News / Model Watch · Coding-Tool Economics
Codex Gets Its 5-Hour Limit Back, Claude Code Cuts 25%: Subscription Quota Anxiety Meets Pay-Per-Token China APIs
Updated Aug 31: OpenAI reset quotas yet again on Aug 30 and revealed the bug haul behind them (details below). Two announcements in six days bookend a strange month for AI coding tools. OpenAI quietly brought back Codex's 5-hour limit on Aug 25 — after spending August handing out quota resets like party favors — and Anthropic's Claude Code will cut limits 25% from Sep 14. If your workflow lives inside these quotas, here's the August timeline, what officially changed, and the metered-API alternative the Chinese model ecosystem offers.
Codex restored its 5-hour usage cap for Plus users on Aug 25 (announced Aug 24 by engineering lead Tibo Sottiaux on X, who admitted "I had mentioned this a while ago, but then postponed it"). August saw three full quota resets (Aug 8, 13, 21) plus a pay-to-reset feature reportedly in testing. Aug 30 update: quotas were reset yet again — the 4th time in August — and OpenAI revealed it fixed 8 quota-draining bugs (a stuck /goal task once ate 15-70% of a weekly allowance); same quota now stretches 10-50% further, and an in-app usage breakdown is in development. Claude Code limits drop 25% from Sep 14 (official ClaudeDevs post, Aug 29). China model APIs bill per token with published rates — no 5-hour windows, no weekly caps — and officially support Claude Code, Codex-style CLIs and 20+ coding tools.
The August timeline, from official posts
- Aug 8 — reset #1. Sottiaux reset usage for all paid Codex and ChatGPT Work users, celebrating the GPT-5.6 Sol rollout and, memorably, clarifying he is not leaving for Anthropic.
- Aug 13 — reset #2. 15M active users milestone; another reset "landing in the next hour."
- Aug 21 — reset #3. 20M active users; every Codex/Work user credited a banked reset usable at will (banked resets expire 30 days after crediting).
- Aug 24-25 — the 5-hour cap returns. After weeks of weekly-cap-only usage, Plus accounts regain the 5-hour rolling window. Official rationale: smoothing compute load keeps weekly quotas generous, and casual users kept burning whole weeks by accident. Business Insider separately reported an unannounced $8 pay-to-reset button in testing for some Plus users.
- Aug 30 — reset #4, with the why revealed. Sottiaux reset all paid quotas again after a bug-fix sweep across 8 usage-accounting issues: compaction kept stale images in context (+~10% usage for image-heavy users), completed /goal tasks kept running, broken tools retried endlessly — one runaway task consumed 15-70% of a weekly allowance; a memory background job ran nearly 15,000 times. Net effect: the same quota now stretches 10-50% further. Architecture-level guards and auto-alerts are in, an in-app usage breakdown is in development — and Tibo teased that "in two-three months, today's Codex will look primitive."
- Aug 29 — Claude Code's cut. Anthropic's ClaudeDevs account announced limits drop 25% from Sep 14; community math reads the fine print as the weekly multiplier going from +50% to +25%, a ~17% effective cut.
Context: this is a pattern, not a week's news. In late June, an official "warroom" investigated Codex usage being consumed faster than expected — the fix note admitted background features (auto-review, subagents) sometimes ran twice or retried aggressively, and dashboards showed activity never charged. In March, limits were reset six times in one month. A community tracker (codex-resets.com) now archives every announcement.
Subscription quotas vs metered China model APIs
| Dimension | Codex / Claude Code plans | China model APIs |
|---|---|---|
| Billing model | Monthly subscription; opaque quota math (credits, multipliers, hidden windows) | Per-token, published rates — e.g. GLM-5.3 $1.40/$4.40 per 1M; DeepSeek V4 with ~50% off-peak discounts |
| Usage windows | 5-hour rolling windows (returning), weekly caps, policy changes month to month | None — run 100 tasks at 3am if you want; concurrency limits only |
| When you hit the wall | Wait for reset, buy credits ($8 reset reportedly testing), or bank resets that expire in 30 days | Top up balance; effective ceiling is your budget, not a window |
| Tool compatibility | Native ecosystem (VS Code, JetBrains, CLI) | GLM Coding Plan officially supports Claude Code, Cline, Roo Code +20 tools; Codex CLI is open source with custom providers |
Rates verified against official docs as of Aug 30, 2026; quota mechanics per official X announcements and press coverage linked below.
The practical takeaway for coding workloads
None of this makes Codex or Claude Code bad tools — they're excellent, and the resets were genuinely generous. But the mechanics matter for anyone running production workloads: when your costs are someone else's quota policy, your capacity planning lives in their changelog. The Chinese model API shelf — GLM-5.3 and the cheaper GLM-5.3-Flash, Kimi K2.6/K3, DeepSeek V4 Flash/Pro, Qwen3.8 — prices the same work per token, with rate cards you can audit and no window that resets under you. For overflow traffic, CI jobs, or the workloads subscription quotas keep squeezing, that's the structural draw: you trade quota anxiety for a meter you control.
Primary sources
- Tibo Sottiaux on X (Aug 24) — 5h limit returns Aug 25, official announcement
- ClaudeDevs on X (Aug 29) — Claude Code limits −25% from Sep 14, official announcement
- codex-resets.com — community archive of every reset announcement (Aug 8/13/21 entries verbatim)
- 9to5Mac (Aug 24) — restoration details and official rationale
- Business Insider — June usage-bug warroom, official fix note · $8 pay-to-reset testing
- Hacker News — restoration thread (119 points)
FAQ (2026)
When did the 5-hour limit return?
Aug 25, for Plus accounts across Codex and ChatGPT Work. Announced Aug 24 on X — with an admission it had been postponed before. Official reasons: smoother compute load, protection for casual users.
How many resets in August?
Four: Aug 8 (GPT-5.6 Sol), Aug 13 (15M users), Aug 21 (20M users + banked reset), Aug 30 (with the 8-bug fix reveal worth +10-50% effective usage). June's usage bug brought a warroom, a fix and compensation; March saw six resets in one month.
Claude Code: how big is the cut?
Officially 25%, from Sep 14. Community math on the multiplier (from +50% to +25% of base) puts the effective cut near 17%. Source: official ClaudeDevs X post, Aug 29.
Do China APIs have usage windows?
No. Per-token published rates only — no 5h windows, no weekly caps. DeepSeek adds time-of-day pricing; GLM Coding Plan bundles monthly quotas that officially support 20+ coding tools.
Can Codex CLI / Claude Code use Chinese models?
Yes — Claude Code via GLM Coding Plan (official support), and open-source Codex CLI via custom OpenAI-compatible providers: GLM-5.3, Kimi K2.6/K3, DeepSeek V4 all reachable through compatible endpoints.
Does this move China API prices?
Not directly — these are subscription-side quota changes. But tighter Western metering plus continued Chinese price cuts historically nudges price-sensitive coding workloads toward metered China model APIs.