Inter-office memo
- To
- Everyone whose Codex limits run out too soon
- From
- Mo, Chief Executive
- Date
- November 16, 2026, 1:45 AM CST
- Re
- Making Codex usage last
How to make Codex usage limits last longer in 2026: 10 ways compared
Start with Settings → Usage or /status, then match the model to the task: on Plus, OpenAI estimates 350 to 3,000 messages every five hours with GPT-6 Luna, against 15 to 160 with GPT-6.1 Sol. Lower effort, Fast mode off and a new chat per task do most of the rest. Claude Code or a local model can take routine work, and Enjoy hands it over mid-conversation.
Disclosure: Enjoy is our product. We checked every command, limit and price below against OpenAI’s, Anthropic’s, Ollama’s and OpenCode’s own pages on November 16, 2026, and we say where another way is the better pick, including the free ones that don’t need Enjoy.
The ways to make Codex’s limits last, compared
Codex doesn’t count messages. “The number of messages you can send depends on the model used, size and complexity of your tasks, and whether you run them locally or in the cloud,” says its pricing page. Plus has a five-hour limit and a weekly one, and “Where both apply, you need allowance remaining in both to continue,” says OpenAI’s usage guide. Pro 100, Pro 200 and Pro 500 “currently have no five-hour usage limit,” but “Your plan still has an included usage allowance.” ChatGPT Work draws on the same pool: it “uses the same pricing, credits, and usage limits as Codex.”
| Way | Cost, checked November 16, 2026 | What it cuts | Best for |
|---|---|---|---|
| See where it goes, in Settings → Usage, /status and /usage | Free | Nothing by itself: it shows what’s left and your pace | Not knowing where your usage went |
| Match the model to the task | Free | A bigger model’s higher cost per message | Clear, repeatable work on Luna or Terra |
| Lower the effort, and skip Max and Ultra | Free | Reasoning, and Ultra’s subagents | Quick, well-scoped changes |
| Leave Fast mode off | Free | Fast’s 2.5x rate, or Ultrafast’s 8x | Work you aren’t waiting on |
| A new chat for each task, compacted when long | Free | History carried with every message | Sessions that run all day or wander |
| Trim what loads with every message | Free | AGENTS.md, MCP servers, long tool output, extra files | Big repositories and many MCP servers |
| Scope the task, and plan the big ones | Free | Broad reading and redone work | Vague or wide changes |
| Rein in subagents and the other draws | Free | Subagents, cloud tasks, images and voice | Heavy users of subagents and the cloud |
| Claude Code or a local model for routine work | Claude Pro $20 a month, or $17 billed yearly; Max $100; a local model is free | Work moves off OpenAI’s limits | Work that doesn’t need Codex |
| Enjoy | A monthly subscription, free for 7 days, with your phone and one computer (pricing) | Shows every account’s limits, and moves work to another agent mid-conversation | Seeing and spreading usage across agents |
Prices are each company’s US list prices, before tax. The first eight ways are settings and habits in Codex itself, and a local model costs nothing extra either.
Why Codex runs out so fast
Context is a big part of it. “Small scripts or routine functions may consume only a fraction of your allowance, while larger projects, long-running tasks, or extended sessions that require the agent to hold more context will use significantly more per message,” says Codex’s pricing page, and “Your prompt, files, chat history, tool results, and ChatGPT’s response all use tokens.” OpenAI’s usage guide adds that “Larger inputs and outputs, higher reasoning settings, Fast mode and tasks with multiple steps can also increase usage.” So the same request can cost very different amounts: “Tasks that look similar can consume different amounts of your allowance.”
It’s a common question on OpenAI’s own forum, and OpenAI’s pricing page answers it with five tips: smaller prompts, fewer source files, output matched to the need, a smaller AGENTS.md and fewer MCP servers. The ways below cover those, and the settings that matter as much.
See where it goes: Settings → Usage, /status and /usage
Best for: anyone who doesn’t know their pace. Start here.
