GPT-6 Astra arrives, and the obvious temptation is simple: if it is the smartest model, why not run everything on it?
Because your usage allowance may die before the task list does.
OpenAI says Work and Codex usage depends on the model, task complexity, context, reasoning settings, speed, tools, and amount of work performed. It also explicitly notes that Astra can consume the included allowance faster than GPT-5.6 Sol.[1][2]
The practical answer is therefore not “always use Astra.” It is “deploy Astra where intelligence has the highest marginal value.”
Codex supports changing the active model and reasoning effort through /model.[3] But this needs one important qualification: it is not a live hot-swap of the model while a turn is already running. A separate request to support switching during an active prompt has been treated as a different feature.[4]
So the useful workflow is:
Work with Sol → finish the turn → switch to Astra for the hard part → finish → switch back to Sol.
1. Model switching is not automatic model routing
/model lets you choose the model and reasoning effort.[3]
That does not mean Codex automatically optimizes every task for your quota by deciding:
- typo fix: Luna
- normal coding: Sol
- mysterious failure: Astra
- implementation after the diagnosis: Sol again
For now, the user still benefits from deciding where Astra is actually worth spending.
Using Astra to fix punctuation in a README is like taking an F1 car to buy milk. It works. It may even be fast. It is just not the reason you bought the machine.
2. Why Astra can burn allowance quickly
The API price table makes the positioning easy to see.[5]
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| GPT-6 Astra | $10 | $50 |
| GPT-5.6 Sol | $4 | $20 |
| GPT-5.6 Terra | $2 | $12 |
| GPT-5.6 Luna | $0.20 | $1.20 |
Astra is 2.5 times the per-token API price of Sol.
But do not turn that into “my ChatGPT weekly allowance always decreases exactly 2.5 times faster.” API billing and the included Work/Codex allowance are different systems. OpenAI says actual allowance use varies with task location, complexity, context, reasoning, speed and tools.[1][2]
Long repository scans, browser work, repeated experiments, large logs and multi-step testing can therefore consume much more than a short request.
If you tell Astra “read everything, think about everything, test everything, and fix everything” for hours, you are using the research director as the loader, receptionist and photocopier too.
3. Exploration is where Astra makes the most sense
The best Astra jobs are not merely “large jobs.” They are jobs where the answer is unknown at the start:
- generate competing hypotheses for an unexplained failure
- infer structural problems across a repository
- find the shared cause of a recurring bug
- rank many candidates by expected value of testing
- design the next experiment from failed results
- compare architectural trade-offs
- make a difficult final decision after evidence has been collected
OpenAI positions Astra as its highest-capability model for complex reasoning, coding, browsing, software engineering and multi-step workflows.[6]
That makes it especially valuable as an exploration and diagnosis model.
4. The strongest pattern is Sol → Astra → Sol
A practical quota-efficient workflow looks like this:
Stage 1: gather cheaply with Sol, Terra or Luna. Collect files, logs, test results, candidate locations and existing implementations.
Stage 2: compress and escalate to Astra. Give Astra a smaller, information-dense package: “Seven candidates remain. These three fixes already failed. Rank the next hypotheses.”
Stage 3: validate with Sol or Terra. Run tests, implement fixes and benchmark alternatives.
Stage 4: return to Astra only if the evidence still does not converge.
In this pattern Astra is not the worker reading one hundred files one by one. It is the research director called in after the evidence table is ready.
5. A simple role split
This is an operational heuristic, not a permanent capability ranking:
| Task | First choice |
|---|---|
| simple search, formatting, repetitive edits | Luna / Terra |
| normal implementation, review, common bug fixing | Sol |
| unknown-cause exploration, hard architecture, hypothesis generation, final hard cases | Astra |
The principle is simple: spend Astra where Astra changes the decision, not where it merely completes routine work.
6. Check usage instead of counting messages
OpenAI recommends checking the usage dashboard for the specific exhausted allowance, credit balance and reset time. In an active Codex CLI session, /status is also available.[2]
This matters because supported agent experiences can share an allowance pool.[2] Message counting is therefore a poor mental model for long-running agent work.
7. “Reset Astra every day” would be nice, but that is not the standard system
A daily full refill would be perfect for exploration: burn the research budget today, wake up tomorrow with a clean slate.
That is not how the current standard allowance works. Codex uses displayed windows that can include five-hour and weekly periods.[2]
Eligible Plus and Pro personal accounts may be offered an immediate paid weekly reset. A completed purchase restores the five-hour and weekly usage, but it pulls the normal weekly allowance forward rather than adding a separate bonus week. The new weekly period starts from the first Work or Codex request after the reset.[7]
Some eligible users may also have banked resets; using a full banked reset refreshes both windows and changes the next weekly reset date.[8]
OpenAI Support does not normally manually reset a correctly exhausted allowance on request.[2]
So no, this is not yet “daily login bonus: Astra 100% restored.”
8. Check your Codex CLI version
OpenAI currently requires Codex CLI 0.153.0 or newer for Astra.[1]
If Astra is missing even though your account should have access, client version is one of the first things to verify.
9. Bottom line
Codex can switch models inside a session, but the useful mental model is switch between turns, not hot-swap a running turn.
If Astra allowance is precious, the optimal policy is straightforward:
Sol for daily work. Terra/Luna for cheap repetitive work. Astra for exploration, hypothesis generation, unknown failures, hard architecture and final judgment. Then switch back immediately.
The real skill is not merely knowing how to use the strongest model.
It is knowing where not to use it.
References (8)
- OpenAI Help Center — ChatGPT Work and Codex help.openai.com
- OpenAI Help Center — Using Codex with your ChatGPT plan / ChatGPTプランでCodexを使う help.openai.com
- openai/codex — codex-rs/tui/src/slash_command.rs (/model: choose model and reasoning effort) github.com
- openai/codex Issue #20842 — Switch model mid-prompt github.com
- OpenAI Developers — Models developers.openai.com
- OpenAI Developers — Model guidance / GPT-6 Astra developers.openai.com
- OpenAI Help Center — Paid weekly Work and Codex rate limit resets help.openai.com
- OpenAI Help Center — How banked Codex resets work help.openai.com
