ChatGPT chat messages and Codex repository tasks are different units, even on the same account. Official docs also state that ChatGPT Work shares usage with Codex. If you dump every investigation into Codex, the agent budget for diffs, tests, and pull requests runs out first.
📑Table of Contents
Keep explanatory work in chat. Send only tasks that must change the repository to Codex. OpenAI Help and the Learn pricing page clearly separate Chat, Work, and Codex, and they document the shared ceiling. Independent write-ups make the same operational distinction: plan in chat, delegate the diff.
What this article will decide
This article lines up four things: the official surface split, the shared five-hour window and plan ceilings, the GitHub @codex review path, and a go / no-go table for the next Codex task. After that you can decide what stays in Chat, when to delegate on GitHub, and what a human still checks before merge.
Split Chat, Work, and Codex with the official definitions
If you need an explanation, use Chat. If you need a finished artifact, use Work. If you need the repository to change, use Codex. The starting surface determines both quota and permissions.
What official Help assigns to Chat, Work, and Codex
OpenAI Help: ChatGPT Work and Codex (retrieved 18 August 2026) draws that line.
- Chat is questions, search, and short drafts. It does not assume machine-side edits.
- Work is long-running outcome work: documents, tables, decks, reports, and Sites. Work on web and mobile runs in the cloud; the desktop app can also use local files.
- Codex is code, debugging, tests, diff review, and repository work. On desktop it stays a separate view. You cannot select it on web or mobile.
The mobile Remote tab can reach a desktop Codex chat, but that history does not appear in web or mobile history.
How the Learn index separates the surfaces
The ChatGPT Learn index lists these as separate entries:
- Get started with Work
- Skills & Plugins
- Codex CLI / IDE / cloud
The product UI shows GPT-5.6 Sol Extra High. Official What’s new notes also attach education plugins and Computer History to both Work and Codex. Treat those as agent-family surfaces, not chat-only extras.
How independent sources cut the same units
Independent sources use the same distinction.
- Aivy (Monika Tantau, verified 10 August 2026): contrasts message-unit chat with repository tasks. If analysis must not change a machine, stay in ChatGPT. If the job is read-repo / write-diff / run-tests, use Codex.
- LatchLoop: Chat, Work, and Codex as one family. Work covers apps, files, and plugins; Codex covers GitHub, worktrees, diffs, and tests.
Codex as a delegate, not a builder
- Emplex: Codex is a delegate, not a sidecar autocomplete tool, and contrasts it with from-scratch builders such as Lovable or Bolt.
History and permission limits that do not sync
The official limits matter.
- Cloud Work syncs across web, mobile, and desktop.
- A locally started chat stays on that device.
- Codex history is not ChatGPT history.
- Workspace controls for Work Cloud, Work Local, and Codex Local are separate.
- The default model for Chat is not the default model for Work and Codex.
For the desktop merge of Work and Codex, see ChatGPT Workとは？Codex統合デスクトップで変わる長時間エージェント運用.
The reader decision is simple. Leave explanations in Chat. Put finished collateral in Work. Change source code only in Codex. Mixed threads make both history and permissions hard to audit later.
Keep Codex budget by splitting chat quota from agent quota
Work artifacts and Codex pull requests draw from the same agent pool. Finish the plan in chat, then hand over a scoped task. That approach keeps implementation budget available.
Official shared usage for Work and Codex
Help documentation says Work follows the same usage structure as Codex, and Voice-started tasks consume the same agentic pool.
The Learn pricing page (retrieved 18 August 2026) states it in one sentence: “ChatGPT Work and Codex share usage.”
Five-hour window and plan ceilings
The personal shape on that day is:
- Free $0
- Go $8
- Plus $20
- Pro from $100 (5x or 20x versus Plus)
Plus uses a five-hour window shared by local messages and cloud chats. The documented ranges are:
- GPT-5.6 Sol 10–100
- Terra 25–200
- Luna 250–2,000
- possible extra weekly caps
Similar tasks still vary by model, context, reasoning, and tool use. Do not treat a fixed message count as a guarantee.
Independent write-ups on the cap
Independent write-ups treat the ceiling as an operations problem.
- daily.dev (Carlos Mendoza, 8 August 2026): Codex Plus is bundled with ChatGPT Plus, but the five-hour cap can stop a long session.
- daily.dev recommends Pro for daily agent work and treats Codex as an async implementation / PR worker, not a replacement for every pre-code step.
- Aivy: every plan has usage limits. Limited on Free / Go, Expanded on Plus, Maximum tasks on Pro.
Dollar tables are snapshots. Aivy’s numbers do not match Learn exactly, so re-open the official page before you change a plan.
Stage split from the public Japanese thread
A public Japanese thread by スナガク (@suna_gaku, 17 August 2026) splits the same work by stage. Chat spends ChatGPT model limits. Work and Codex spend the agent quota.
- Research, plans, issue text, and mid-PR checks stay in Chat plus the GitHub plugin.
- Code edits and final review stay on Codex.
- Large-repo investigation on Codex burns input tokens that you wanted for implementation.
- A finished ChatGPT thread can be attached from the Codex
@menu as ChatGPT conversation context. - Name
owner/repositoryor the URL, not onlyPR #123, so the agent does not touch the wrong repo.
Three-step cut before you open Codex
Use this three-step cut:
- Do not send a large repository survey to Codex.
- Keep the plan in Chat until the change set is written down.
- Hand Codex a brief that names the scope, the off-limits paths, and the verification method.