“Check Settings → Usage for your remaining allowance and reset times,” says OpenAI’s usage guide, in ChatGPT or the ChatGPT desktop app. In the Codex CLI, “If you want to see your remaining limits during an active Codex CLI session, you can use /status,” per the pricing page. For how fast it goes, /usage shows “daily, weekly, or cumulative ChatGPT token activity,” or type /usage daily, /usage weekly or /usage cumulative, per Codex’s command docs. To keep your limits on screen, run /statusline, whose items “include model, model+reasoning, context stats, rate limits,” and token counters.
OpenAI’s own advice: “Check the dashboard every week or two to understand your pace and remaining capacity. If usage is higher than expected, consider whether a smaller model or tighter task scope would still produce a useful result.”
Price: free.
Watch for: OpenAI’s message counts are estimates, not a budget: “These estimates are not fixed message limits.” The dashboard is what tells you where you stand, and before a big task, OpenAI suggests you “Check Settings → Usage before a large task to see your remaining allowance, usage windows and reset times.”
Match the model to the task
Best for: clear, repeatable work that doesn’t need OpenAI’s biggest model.
OpenAI’s own estimates show how much the model matters. On Plus, its pricing page estimates these local messages every five hours: 5 to 45 with GPT-6 Astra, 15 to 160 with GPT-6.1 Sol, 35 to 285 with GPT-6.1 Terra and 350 to 3,000 with GPT-6 Luna. Its models page says “Astra suits the most demanding workflows. GPT-6.1 Sol offers near-Astra performance at a lower cost than Astra,” and suggests “Luna, for clear, repeatable tasks,” those “when you know what a good result looks like, such as extraction, classification, transformation, and structured summaries.”
In the CLI, choose with /model, or start with codex -m gpt-6-luna. Set a default with model in config.toml, which the ChatGPT desktop app, the CLI and the IDE extension share. In the desktop app, move the Power control toward “Faster for faster, lower-cost work,” or open Advanced to pick a model. For subagents, Codex’s subagent docs suggest gpt-6-luna “when you want a faster, lower-cost option for lighter subagent work.”
Price: free. Codex’s pricing lists GPT-6.1 Sol and GPT-6 Luna on Plus, and GPT-6.1 Terra on Plus, Pro, Business, Enterprise and Edu. Free and Go get “GPT-6 Luna at Standard speed in the desktop app, subject to rollout.”
Watch for: Luna isn’t for everything. OpenAI’s subagent docs say “For most tasks in Codex, start with gpt-6.1-sol,” and its models page says “Keep Astra for your most demanding work.” And switching doesn’t bring back what you’ve spent: “Switching models does not restore allowance in a shared usage pool.”
Lower the effort, and skip Max and Ultra
Best for: quick, well-scoped changes.
“Higher effort can improve results for complex tasks, but it takes longer and uses more tokens,” says Codex’s models page. “Start with the default effort and increase it when the task needs deeper planning or analysis.” It suggests “Light in the ChatGPT desktop app, ChatGPT Work on the web, and IDE extension, or Low in the CLI,” for “quick, well-scoped tasks.” In the CLI, /model --reasoning=low lowers it and “keeps the current model,” per the command docs, and model_reasoning_effort in config.toml sets a default.
Max and Ultra go further. “Max gives the selected model more time to reason about a single task. Use it for the hardest problems, when depth matters more than speed or usage.” Ultra “uses subagents to handle separate parts of a complex task in parallel,” and OpenAI says plainly: “Most tasks do not need Max or Ultra.”
Price: free.
Watch for: less reasoning means less checking. Codex’s best practices keep “Medium or High for more complex changes or debugging.” Effort and model trade off, too: “Astra at Low effort can outperform Sol at High effort,” says OpenAI’s usage guide, and “Reasoning efforts don’t map exactly between model generations,” so try a familiar task at a lower setting first.
Leave Fast mode off
Best for: work you aren’t sitting and waiting for.
“You can increase model speed in Codex in exchange for higher usage,” says Codex’s speed page, and the exchange rate is steep: “For supported models, Fast mode uses included subscription limits at 2.5x the Standard rate.” Ultrafast, on Pro 500 and some Enterprise and Edu plans, “uses included subscription limits at 8x the Standard rate.” In the CLI, type /fast to turn Fast on or off, and add the Fast mode item with /statusline to see it in the footer.
