Claude Code vs Codex Rate Limits Compared (2026)

If you run Claude Code and Codex side by side — and plenty of people do — you eventually hit both walls and start asking which limit system is actually fairer. The two tools meter differently, publish different amounts of detail, and as of September 14, 2026, they're in different post-promotion worlds. Here's the side-by-side of how the limits really work.

How Claude Code's limits work (post–September 14)

Claude Code meters usage on a rolling five-hour session window plus a weekly window. The weekly reset day is assigned to your account, not chosen by you, and it isn't rolling — Pro carries one weekly limit across all models; Max carries two (one across all models, one Sonnet-only, per Anthropic's Max page). On Team and Enterprise, the allowance is shared with Claude chat and Cowork from the same per-seat pool. Anthropic publishes the shape of both windows and no absolute figures — the only reliable number is your own /usage readout.

The big 2026 change: Claude Code ran with a temporary 50% boost to weekly limits from May 2026, extended repeatedly — announced August 19 through August 31, then last extended through September 13 — and it expired around September 14, 2026 without an announcement, simply by the date passing. Anthropic's support documentation never contained the boost at any point. Reports after the expiry put the settled level at roughly +25% over the old baseline — a permanent but smaller number than the temporary +50% — but treat that as reports rather than documented policy: no support page confirms it. Complaints about the effective cut surfaced widely as the date passed.

How Codex's limits work

Codex meters on a rolling five-hour window plus a weekly cap sitting on top of it. OpenAI publishes ranges, not fixed numbers — and the ranges are wide. From OpenAI's pricing docs (read August 21, 2026 via the Codex usage-limits tracker):

Two things stand out. First, Pro tiers are exact 5x/20x multipliers of Plus. Second, model choice moves your ceiling far more than your plan does: on Plus, Luna gives 25x the messages of Sol — more than upgrading Plus to Pro 5x would buy you. The weekly limit that sits above the window exists but has no published size at all; OpenAI's own footnote only says "additional weekly limits may apply."

The quota is also not perfectly reliable. On September 9, 2026, an open bug report in openai/codex (#44233) described a weekly quota dropping from 21% remaining to 0% while idle, with the reset timestamp changing unexpectedly — labeled a bug, still open at time of writing. Plan your week assuming the meter can surprise you.

Which workload suits which

• Steady, all-week agent work → Claude Code's weekly pool. A weekly ceiling with a fixed reset rewards consistent use; you plan around one number per week. But the post-boost reality means less headroom than you had in August, and Fable-class models drain it faster than other models by Anthropic's own documentation.
• Bursty, model-flexible work → Codex's 5-hour windows. Five hours roll, so a hard morning doesn't wreck your week — and dropping from Sol to Terra or Luna multiplies your allowance instantly. The tradeoff: ranges instead of numbers, an invisible weekly cap, and the occasional idle-quota bug.
• Shared allowance teams: Claude's Team/Enterprise pool includes chat and Cowork — one heavy chat user eats agent headroom.

What to do when you hit the wall

On either tool, the first move is diagnosis: Claude Code's /usage (weekly vs session tells you which ceiling you hit), Codex's status output for window and weekly state. Then, in order: route routine work to a cheaper model (Luna over Sol, Sonnet over Fable), shrink context churn so the same files aren't re-read endlessly, and shift movable tasks past the reset — Claude's weekly reset schedule and Codex's limit mechanics cover the timing — waiting for the reset, switching models, or buying credits are the remaining options — and our Claude vs Codex price comparison covers what upgrading actually buys.

The meshcode angle

meshcode is a native desktop app for Mac and Windows designed for exactly this situation: run Claude and Codex side by side in separate panes, connecting your own CLIs — each provider bills you directly, and meshcode charges no extra token fee on those tokens. When one wall comes down, the other pane keeps working. For meshcode's built-in model, the balance is prepaid — top up from $1, a flat 5% + $0.50 fee at top-up time, tokens at cost, no subscription and no overage possible. Budget-friendly, not the floor.

See the pricing yourself: meshcode is free to start — download it

👉 Download meshcode — Mac, Windows