Codex has no price per task, only a rolling 5-hour window and a weekly cap, and the same Plus plan gives you anywhere from 15 to 160 GPT-6.1 Sol messages per window. OpenAI sells no separate Codex subscription. It is included in ChatGPT plans from $0 to $500 a month, so what you buy is an allowance, and what you plan is how fast you burn it.
Short answer: a solo developer who codes with Codex a few hours a day fits on Plus at $20 or Pro at $100, and moves to Pro 500 only if Ultrafast speed is worth burning the allowance eight times faster. A five-person team is cheaper per head on Business at $20 a user billed annually than on five Pro 100 seats, but the allowance per seat is the Plus-sized one. The API-key path is billed per token, has no window, and skips cloud features such as GitHub review and Slack. The numbers below come from OpenAI's help pages and October coverage; the planner math is mine and labelled as derived.
A rate limit is not a rate
Per-token APIs price what you use. Codex plans price access to a pool. Since April 2026 OpenAI meters Codex as token credits against a rolling 5-hour window plus a weekly cap. The 5-hour window is the one you feel: you hit it mid-afternoon, wait, and carry on. The weekly cap is the one that ends your Thursday.
The message counts people quote are estimates, not guarantees. OpenAI publishes ranges per model, and the range is wide because a message that reads one file costs far less than a message that rewrites a module. Cloud tasks are the sharper edge: OpenAI says cloud chats may use more allowance than local messages, and you cannot change the default model for Codex cloud chats. I could not find a published weekly cap figure in my sources, so I do not give one.
When you run past the window you can buy credits. The going estimate is about $0.04 per credit. That makes overflow predictable in dollars even when the messages-per-credit rate is not, and it sets a ceiling on how much a bad day costs you. Credits bought on Pro 100 or Pro 200 do not unlock the Pro 500 Ultrafast model.
The plans and what a window buys
| Plan | Price | Allowance relative to Plus | Notes |
|---|
| Free | $0 | Smallest | Entry access |
| Go | $8 a month | Small | Entry paid tier |
| Plus | $20 a month | 1x | Cloud floor; Linear and Slack integrations need a paid plan |
| Pro 100 | $100 a month | About 5x | Pro is quoted as 5x Plus |
| Pro 200 | $200 a month | Reported 10x from 30 Oct 2026 | Reopened; sources disagree on the date, one says grandfathered to 29 Oct |
| Pro 500 | $500 a month | 25x | Includes GPT-6 Astra Ultrafast: 8x faster token generation, 8x allowance burn |
| Business | $20 per user a month annual, $25 monthly | Plus-sized per seat | Two or more users |
October guidance puts the Plus and standard Business 5-hour window at these message ranges:
| Model | Messages per 5-hour window (Plus) | Derived: Pro 100 (5x) |
|---|
| GPT-6 Astra | 5 to 45 | 25 to 225 |
| GPT-6.1 Sol | 15 to 160 | 75 to 800 |
| GPT-6 Sol | 15 to 150 | 75 to 750 |
| GPT-6 Luna | 350 to 3,000 | 1,750 to 15,000 |
The Pro 100 column is my multiplication of OpenAI's stated 5x, not a published figure. Treat it as a planning number. Older Sol, Terra and Luna generation estimates were 10 to 100, 25 to 200 and 250 to 2,000 for Plus, which shows how much these ranges move when the models change. Re-read the help page each quarter.
Pro 500 and Ultrafast: speed you pay for twice
Pro 500 costs $500 a month, gives 25 times the Plus allowance, and includes the Astra Ultrafast model. Ultrafast generates Codex tokens eight times faster and burns allowance at eight times the rate. Divide the 25x pool by the 8x burn and you get the equivalent of roughly 3x Plus when you run Ultrafast flat out. On GPT-6.1 Sol ranges that is about 47 to 500 messages per window (375 to 4,000 divided by eight), again my derivation.
Read that against the price. Pro 100 at 5x costs a fifth of Pro 500. If you run Ultrafast all day, Pro 500 gives you less raw allowance per dollar than Pro 100 gives you on standard speed. The product is for people whose bottleneck is waiting for tokens, not running out of them: a founder in a live debugging session, a consultant on an hourly rate where eight times faster pays for itself. If your bottleneck is thinking time, spend $100.
The API-key path for the same model family is different again. GPT-6.1 Sol lists at $2.00 per million input tokens, $0.10 cached input and $10.00 per million output. A task that sends 200K input tokens (160K of them cached) and returns 20K output costs about $0.30 on those prices: $0.08 uncached input, $0.016 cached input and $0.20 output. Those token counts are my assumption. At roughly thirty cents a task, a $20 Plus plan equals about 67 such tasks by API, and the plan gives you a recurring window instead. The API path also has no 5-hour wall, and it excludes the cloud features: GitHub review and Slack are plan-only.