Price: free. Leaving it off costs nothing; turning it on costs usage.
Watch for: Fast stays on until you turn it off. The command docs say /fast “toggles the tier and saves the selection,” and service_tier = "fast" in config.toml makes it the default, so check both if your week is going faster than you expect. On Pro 500, Ultrafast “uses your included usage first, then your available credits after that allowance runs out.”
Start a new chat for each task, and compact long ones
Best for: sessions that run all day, or wander from one problem to the next.
History rides along with every message. “Chats accumulate context, decisions, and actions over time,” say Codex’s best practices, and among the common mistakes: “Using one chat for an entire project instead of one chat per coherent outcome. This leads to bloated context and worse results over time.” In the CLI, /new will “Start a new chat inside the same CLI session,” and /clear will “Clear the terminal and start a fresh chat.” To go back to an earlier one, /resume will “resume a saved chat.”
When you carry on with a long task, /compact will “Summarize the visible chat to free tokens,” per the command docs: “Codex replaces earlier turns with a concise summary, freeing context while keeping critical details.” Codex also compacts on its own as a chat grows, and model_auto_compact_token_limit in the configuration reference sets the “Token threshold that triggers automatic history compaction.”
Price: free.
Watch for: don’t split a problem that’s still one problem. “If the work is still part of the same problem, staying in the same chat is often better because it preserves the reasoning trail. Fork only when the work truly branches.”
Trim what loads with every message
Best for: big repositories, a long AGENTS.md, several MCP servers, or long command output.
Some context loads every time. AGENTS.md “loads into context automatically,” say the best practices, and “A short, accurate AGENTS.md is more useful than a long file full of vague rules.” The pricing page says to “Reduce the size of your AGENTS.md,” and in a larger project, to “control how much context you inject through AGENTS.md files by nesting them within your repository,” since Codex reads them only from the project root down to the folder you start it in. “If AGENTS.md starts getting too large, keep the main file concise and reference task-specific markdown files.”
MCP servers cost on every message too: “Every MCP server adds more context to your messages and uses more of your limit. Disable MCP servers when you don’t need them.” Run /mcp to list them, and in config.toml, enabled = false will “disable a server without deleting it,” per Codex’s MCP docs, while enabled_tools keeps a “Tool allow list.” For long command output, tool_output_token_limit sets the “Token budget for storing individual tool/function outputs in history.” And name files rather than pasting them: “Provide only relevant files and, when possible, narrow the sources or date range.” In the CLI, Codex’s prompting guide says to “mention paths explicitly, or attach files with /mention and @ path autocomplete.”
Price: free.
Watch for: a long AGENTS.md may already be cut short. Codex “stops adding files once the combined size reaches the limit defined by project_doc_max_bytes (32 KiB by default),” says its AGENTS.md page, and it reads from the root down, so a long root file can crowd out the more specific ones below it. In the IDE extension, mind your tabs: it “automatically includes your open files as context.”
Scope the task, and plan the big ones
Best for: vague requests, and changes across many files.
What you ask for sets what Codex reads and writes. The pricing page says to “Be precise with the instructions you give the agent, but remove unnecessary context,” and to “Define the audience, format, and length, and separate required work from optional improvements.” The best practices suggest “A good default is to include four things in your prompt”: the goal, the files that matter, the constraints, and when it’s done, “such as tests passing.” That “helps Codex stay scoped, make fewer assumptions, and produce work that’s easier to review.” For a big or fuzzy change, “Plan mode lets Codex gather context, ask clarifying questions, and build a stronger plan before implementation. Toggle with /plan or Shift+Tab.” And for work you repeat, “stop relying on long prompts or repeated back-and-forth. Use a skill.”
Price: free.
Watch for: planning is a conversation with the model too, and it draws on the same allowance. OpenAI suggests it “If the task is complex, ambiguous, or hard to describe well,” so for a one-line fix, skip it and say exactly what you want. If a task drifts, OpenAI’s prompting guide says to “Narrow or stop the task if it starts doing work you no longer need.”
