ChatGPT Work + Codex Usage Limits

ChatGPT Work and Codex usage limits now belong in the same planning conversation. OpenAI says Codex, ChatGPT Work, its spreadsheet agent, and Workspace Agents can draw from the same agentic usage and credit pool when those features are available on your plan. For founders and operators, the practical answer is simple: use ChatGPT Work for broad business tasks, use Codex for technical execution, and treat long-running agent work like a budgeted resource.

Quick verdict

If you are trying to do this	Start here	Watch this limit
Research, summarize, draft, or turn scattered work into a finished deliverable	ChatGPT Work	Agentic usage allowance
Fix code, review a repo, run tests, or create a pull request	Codex	Codex task credits and model choice
Clean data, build a small internal tool, automate a workflow, or create a technical report	Codex inside ChatGPT Work	Task size and session length
Run a serious codebase with local files, scripts, tests, and memory	Claude Code	Your Anthropic plan and local execution setup
Build an AI operating system for recurring work	Second Brain AI plus the right agent	The workflow design, not only the plan limit

The mistake is treating limits as a pricing footnote. Limits shape what you delegate, how you split work, and whether the result is worth the spend.

What changed with ChatGPT Work and Codex

OpenAI launched ChatGPT Work on July 9, 2026 as a broader work agent. The pitch is no longer just chat, search, or file analysis. ChatGPT Work can act across apps and files, stay with a project for longer stretches, and move a goal toward finished work.

Codex now sits inside that larger story. OpenAI's help docs say Codex usage can share the same agentic usage and credit pool as ChatGPT Work and other agent features. OpenAI's product writing also shows Codex moving beyond software teams into reports, spreadsheets, presentations, contracts, workflow automation, and lightweight tools.

That matters because the search intent is changing. People will ask:

Is Codex part of ChatGPT now?
Does ChatGPT Work use my Codex limits?
How many Codex tasks do I get?
Should I use ChatGPT Work or Codex for this job?
When should I stop using OpenAI and move deep repo work to Claude Code?

Those are buying questions, not trivia questions.

The simple model

Think about the work in three layers.

Layer	What it means	Best tool
Business surface	The work involves documents, meetings, files, research, browser context, and team apps	ChatGPT Work
Technical execution	The work needs code, repo access, scripts, tests, or generated tools	Codex
Operating system	The work needs persistent memory, repeatable workflows, local files, and review loops	Claude Code or Second Brain AI

Most founders will use all three. ChatGPT Work becomes the front door. Codex handles technical tasks inside that front door. Claude Code or a structured Second Brain handles deeper systems that need memory and repeatability.

How Codex usage is counted

OpenAI gives the clearest operational answer in its help docs: Codex usage depends on the plan, the model, the size of the task, the complexity of the work, and where the task runs.

A small script can use a small part of your allowance. A larger repo task, long session, or complex model run can use more. OpenAI's Codex rate card says a typical Codex task using GPT-5.5 may consume a range of credits, and fast mode can consume credits at a higher rate for supported models.

That means you should not plan Codex like a fixed number of messages.

Plan it like this:

Define the outcome.
Split the work into verifiable tasks.
Use cheaper or smaller runs for exploration.
Save expensive runs for execution.
Stop the agent when the next step needs human judgment.

This is the same way good operators manage contractors. You do not pay someone to wander. You pay them to finish a scoped task.

ChatGPT Work vs Codex: which one uses the budget better?

Use ChatGPT Work when the work is messy and cross-functional.

Good examples:

Turn meeting notes, docs, and Slack context into a decision memo.
Create a first draft of a board update.
Compare vendors from websites, PDFs, and pricing pages.
Build a weekly operating brief from connected sources.
Convert scattered research into a clear plan.

Use Codex when the work has a technical object.

Good examples:

Fix a bug in a repo.
Create a script.
Add a small internal tool.
Clean a dataset.
Review code.
Turn a workflow into automation.
Generate a spreadsheet model or technical report.

The boundary is the artifact. If the output is a memo, deck, doc, or operating plan, start in ChatGPT Work. If the output is code, a tool, data cleanup, or a technical artifact, use Codex.

The founder workflow I would use

Start every agent task with this five-line brief.