Planner: a solo developer and a five-person team
Pick by how many windows you exhaust, not by how many hours you code. A good habit: for one week, note every time Codex tells you the window is spent. Zero times means you are on the right plan or an oversized one. Twice a week means you are near the edge. Daily means move up a tier.
| Profile | Plan | Monthly cost | Reasoning |
|---|
| Solo, part-time, local edits | Plus | $20 | Cloud floor; Luna-class models stretch to 350 to 3,000 messages a window |
| Solo, full-time, heavy cloud tasks | Pro 100 | $100 | 5x pool for 5x the price; one account, one window |
| Solo, latency-bound, billable by the hour | Pro 500 | $500 | Only if Ultrafast speed pays for itself |
| Solo with a pipeline script | API key | Metered | No window, no cloud features |
| Team of five, mixed use | Business, annual | $100 ($20 x 5) | Plus-sized allowance per seat; $125 billed monthly |
| Team of five, one heavy user | Business for four, plus a personal Pro 100 | $180 | Heavy user gets 5x; others stay at seat price |
The five-person Business plan costs the same $100 as one Pro 100 seat, and delivers five separate Plus-sized windows. If your usage is spread evenly, five windows beat one window five times the size, because the bursts do not collide. If one person uses 80% of the team's Codex time, the pooled seat does not exist and that person hits their own wall. Fix it with a Pro seat for them, not with a bigger plan for everyone.
Model your own mix on the Agent Run Cost Simulator, new today, and price API-key work on the AI Prompt Cost Calculator. For the cheaper end of the model market, the Haiku 5.5 pricing breakdown shows what a per-token budget looks like at $0.10 and $0.50.
Three signals say you are on the wrong plan. If you hit the 5-hour window more than twice a week, the window is your bottleneck and a higher tier or credits pay back. If you never hit it and your weekly usage sits well under what you expect the cap to be, you are paying for headroom you do not use, and you can drop a tier. If your overflow credits regularly cost more than the gap to the next plan, upgrade. At about $0.04 a credit, the arithmetic is easy to run once you know how many credits a typical heavy day consumes, and your usage page shows that number.
The verdict for most readers of this blog: start on Plus, track window hits for two weeks, and move to Pro 100 when you hit it twice a week. Skip Pro 500 unless you can name the hourly value of eight times faster. Skip the API-key path unless you run Codex from scripts and do not need the cloud features. Revisit in November, because OpenAI is still changing the numbers: the Pro 200 allowance change reported for 30 October is only the latest example.
Burning the window slower
You cannot raise the allowance, but you can lower the burn per task. Three habits matter more than the plan tier.
Keep instructions short and stable. Every message carries your standing instructions. A bloated AGENTS.md or CLAUDE.md-style file costs every call, and a stable prefix is what caching rewards. The CLAUDE.md Starter Templates Pack shows tight versions of those files that also carry over to Codex-style agents.
Scope tasks before you send them. A vague request makes the agent explore. A request that names the file, the function and the test it must pass costs a fraction. Cloud tasks suffer most from vagueness, because they spin up an environment each time.
Watch spend as an operations metric. The AI Agent Ops Bundle includes the logging and budget-guard patterns that turn "I hit the limit again" into a chart. When the chart says you hit the window on Tuesdays after a particular kind of task, you change the task, not the plan.
Quick answers
How much does Codex cost?
There is no separate subscription. Codex is included in ChatGPT plans: Free $0, Go $8, Plus $20, Pro from $100 and Business at $20 per user a month billed annually or $25 monthly. The API-key path bills per token instead.
How many Codex messages do I get in 5 hours?
On Plus, OpenAI's October guidance gives 15 to 160 for GPT-6.1 Sol, 15 to 150 for GPT-6 Sol, 5 to 45 for GPT-6 Astra and 350 to 3,000 for GPT-6 Luna. Pro is about five times that.
What does Pro 500 add?
$500 a month buys 25 times the Plus allowance and access to Astra Ultrafast, which generates tokens eight times faster and burns allowance eight times faster. Credits bought on Pro 100 or 200 do not unlock it.
What do extra credits cost?
Overflow credits cost roughly $0.04 each. Cloud chats may use more allowance than local messages, so the credits per task vary.
Every product mentioned is available at wowhow.cloud — pay once, ship forever.
Originally published at wowhow.cloud
Top comments (0)