How this calculation works
Measure a comparable block of work, keep the model, effort and speed fixed, and enter its meter changes. Exact profile matching prevents one model’s sample from silently being reused for another. The slowest and fastest observed rates form a scenario range, not a statistical confidence interval.
Observed rate = used percentage-point increase / active minutes; quota time = remaining points / rate; stop before the earliest relevant resetA worked example
Fictional demonstration: 120 active minutes use 100 primary percentage points and 18.5 secondary points. With the primary meter at zero and secondary at twenty, a two-meter calculation is primary-limited to 120 active minutes. A primary reset can occur later than that; waiting time is different from active capacity.
Good to know
Unofficial independent estimator, not affiliated with or endorsed by Codex. Usage can vary with plan, model, reasoning, context, caching, tools, attachments, parallel tasks, server load, promotions and policy changes. Results are conditional projections, not guaranteed maximum hours, requests or access. Account meters and provider notices take priority. Sources checked 9 November 2026.
Reference: OpenAI — Codex pricing and usage guidance (opens in a new tab) ↗
What Codex officially says
OpenAI publishes broad five-hour local-message estimates for Plus and Standard Business, while Pro currently has no five-hour cap. Work and Codex share included usage. Task shape changes consumption; API token prices do not determine subscription capacity. Fast and Ultrafast also change included usage.
Use the quota category and reset time shown in your own account. The model names and effort labels help organize measurements; they do not grant access to a model or assert that every listed setting exists on every plan.
The Claude usage limit calculator covers a separate account or billing model. Compare its own documented scope rather than carrying a quota percentage or hosted rate across providers.
Read more: OpenAI — Codex pricing and usage guidance (opens in a new tab) ↗
What community reports can and cannot tell us
A poster reports 36% weekly use over eight hours while mixing models and subagents. Because there is no controlled per-model token ledger, that is a workload anecdote, not a Sol or Astra burn-rate benchmark. Parallel agents can make wall-clock hours especially misleading.
This linked discussion was checked on 9 November 2026. It is a selected self-reported example, not a representative survey. Different accounts, promotions, workloads and observation windows cannot be combined into a reliable universal hourly rate.
Read more: Community report — mixed-model weekly burn (opens in a new tab) ↗
Measure a profile without confusing its clocks
Record used percentage before and after a short comparable task without crossing a reset. If your meter shows remaining capacity, subtract the later remaining value from the earlier one. A weekly meter from seventeen to twenty percent used represents three percentage points, not twenty.
Measure active work consistently. If another device or agent is using the shared pool, pause it or include its work in the measured scope. Repeating a measurement with the same model, effort and speed gives a useful observed range; one sample gives only one observed pace.
How to reduce Codex usage
Give the agent a narrow file or feature scope, a concrete acceptance check and a stopping point. Request a concise handoff when changing tasks instead of carrying a large unrelated history.
Inspect the usage dashboard or /status before and after a comparable task. Match model, reasoning and speed to the job; lower-cost models can handle mechanical edits while demanding work can justify more effort.
Avoid dumping whole repositories, giant logs or unrelated tool catalogs into every task. Review which integrations are needed, and avoid repeating a full build or wide investigation when a narrower verified check answers the question.
Read the SVG forecast and plan the next block
The usage lines show a constant-pace projection up to the entered next reset. Each activity tile represents an equal block of that active-time interval, not a calendar day or imported account history. Blocks, grouped and cumulative views reorganize the same scenario; equal intensity reflects its constant-rate assumption.
A cautious pre-reset estimate uses the fastest matching sample and whichever tracked pool runs out first. The slower-sample endpoint is an optimistic scenario, not an upper bound guaranteed by the provider. If the next task is longer, has more files or changes effort, take a new sample.
Independent estimate disclaimer
Unofficial independent estimator, not affiliated with or endorsed by Codex. Usage can vary with plan, model, reasoning, context, caching, tools, attachments, parallel tasks, server load, promotions and policy changes. Results are conditional projections, not guaranteed maximum hours, requests or access. Account meters and provider notices take priority. Sources checked 9 November 2026.
The calculation runs on your device and does not connect to your AI account. Demo observations and request caps are fictional. A zero or rounded meter change cannot establish unlimited usage; a reset or decrease during measurement requires a fresh sample.
A few questions, answered
Does selecting Ultra apply a fixed quota multiplier?
No. Where your client offers Ultra, record a separate matching sample. Ultra is not the same as Ultrafast speed, and its implementation can involve multiple agents. No universal multiplier has been established here.
Why can a five-hour limit end after two hours?
The window governs when allowance resets. It is not a guarantee of five active hours. A workload can consume its allowance before the window ends, while weekly or other feature limits can bind independently.