Quota is a range.
OpenAI publishes estimated local messages per five hours. We convert the endpoints with 100 ÷ messages. These are indicative model-level ranges, not per-task measurements or statistical confidence intervals.
THE MODEL EXPLORER
More clarity. Less guesswork.
Compare AI capability and usage before you choose a model.
Sources checked
SELECTED CONFIGURATION
OpenAI · GPT-6 series
Output tokens / benchmark task
4.4ktokens
938 reasoning + 3.5k answer tokens
Reasoning effort
Measured on Artificial Analysis benchmarks. Output includes reasoning. Tokens are not subscription quota. How we calculate this
Capability vs. token use
Higher is more capable. Further left uses less.
Swipe to explore the chart ↔
Output tokens per benchmark task · includes reasoning →
The dotted frontier connects configurations with no higher-scoring alternative using fewer output tokens. This is token efficiency on these benchmarks, not a quota recommendation.
Scroll the table to see every metric ↔
| Model & reasoning | Tokens / taskMeasured output | IntelligenceAA Index | Reasoning tokensBenchmark | Compare configuration |
|---|---|---|---|---|
2.1k | 20.9 | 515 | ||
2.5k | 21 | 559 | ||
3.4k | 33.9 | 697 | ||
3.9k | 27.5 | 1k | ||
4k | 33.5 | 837 | ||
4.4k | 45.8 | 938 | ||
4.5k | 25 | 1.5k | ||
5.5k | 30.1 | 1.9k | ||
6.5k | 39.8 | 2.4k | ||
7.9k | 39.2 | 2.7k | ||
9.6k | 49.6 | 3.4k | ||
10.2k | 42.8 | 5k | ||
11.2k | 29.5 | 6.2k | ||
11.3k | 34.2 | 5.4k | ||
11.8k | 50.9 | 4.7k | ||
13.3k | 42.3 | 5.9k | ||
13.9k | 32.1 | 7.2k | ||
16k | 44.1 | 9.3k | ||
16.9k | 52.4 | 8.6k | ||
19.6k | 44 | 10.1k | ||
19.8k | 32.1 | 12.8k | ||
20.4k | 38 | 11.5k | ||
23.6k | 34.6 | 14.7k | ||
27.2k | 33.9 | 18.7k | ||
27.2k | 52.7 | 16.7k | ||
29.3k | 47 | 17.3k | ||
31.2k | 47.5 | 21k | ||
38.9k | 42.1 | 26k | ||
41.2k | 37.3 | 28.1k | ||
50.5k | 37.3 | 39.3k |
Quota ranges repeat across reasoning levels because effort-specific rates are unpublished.Download data
YOUR ALLOWANCE
Working with a partly used quota?
Scale the published message range to the quota you have left.
UNDERSTAND THE NUMBERS
A comparison you can inspect.
Three different measurements, with their boundaries kept visible.
Capability is benchmarked.
Scores come from the Artificial Analysis Intelligence Index v4.3.2. Higher scores mean stronger performance across its evaluation mix. A score twice as high does not mean twice the intelligence.
Your work will vary.
Context, caching, tools, reasoning and subagents change consumption. Benchmark tasks differ from your projects. API prices and purchased-credit promotions do not determine your included quota.
What’s included, and what isn’t +
This release covers six OpenAI models used in Codex, five benchmarked reasoning levels, and Plus, Pro 5× and Pro 20×. Only configurations with comparable, sourced benchmarks appear here.
The quota view covers local Codex messages at Standard speed, including local CLI and IDE use with ChatGPT authentication. It does not estimate cloud tasks, Fast mode, API-key spending or weekly quota. A local message may start a multi-step task; it is not a fixed unit of work.
The same family range is intentionally shown for every reasoning level. Changing effort updates the measured capability and token values, but cannot manufacture an effort-specific quota estimate. Ultra and non-reasoning modes are excluded from this comparison because they are not part of this set of five comparable benchmark levels.
Quota dots use the geometric midpoint purely to position ranges on a logarithmic chart. The selected horizontal line shows the full published range. Overlapping ranges do not establish which configuration uses less quota for the same task. The token frontier is computed only from the visible benchmark configurations.
Scores are rounded to one decimal, token counts to whole tokens. Sources are checked before a dated snapshot is published. After 14 days, the page labels the snapshot as needing review. These data are not live account telemetry.
All controls run in your browser. No login, tracking cookies, API key or account data is required. Share links contain the model, comparison, plan and display filters only.
TRACEABLE BY DESIGN
Go straight to the source.
Snapshot reviewed 26 Sept 2026.