Stop Hitting Codex Usage Limits (2026): Resets, Free Tiers & Token Habits

Codex quit halfway through a refactor and told you to wait five hours. You barely typed twenty messages. The burn is not your prompts. It is context you re-send on every single turn.

This guide covers three things: how the limits actually work, how to redeem a free reset today, and the config changes that stretch a $20 plan.

How Codex limits work in 2026

OpenAI documents the plan mechanics at learn.chatgpt.com/docs/pricing and in the Codex rate card.

• Two windows, not one. A rolling five-hour window caps messages, and a weekly cap sits on top of it. The five-hour window resets five hours after the usage that consumed it. The weekly window resets seven days after your first message that week.
• Every surface shares the pool. Local CLI messages, IDE sessions, and cloud chats all draw from the same five-hour window. Switching models does not reset it.
• Pricing is token-based since April 2, 2026. OpenAI moved Codex off per-message pricing and onto API-style token rates for Plus, Pro, and Business.
• Credits are the paid escape hatch. Plus and Pro users buy extra credits under Settings → Usage → Credits. Rates run 125 credits per million input tokens on GPT-5.6 Sol down to 5 on Luna, with cached input at roughly 10% of standard.
• Cached input is the cheap path. Resuming an existing session reuses cache. Starting fresh on the same code pays full input rates again.

The plan ladder: Free at $0, Go at $8, Plus at $20, Pro from $100 with 5x or 20x the Plus limits, and Business at $20 per user.

Free ways to cut token burn

1. Use /clear between unrelated tasks. It starts fresh and stops re-sending old context. /compact summarizes instead, and it costs tokens to run, because it must read everything it condenses.
2. Plan in one session, implement in another. Write the plan to a file, clear, then say “implement PLAN.md”. The implementation runs on a lean context.
3. Cap tool output. Set tool_output_token_limit = 12000 in config.toml. Verbose command output is the single largest avoidable cost, and truncating it changes nothing about code quality.
4. Keep AGENTS.md short. It injects on every turn alongside the system prompt and every tool schema. Long memory files bill on questions they have nothing to do with.
5. Drop reasoning effort for routine work. model_reasoning_effort = "medium" cuts thinking tokens, which bill as output.
6. Match the model to the job. Smaller GPT-5.6 tiers cost a fraction on output. Save the top model for architecture and hard debugging.
7. Resume sessions instead of restarting. Cached input runs about 90% cheaper than fresh input.
8. Run /status before you panic. Confirm the block is really a usage limit and not an outage or auth failure.

Banked resets: the free lever most people never redeem

OpenAI shipped rate-limit reset banking on June 11, 2026 with Codex app 26.609. It matters because it is free capacity sitting in accounts unused.

• Every Plus and Pro subscriber received one complimentary banked reset at launch.
• A referral window in June granted additional resets, up to three referrals per user.
• Resets expire 30 days after they are granted. Unredeemed resets vanish.
• Redeem from the limit prompt in the app, or run /usage in the CLI. Reset banking needs CLI v0.135 or newer.
• You cannot buy resets. They come from grants and invitations. Credits are the paid option; resets are not.

Treat a banked reset like a limited refresh token. Spend it on the deploy you need finished tonight, not on the first wall you hit on a Tuesday afternoon.

When to pay instead

Three paths exist once the window empties, and they suit different situations.

If you overflow every week, credits are a patch and Pro is the fix. Pro gives 5x or 20x the Plus limits for a fixed price.

The 10-minute setup

1. Run /status and /usage to see your real position and any banked reset.
2. Redeem an expiring reset before day 30, or lose it.
3. Add tool_output_token_limit = 12000 to config.toml.
4. Cut AGENTS.md to the essentials.
5. Adopt the two habits that need no install: /clear between tasks, and plan-then-implement in separate sessions.

Those five changes do more for your weekly cap than any plan upgrade.

See Also

• Stop Hitting Claude Code Usage Limits: Best Tools and Settings (2026) — the sibling guide, with eight verified open-source tools that measure and compress context. Most of them work regardless of which agent you run.