There is no single Codex message limit that applies to everyone. OpenAI currently publishes model-specific ranges for several paid plans, but your actual consumption also changes with repository size, context, reasoning, tools, caching, and whether work runs locally or in the cloud.
As of August 16, 2026, the most reliable operating rule is simple: use the public table to budget, then check your authenticated Codex Usage dashboard before handing over work that cannot pause. In an active Codex CLI session, /status shows remaining limits.
The public ranges are not your account balance
The current OpenAI Codex pricing documentation lists these estimates for local messages in a five-hour window:
| Model | Plus / Business | Pro 5x | Pro 20x |
|---|---|---|---|
| GPT-5.6 Sol | 10–100 | 50–500 | 200–2,000 |
| GPT-5.6 Terra | 25–200 | 125–1,000 | 500–4,000 |
| GPT-5.6 Luna | 250–2,000 | 1,250–10,000 | 5,000–40,000 |
Those wide ranges are intentional. A focused edit in one file might consume a fraction of the allowance used by a repository-wide migration that reads many files, keeps a long conversation in context, runs tools, and produces detailed output. Prompt length alone does not predict the cost.
The same page says local messages and cloud chats share a five-hour window where the table applies, and that additional weekly limits may apply. A newly available five-hour window therefore does not guarantee that a weekly bucket has also recovered. It also currently marks the cloud-chat and code-review columns unavailable for the GPT-5.6 rows above, so the local-message values should not be copied into a promise for every Codex surface.
Free and Go are included Codex routes, according to OpenAI's current ChatGPT-plan guidance, but the current pricing table does not provide a durable Free- or Go-specific message range. If your account shows a banner or limit, that account UI is stronger evidence than an old screenshot or forum count. Readers choosing Go specifically can use the separate ChatGPT Go Codex usage guide.
Find the meter before choosing a fix
“Codex limit” can refer to three different contracts. Treating them as one is the fastest way to buy the wrong plan or debug the wrong system.
| How Codex is authenticated | Meter owner | Evidence to check | What a limit means |
|---|---|---|---|
| ChatGPT personal plan | The signed-in account and its agentic usage | Codex Settings → Usage; CLI /status | Included plan capacity, a reset window, or credits where offered |
| Business, Enterprise, or Edu workspace | The seat, workspace policy, or credit pool | Workspace Usage plus owner/admin controls | A personal setting may not be able to change capacity |
| API key | API organization and project | API usage, billing, project limits, response data | Token billing and API rate limits, not a ChatGPT-plan refill |
OpenAI says Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents draw from the same agentic usage and credit pool when those features are available. That can explain why Codex capacity changed even when your recent work happened in another agentic surface. It does not mean every ordinary ChatGPT cap is the same meter.
If the symptom is that a second account appears to inherit the first account's limit, do not assume a device-wide pool. Check the active account, workspace, API key, and local client state using the cross-account Codex limit checklist. Rotating accounts to keep working is not a sound limit-management strategy.
Why one task can cost much more than another
Most current customers now use a token-based credit rate card. Credits are a consumption unit, not a fixed number of messages. OpenAI's current Codex rate card prices input, cached input, and output tokens separately. For GPT-5.6 Sol the current rates are 125, 12.5, and 750 credits per million tokens; Terra is 50, 5, and 300; Luna is 5, 0.5, and 30.
OpenAI describes a typical GPT-5.6 task as using roughly 5–40 credits, but that estimate is not a conversion table. A task can land outside the typical shape because of its context, token mix, tools, agents, reasoning, or output. A few practical signals matter more than message count:
- a bounded task with a small file set and short acceptance test is easier to budget;
- a long session keeps more history and retrieved material in play;
- broad repository exploration and repeated tool output increase the working context;
- Fast mode consumes credits at a higher rate on supported models;
- image generations currently use included limits about three to five times faster on average than comparable non-image turns, depending on image size and quality.
Model choice changes the slope. Sol is for the hardest ambiguous work, Terra is the general production workhorse, and Luna is optimized for fast, focused, high-volume tasks. A lighter model is not a substitute for judgment, but it is often the right choice for routing, extraction, classification, repetitive edits, or narrow verification.
Check capacity before the expensive part of the task
Open the Usage dashboard and record the active plan or workspace, model, remaining five-hour capacity, any weekly limit, and the displayed reset time. If you are in the CLI, run /status in the same authenticated session. If the browser and CLI appear to disagree, confirm that they use the same account and workspace before drawing a conclusion.
Then give the task a checkpoint that still has value if Codex stops. For a migration, the first outcome might be a verified dependency map rather than a full implementation. For a refactor, finish one package and its tests before expanding. For an investigation, save confirmed observations separately from hypotheses.
This is not merely prompt hygiene. A recoverable task boundary prevents a limit from turning unfinished context into lost work. It also makes model selection easier: use the expensive reasoning route only where ambiguity or risk justifies it, and move clearly bounded work to a lower-consumption route.
What to do after a limit appears
OpenAI's current limit guidance supports several branches, but they are plan-dependent.
Wait for recovery when the task is not urgent and its state is safely recorded. Use the reset time shown in your account rather than assuming a universal clock.
Reduce the task or choose a smaller model when capacity remains but the current pace is too expensive. Remove irrelevant context, narrow the file scope, and ask for a result that can be validated independently.
Buy additional credits only when the account or workspace offers that option. Plus and Pro users who exhaust included usage can purchase credits where supported. Flexible-pricing Business, Edu, and Enterprise workspaces can purchase workspace credits; a member may need an owner or admin to act.
Use an API key deliberately when the job fits local CLI, SDK, or IDE work and you accept API token billing. OpenAI's pricing page says the API-key route does not include all cloud-based integrations available through a ChatGPT plan. It is a different product contract, not a hidden extension of a Plus allowance.
Reset news is not the normal limit contract
OpenAI added rate-limit reset banking for eligible Plus and Pro users in June 2026 and has run time-limited invitation rewards and account-specific reset events. Those announcements do not prove that every account has a free reset today. Eligibility, expiration, and the current promotion are visible in the account surface when they apply.
That distinction matters whenever a headline says limits were “removed” or “reset.” A mass adjustment can be real without becoming the permanent five-hour or weekly rule. Check the current dashboard and current official documentation before planning critical work around a reset.
One near-term change is already dated. The current rate card says GPT-5.4 and GPT-5.4 mini will retire from ChatGPT-authenticated Codex on August 31, 2026, with GPT-5.6 Terra and GPT-5.6 Luna as the respective migration paths. OpenAI says API access and Codex use with your own API key are not affected by that retirement. If an old capacity plan depends on GPT-5.4, recheck the model list and Usage dashboard before the date.
The useful answer to a Codex limit is therefore not one quota number. It is a matched set of evidence: the right billing surface, the current account meter, a task boundary that can survive interruption, and a continuation path your plan actually supports.