Rein in subagents, cloud tasks and the other draws
Best for: heavy users of subagents, Ultra, cloud tasks, images and voice.
Every subagent spends from your allowance. “Subagent workflows consume more tokens than comparable single-agent runs because each subagent does its own model and tool work,” say Codex’s subagent docs, and “If you don’t configure a subagent model or model_reasoning_effort, the subagent inherits the parent agent’s model and reasoning effort,” so a session on Astra at high effort starts its subagents there too. Set a lighter model for them in your prompt or config.toml, and agents.max_concurrent_threads_per_session caps how many run at once.
Other things draw on the same pool, per the pricing page: “Cloud tasks may use more of your allowance than local messages,” “Image generations use included limits 3-5x faster on average than similar turns without image generation,” and “Voice in Desktop uses your existing Codex usage budget at $0.05 per minute.” From December, ChatGPT joins them: “Plus, Pro, and Business plans’ usage allowance will be shared across ChatGPT and Codex,” OpenAI says, though “Instant chat will not draw from your usage.” Our memo on the change explains it.
Price: free.
Watch for: subagents can keep the main chat lean as well as cost more. OpenAI warns that if you flood the main chat with “noisy intermediate output such as exploration notes, test logs, stack traces, and command output, the session can become less reliable over time,” and it suggests subagents “for read-heavy tasks such as exploration, tests, triage, and summarization.” Keep them for that, on Luna where the job allows.
Hand routine work to Claude Code or a local model
Best for: work that doesn’t need Codex, so Codex’s limits go to work that does.
Other agents have limits of their own. Claude Code “is included in all paid plans,” says Claude’s pricing, and it reads your existing instructions: Claude Code “can read AGENTS.md as your project instructions, so a repository already set up for other coding agents works without adding a CLAUDE.md,” per its docs.
For no limit at all, run a model on your computer. “Codex can run against a local ‘open source’ provider such as Ollama or LM Studio when you pass --oss,” say Codex’s docs. Ollama’s Codex guide sets it up with ollama launch codex, and you “Type /model inside Codex to switch between your installed, tool-capable Ollama models without restarting.” One built for this is Recursive’s Flint, “a 39.8 billion-parameter mixture-of-experts model from Recursive, built for coding agents.” If you’d rather use another agent, OpenCode’s docs say that with Ollama running, “OpenCode probes http://127.0.0.1:11434, discovers completion models, and adds them to /models without an account connection.” Our guides to using Claude Code and Codex together and running a coding agent on a local model compare the ways to split the work.
Price: Claude Pro is $20 a month, or $17 a month billed yearly, Max is $100 and Ultra $200. A local model is free, on your own computer.
Watch for: Claude has limits too: “Most accounts have a limit that resets on a rolling five-hour session window, and paid plans add weekly limits on top,” and Claude Code “shares the same usage limits as the rest of your plan.” Each agent starts a conversation of its own, so tell it where things stand. Flint’s download is 23 GB, and Ollama says “Use a context window of at least 64k tokens for Codex.”
Enjoy
Best for: keeping every Codex account’s limits in view, and moving routine work to another agent without starting over.
Enjoy runs Codex, Claude Code, Grok Build, Antigravity, Muse Code, and local models through OpenCode, on Mac, Windows and Linux, on the plans you already have. For usage, it adds:
- Every account’s limits. The agent settings show Codex’s limits as bars, such as 5 hours and Weekly, with when each resets, plus your credits and free resets, and the account menu shows the same bars for each signed-in account (how).
- Your pace, week by week. AI plans in Settings shows a bar for each of the last four weeks for each agent account with a weekly limit, Codex included, the check OpenAI suggests every week or two (how).
- Model, effort and speed per message. The agent settings’ Model, Effort and Speed columns change them for your next message, and pointing to a speed shows “what it costs with your agent” (how).
- Compact from the menu. A Codex conversation’s ⋯ menu has Compact conversation, which summarizes the earlier part now (how).
- Another agent, mid-conversation. Switch to Claude Code or a local model for the routine part, and “The new agent gets the conversation so far.” Switch back, and Codex “continues where it left off” (how).
