscription. The account meter reported pro but did not expose whether it was the $100 or $200 tier. We therefore cannot assign these results to either tier.

Both tasks used synthetic Python code, with requirements and acceptance tests frozen before any model request. We instructed the agent to fix the implementation without changing tests, installing dependencies, accessing the network or using subagents. Each task had one run and no human correction prompt.

*These are account readings surrounding each task, not attributable task costs. We read 56% at 04:25:24.733, 57% at 04:26:16.471 and 57% at 04:27:42.381. The reported weekly reset was October 3, 2026 at 16:58:30 UTC. The five-hour reading was unavailable. No reset occurred between these readings.

The elapsed meter observation was 137.648 seconds. Other Codex tasks and the editorial coordination task were running on the same account. The meter reports whole percentages. Consequently, the first row does not establish that interval merging costs one percentage point, and the second does not establish that ledger work is free. Neither row supports a conversion into tokens, dollars or a universal prompts-per-plan figure.

Both intentionally broken baselines failed; the final implementations passed all 23 supplied checks with tests and requirements unchanged. These are two small fixtures, not 23 independent projects. The timing includes CLI startup and tool execution, and should not be treated as general model latency. Fixture design and execution were assisted by Codex.

Download the reproducible evidence pack (ZIP): starting files, prompts, final implementations, source patches, acceptance output, timestamps, sanitized meter readings and SHA-256 manifest. To check a final implementation, enter either task’s after folder and run python3 -m unittest -v.

We did not purchase credits or run paid API tests for this observation. We later ran three paired coding fixtures using existing ChatGPT Pro and Claude Pro subscriptions; Codex CLI and Claude Code each passed six of six scored attempts on the frozen checks. The later runs also lacked isolated account meters, and Claude Max was not tested. Neither test set measures per-task Astra allowance or cross-provider subscription endurance. The Codex vs Claude Code comparison links the paired evidence packs.

Why two prompts can consume different amounts

A short question about one function and a repository-wide investigation can require very different amounts of reading, reasoning and tool work. Higher effort, larger context and longer tasks can increase consumption. Fast mode also has a usage multiplier. OpenAI currently documents 2.5x usage for Fast on its GPT-6 models. Source: model rates and Fast mode.

Use Astra when the task warrants it. For simpler work, OpenAI’s current model lineup also includes Sol and Luna. An editorial workflow worth trying is to give bounded formatting or straightforward edits to a lighter model, then reserve Astra for a difficult investigation. That is a recommendation to test against your work, not a measured saving from our experiment. Current model guidance.

Do not confuse a long-context problem with a quota problem. Compaction can summarize earlier work so a task can continue within context constraints. It does not reset the five-hour or weekly allowance. A task that needs fewer irrelevant files may become easier to handle, but there is no fixed percentage saving promised here.

GPT-6 Pro in Chat has different limits

OpenAI lists 50 weekly GPT-6 Pro messages for Pro $100, shared with GPT-5.6 Sol Pro. Pro $200 has 200 weekly GPT-6 Pro messages, a separate 170-per-day GPT-5.6 Sol Pro allowance and a combined daily ceiling of 200. These are Chat rules; the Astra Work/Codex estimates above do not replace them. Source: Chat model limits.

What to check when Astra stops working

1. Identify the surface and model. GPT-6 Pro in Chat, Astra in local Codex and a hosted Codex job can have different availability and limits.
2. Read every applicable usage window. Record whether each meter shows used or remaining, plus its reset time. A weekly limit can remain exhausted after a shorter window resets.
3. Check simultaneous activity. Other tasks can consume the same account allowance. Closing a tab does not prove every remote task has stopped.
4. Check client version and account. Update an older CLI through the supported channel and confirm the intended subscription is signed in.
5. Distinguish a limit from an execution failure. A denied command, missing dependency or unavailable model is not automatically a quota problem.
6. Choose a supported next step. Wait for the displayed reset, reduce unnecessary work or use a suitable available model if capacity remains. Eligible paid credits are an additional purchase, not included unlimited access.

For broader distinctions between quota, context, credits and billing, see our AI coding plan limits guide. Check any older examples there against the current provider rules linked here.

Frequently asked questions

How many Astra prompts do I get?

There is no single guaranteed count in Work/Codex. OpenAI publishes the plan-dependent estimates above; workload and applicable shared limits determine actual availability.

Does starting a new task reset my allowance?

No. A new task can reduce irrelevant conversation context, but it does not create a new subscription pool.

Can I use the API to keep working?

API access is a separate billing route with its own limits. Confirm model availability and the charges before using it. A ChatGPT subscription is not a general API credit balance.

Does Kingy’s experiment show the cost of one Astra task?

No. It shows two completed tasks and the actual surrounding account readings. Rounded meters and concurrent work prevent a defensible per-task usage calculation. A stronger follow-up would use a quiet account, a verified tier and repeated observations across complete usage windows.