A subscription is not an API wallet
Paying a monthly fee does not establish a fixed token balance or an unlimited number of agent tasks. This guide covers individual plan prices and the Work/Codex allowance, checked on September 30. It is not a review of a private account or a claim that we have personally measured every tier.
ChatGPT Learn lists the following individual monthly USD prices. Your checkout, region, taxes, and billing channel determine the actual invoice.
| Plan | Monthly price |
|---|---|
| Free | $0 |
| Go | $8 |
| Plus | $20 |
| Pro | $100 / $200 / $500 tiers |
The important Pro change
The current pricing documentation says Pro plans have no five-hour limit. Do not copy an older five-hour explanation into a current Pro comparison. That does not mean unlimited usage: use the account's current allowance and reset information. API-key usage has separate token billing.
For Plus, the page provides these estimated local messages per five hours, also applicable to Standard Business seats:
| Model | Estimated local messages / 5h |
|---|---|
| GPT-6.1 Sol | 15-160 |
| GPT-6 Sol | 15-150 |
| GPT-6 Astra | 5-45 |
| GPT-6 Luna | 350-3,000 |
Cloud tasks can consume more; weekly limits may also apply. These ranges cannot be converted into a promised number of articles, pull requests, or research reports.
Check the speed setting and the shared allowance
The speed documentation says Work and Codex share usage. Fast consumes included allowance at 2.5 times Standard, while purchased credits are billed at twice Standard. Astra Ultrafast uses included allowance at eight times Standard and purchased credits at six times, with access on Pro $500 and eligible Enterprise/Edu plans.
Those multipliers describe consumption, not a guarantee that a task finishes that many times faster. A faster response may be valuable during a live debugging session; it may be less valuable for a task you were going to leave running while doing something else.
Model guidance also notes that higher reasoning effort takes more time and uses more tokens. Keep the setting visible in your comparisons. Comparing a light run against a much deeper run tells you about both settings, not just the model names.
Measure your own usable capacity
- Before a work block, record the plan, model, reasoning, speed mode, available allowance, and displayed reset time.
- Choose a bounded task with a written acceptance rule. Avoid mixing unrelated background jobs into the same measurement.
- After it finishes, record remaining allowance, retries, and minutes spent checking or repairing the result.
- Repeat across several days and task types. Report accepted outputs and consumption together, including failed attempts.
Illustration only: if a displayed remaining allowance moves from 80% to 72%, the observed change is eight percentage points for that measurement window. It does not establish eight dollars, a fixed token count, or the next task's consumption. Other simultaneous work or a reset can invalidate the comparison.
Upgrade for an observed bottleneck
Keep a simple log: date, task, accepted or rejected, allowance change, review time, and whether a limit actually stopped work. A higher tier is easier to assess when that log shows which interruption it would remove. A large headline message estimate alone is not evidence of better value.
Use the API pricing article for API budgets and the Claude subscription guide for the other provider's rules. Their allowance systems should be measured separately.