When a limit arrives anyway, Enjoy offers your free resets, moves the conversation to another of your accounts, or waits for the reset, as our guide to what to do when Codex hits its limit explains.
Price: Enjoy is one monthly subscription, and the first seven days are free, with a card to start. It includes your phone and any browser and one computer; each other computer is $9.99 a month, and each teammate you invite needs a subscription of their own. See pricing.
Watch for: Enjoy doesn’t change OpenAI’s limits or make Codex use fewer tokens on its own: the settings above do that. Its count of what each conversation used is for Claude Code only, since “Other agents don’t report what each turn used yet,” and when you compact a Codex conversation, Codex decides what to keep. An agent you switch to reads the conversation so far on its own plan. Enjoy runs the Codex CLI, which OpenAI includes from Plus. Enjoy isn’t open source, and it’s a monthly subscription on top of your AI plans once the free week ends.
When to stay with the defaults, or upgrade
The defaults are a fair start. OpenAI says “For most tasks in Codex, start with gpt-6.1-sol,” and “Start with the default effort and increase it when the task needs deeper planning or analysis.” Keep Astra, and higher effort, for “the hardest end-to-end work,” and save on the routine work around it.
If you’ve done all this and still run out most weeks, the plan is too small. Codex’s pricing page lists Pro at “Plans at $100, $200, or $500 USD per month,” against Plus’s $20, and “Pro plans currently have no five-hour limit.” For the odd heavy week, “ChatGPT Plus and Pro users who reach their usage limit can purchase additional credits to continue working without needing to upgrade their existing plan.” Our guides to ChatGPT Plus vs Pro for Codex and what to do when Codex hits its limit help you choose.
How we compared
We read Codex’s pricing, models, best practices, prompting, command, speed, AGENTS.md, MCP, subagent and configuration pages, OpenAI’s help center, Anthropic’s pricing and Claude Code docs, and Ollama’s and OpenCode’s pages on November 16, 2026, and every one is linked below. We quote each company’s own words for its limits, and leave out what it hasn’t documented, including other sites’ estimates of how much each habit saves. Plans and limits change often, so we recheck this guide against its sources at least once a month, and the date at the foot of the page says when we last did.
15.Questions.
Why does Codex use up my limit so fast?
- Mostly because of the model, the context and the settings. OpenAI says “Model choice, context, reasoning, tool use, retrieval, and caching all affect usage,” and that sessions that hold more context “will use significantly more per message.” A bigger model, higher effort, Fast mode and subagents each use more. Check Settings → Usage to see which limit is going.
Which Codex model uses the least of my limit?
- GPT-6 Luna. On Plus, OpenAI estimates 350 to 3,000 local messages every five hours with Luna, 35 to 285 with GPT-6.1 Terra, 15 to 160 with GPT-6.1 Sol and 5 to 45 with GPT-6 Astra. It suggests Luna “for clear, repeatable tasks,” and says “For most tasks in Codex, start with gpt-6.1-sol.”
Does Fast mode use more of my Codex limit?
- Yes, 2.5 times as much. OpenAI says “Fast mode uses included subscription limits at 2.5x the Standard rate,” and Ultrafast, on Pro 500, uses them at 8 times. Type /fast in the Codex CLI to turn it off, since Codex saves the choice for next time.
Does /compact save Codex usage?
- For the messages after it, it can. OpenAI describes /compact as a way to “free tokens,” and says sessions that hold more context “will use significantly more per message.” For an unrelated task, start a new chat with /new or /clear instead, and pick up the old one later with /resume.
Will chatting in ChatGPT use up my Codex limit?
- From December, on Plus, Pro and Business, yes, unless you chat with Instant. OpenAI says the allowance “will be shared across ChatGPT and Codex,” and “Instant chat will not draw from your usage.” ChatGPT Work already shares Codex’s limits. See our memo.
Can another agent take the routine work?
- Yes. Claude Code on a Claude plan has limits of its own, and a local model through Ollama or OpenCode has none. Enjoy moves a conversation to one of them with the conversation so far, and back to Codex for the hard part.