Goal:
Inputs:
Output format:
Definition of done:
Stop conditions:

Example:

Goal: Create a weekly customer insight memo from support tickets.
Inputs: CSV export, product changelog, last week's roadmap notes.
Output format: 1-page memo with themes, quotes, risks, and next actions.
Definition of done: I can send it to the team without rewriting the structure.
Stop conditions: Stop if the data is missing columns or if customer names appear in raw quotes.

This saves usage because the agent does less guessing.

Then route the work:

Step	Tool	Why
Frame the outcome	ChatGPT Work	It can reason across the business context
Create or clean technical artifacts	Codex	It can execute code-like work
Store the repeatable workflow	Second Brain AI	The same task should improve over time
Run deep repo changes	Claude Code	Local tests, files, and memory matter

The real gain comes when the task becomes reusable. A one-off Codex run saves time once. A stored workflow saves time every week.

How to avoid wasting Codex credits

Most wasted usage comes from vague prompts and oversized tasks.

Use these rules:

Ask for a plan before execution when the task is large.
Give the agent the exact files, URLs, or data source.
Require a short diff summary before any broad change.
Ask for tests or verification steps in the same task.
Split exploration from implementation.
Stop after a useful checkpoint.

Bad prompt:

Improve our onboarding.

Better prompt:

Read the onboarding checklist and the last 20 support tickets. Find the 5 most common setup blockers. Return a table with blocker, evidence, fix, and owner. Do not edit files yet.

That second prompt is cheaper because it narrows the job.

When to upgrade, buy credits, or change tools

Upgrade only when the limit blocks a recurring workflow that already saves real time.

Use this test:

Signal	What to do
You hit limits while experimenting	Split tasks better first
You hit limits on one-off curiosity	Do nothing
You hit limits on weekly work that saves hours	Upgrade or buy credits
You hit limits on repo work with lots of local context	Move that loop to Claude Code
A team member burns credits with vague tasks	Fix the workflow brief before buying more

Credits are useful when the workflow is already proven. They are a bad fix for unclear delegation.

Where Second Brain fits

ChatGPT Work and Codex can execute. They do not automatically give your company a memory system.

That is where the Second Brain layer matters.

The compounding system is:

Capture the work pattern.
Turn it into a repeatable workflow.
Store prompts, inputs, outputs, decisions, and review notes.
Link the workflow to the right agent.
Improve the workflow each time it runs.

If your team is only buying more agent credits, you are renting intelligence. If you capture the workflow, you are building operational memory.

For that layer, start with the AI Chief of Staff workflow. If you want the full operating system around your agents, use Second Brain AI.

FAQ

Is Codex now part of ChatGPT?

Yes, in practical usage terms. OpenAI says Codex is included across ChatGPT plans, and Codex usage can draw from the same agentic usage and credit pool as ChatGPT Work when the features are available on your plan. Codex still has its own surfaces for technical work.

Does ChatGPT Work use the same limits as Codex?

OpenAI says Codex, ChatGPT Work, its spreadsheet agent, and Workspace Agents can draw from the same agentic usage allowance for Plus and Pro users when supported on the plan. Business and Enterprise workspaces can also involve credits and spend controls.

How many Codex tasks do I get?

There is no single task count that fits every plan and task. Usage depends on the model, task size, complexity, session length, and where Codex runs. Small tasks use less. Long-running work and large repos use more.

Is Codex worth it for non-developers?

Yes, when the work has a technical artifact. OpenAI shows knowledge workers using Codex for reports, spreadsheets, presentations, contracts, automation, data analysis, and lightweight tools. If the work is only writing or synthesis, start with ChatGPT Work.

Should founders use ChatGPT Work, Codex, or Claude Code?

Use ChatGPT Work for broad business delegation. Use Codex for OpenAI-native technical execution. Use Claude Code when the repo, terminal, tests, hooks, memory files, and local workflow are central to the job.

What is the safest way to manage usage?

Scope every task before execution. Define the goal, inputs, output format, definition of done, and stop conditions. Then split large work into smaller checkpoints.

Official sources checked

Bottom line

ChatGPT Work is the broad work surface. Codex is the technical execution layer. Claude Code remains the stronger deep repo system.

The founder move is to stop treating agent limits as an annoyance and start treating them as an operating design constraint.

Use expensive agent work for scoped execution. Use cheaper thinking for planning. Store the workflows that repeat. That is how you turn usage into compounding advantage instead of another SaaS bill.