If search cost inside a large repo is the next bottleneck, read Codexのコード探索をBM25前段にしてトークンを抑える設計. The practical judgment in this section is simple: do not spend the agent pool on slides or unscoped exploration.
Quota comparison across Chat, Work, and Codex
| Surface | Main unit | Quota it spends | Good fit | Poor fit |
|---|---|---|---|---|
| Chat | Messages | ChatGPT model limits | Research, drafts, issue text, option comparison | Direct source edits |
| Work | Outcome tasks | Same agent pool as Codex | Docs, tables, recurring updates, plugin work | Burning implementation quota on collateral |
| Codex | Repository jobs | Same agent pool (5-hour window) | Diffs, tests, PRs, @codex review |
Unscoped mega-investigations |
Sources for the quota split
Sources:
- OpenAI Help
- Learn pricing
- Aivy
- daily.dev
- スナガク (as of August 2026)
Delegate on GitHub, then keep a human review boundary
Set cloud access and review permissions before you mention a task in a pull request. One comment is not a blank cheque.
Learn’s GitHub documentation (retrieved 18 August 2026) lists this order:
- Configure Codex cloud on the target repository.
- Turn on Code review. You need GitHub push or admin rights.
- For a manual request, comment
@codex review. After the 👀 reaction, Codex posts a review focused on P0 / P1.
Automatic reviews and follow-up prompts
- If Automatic reviews is on, newly review-requested pull requests get a review without an extra comment.
- Ask for a fix with
@codex fix the P1 issue. With permission, Codex may push to the branch. - Any other
@codex …prompt is a cloud chat that uses the pull request as context, for example@codex fix the CI failures.
Repository guidance and review scope
Put repository-specific guidance in a nearby AGENTS.md under ## Code Review Rules.
- You can keep both a root file and a file closer to the change.
- Leave mechanical lint in CI.
- Narrow the pass with a phrase such as
@codex review for issues in the database migration. - Security Review is still a research preview via
@codex security review. - Findings can overlap ordinary code review.
Keep a human review boundary
Official docs say review rules do not replace tests, branch protection, or required approvals.
- Emplex makes the same operational point: write a scoped brief, then read the diff before merge.
- A misunderstood intent grows during long autonomous runs.
- GenAI PM’s 16 August 2026 wiki also tracks GitHub review bots as part of the Work / Codex execution stack, including post-CI risk scoring. That is not extra chat quota.
Minimum brief before you delegate
Write at least these four lines before you delegate:
- Paths in scope and paths that must not change
- The test or reproduction steps you expect
- Human checks before merge (permissions, secrets, breaking migrations)
- The target repository as
owner/nameor a URL
Do not treat a review comment as the required approval. Official GitHub comments flag P0 / P1 only. If you also want a separate vulnerability scan path, see OpenAI Codex Security — 脆弱性スキャン CLI / TypeScript SDK を GitHub 公開（Apache-2.0）.
Decide whether this task should go to Codex now
If you cannot write the change set, do not send the task to Codex. The pass condition is a named diff plus a named verification method.
Use this checklist. Any empty box sends the work back to Chat.
- The desired output is a repository diff or pull request, not more conversation
- You can describe in-scope and off-limits paths in one paragraph
- You have a verification method: test, reproduction, or a visual check
- Large research and option comparison already finished in Chat
- Work collateral has not already drained the same agent pool
- You can name the target repository as owner / name or a URL
- A human already knows what to inspect before merge
Typical Chat, Work, and Codex splits
Typical splits:
- API design comparison stays in Chat
- A slide deck or weekly table is Work
- A failing test, a local type-error fix, or
@codex reviewon a review-requested pull request is Codex - “Read the whole repo first” is a poor Codex job because input tokens spend implementation budget
Plan ceilings are not a fixed message count
Third-party price tables and “N messages” claims are not adoption guarantees. Official pricing says consumption moves with model and tool use.
daily.dev’s Pro recommendation assumes daily long agent sessions. If you only ship a few small pull requests a week, cut the tasks to fit the Plus five-hour window before you change plans.
The useful technology choice is not a new model name. It is who owns the remaining quota. Plan in chat, delegate on GitHub, then merge only after a human reads the diff. That order keeps the agent pool for implementation and final review.
FAQ
Caps and plans
Review and surfaces
Mixing research with implementation
What to do next
Pick one next job. If you cannot write the plan, stay in Chat until you have one paragraph, then stop. If the change set is written, confirm GitHub settings and send one @codex review or one implementation task. Before merge, a human should check permissions, secrets, and breaking migrations.
Third-party price and message-count tables in this article are snapshots from the research date. Open the official pricing page before you change a plan.
Related articles:
Author
krona23
Over 20 years in the IT industry, serving as Division Head and CTO at multiple companies running large-scale web services in Japan. Experienced across Windows, iOS, Android, and web development. Currently focused on AI-native transformation. At DevGENT, sharing practical guides on AI code editors, automation tools, and LLMs in three languages.
🔥 Most Popular
- Claude Pricing: Free, Pro, Max & Team Plans Compared (August 2026)
- Claude Desktop Won't Install? Windows & Mac Fixes That Worked (2026)
- Cursor Pricing 2026: Plans & Real Costs After 3 Years of Pro
- n8n vs Dify vs Zapier vs Make: Tested in Production (2026)
- AI Code Editor Comparison 2026: 6 Tools Tested, Why I Use Zed + Claude Code
Leave a Reply