Developers have been comparing Codex usage meters in public threads for weeks — some accounts drain in days, others survive a full week on the same plan tier. On August 21, 2026, OpenAI Codex lead Tibo Sottiaux posted a direct explanation: after investigating, many of the affected users were on sub2api — routing a personal ChatGPT subscription through a proxy that re-serves it as shared API traffic. That pattern is not supported and gets flagged by fraud-prevention systems.
The same post drew a line around what is supported: Sign in With ChatGPT through official clients or open-source tools (Pi, OpenCode, and similar) that authenticate with your account and consume your included usage.
Community pushback is real too — plenty of builders on the official ChatGPT app report faster limit drain since GPT-5.6 Sol landed. Those are separate complaints from the sub2api fraud flag, but both matter if you are budgeting agent time this month.
TL;DR — sub2api vs Sign in With ChatGPT
| Question | Answer |
|---|---|
| Did OpenAI secretly change Codex limits? | Tibo says no — limit changes would come with community engagement and transparency. |
| Why did some accounts hit limits faster? | Many traced users were on sub2api — subscription converted to shared API traffic. |
| Is sub2api supported? | No. Fraud-prevention systems flag subscription resale and pooling. |
| What is supported? | Sign in With ChatGPT in official Codex/Work clients or OSS clients (Pi, OpenCode) on your account. |
| Are official-app users affected too? | Community reports say yes — Sol and larger context increase burn on the shared Work/Codex pool. |
| What should teams do? | Stop relaying subscriptions; use OAuth or a real API key; route heavy tasks like AT&T's LiteLLM playbook. |
What Tibo posted
Tibo's August 21 thread (paraphrased for clarity — see his post on X for exact wording) covered four points:
- Investigation — OpenAI looked into messages claiming Codex usage limits were different across accounts.
- Transparency — limit policy changes are not something OpenAI does quietly; they would engage the community first.
- sub2api pattern — talking to affected users, many were using sub2api: converting a subscription into API traffic, then re-serving or sharing it across many users.
- Supported path — you are fine using your subscription through Sign in With ChatGPT, whether in OpenAI's official clients or OSS clients that sign in with your account and use included usage.
That is the clearest on-the-record distinction OpenAI has drawn between personal OAuth sessions and subscription pooling proxies — at least for Codex and ChatGPT Work.
What sub2api actually does
sub2api sits in the same category as newapi and other relay panels explainx.ai mapped in the AI token black market post and the July token relay market anatomy follow-up.
The mechanics, simplified:
Paid ChatGPT/Codex seat → sub2api gateway → many downstream "API users"
One subscription (or a farm of them) backs an OpenAI-compatible endpoint. Resellers meter tokens, undercut official API list prices, and hide the upstream account. Consumers get cheap inference; OpenAI sees one subscriber's credentials serving traffic that looks like a small API business.
That is not the same as:
Your ChatGPT account → Sign in With ChatGPT → Codex CLI on your laptop
The second path is one human, one account, one quota bar — exactly what Tibo endorsed.
Why fraud systems care
Subscription pooling breaks the economic model OpenAI priced into Plus, Pro, and team plans. Included Codex and ChatGPT Work usage share one agentic pool; reselling that pool as pseudo-API capacity is terms abuse and looks like the same account-farm patterns labs already fight in distillation campaigns.
Tibo's wording — flagged by fraud-prevention systems — implies automated classifiers, not a manual support ticket. If your endpoint suddenly throttles or your account gets restricted, check whether any tool in your stack proxies through a shared subscription gateway before assuming OpenAI changed the quota formula.
Sign in With ChatGPT — the supported integration model
Sign in With ChatGPT is OpenAI's OAuth flow for third-party apps to act as you, with your plan limits. Supported surfaces Tibo named explicitly:
| Client type | Examples | What you get |
|---|---|---|
| Official OpenAI | Codex app, CLI, IDE extension, ChatGPT Work mode | Full harness, your quota |
| OSS with OAuth | Pi, OpenCode, other community clients with ChatGPT sign-in | Included usage on your account |
This is the same integration philosophy OpenAI expanded in Codex as a platform — embed the harness, authenticate the user, respect sandbox and approval boundaries — but Tibo's August post adds a compliance line: OAuth for yourself is in; subscription gateways for a crowd are out.
Codex inside another harness
Using Codex through your account inside a different product is still on the supported side when OAuth is yours. explainx.ai's OpenAI Codex plugin for Claude Code guide walks that path: Claude Code owns the UI; Codex runs delegated tasks against your ChatGPT/Codex entitlement.
Do not confuse that with pointing Claude Code at a Telegram-sold sub2api URL — same IDE, completely different trust and terms posture.
Sol-era drain — the complaint Tibo did not fully resolve
Pushback in replies is worth separating from the sub2api story:
- Official-app users report hitting weekly caps faster after GPT-5.6 Sol became default in Codex and Work.
- Sol ships with a larger context window; Tibo himself posted config for 1M-context Codex through ChatGPT accounts — with a warning that defaults were tuned carefully for a reason.
- OpenAI has been issuing manual usage resets around milestones — 8M users, 15M users, August 9 celebrations — but those are gifts, not a published new baseline. See Limit Reset Day for how 5-hour, weekly, and banked buckets interact.
If you are on official clients only and still burning out in 2–3 days:
- Check whether you enabled 1M context or Sol Ultra on long-running loops.
- Remember Work and Codex share one pool — a deck-building Work session and a repo refactor both draw from the same meter.
- Watch for auto-review and background agent passes eating quota silently — Codex auto-review controversy covers that thread.
- For team infrastructure, route cheap steps to cheaper models the way AT&T cut coding spend 56% with LiteLLM — included Codex usage is finite; not every subtask needs frontier Sol.
What developers should do
If you are on sub2api or a similar relay today
- Stop routing production work through pooled subscription endpoints — account bans and sudden cutoffs are common in gray-market threads.
- Migrate to Sign in With ChatGPT in an official or documented OSS client for personal included usage.
- Buy real API capacity for anything that needs SLAs, logging, or team billing — subscription resale is not an enterprise contract.
- Read the June token black market post for why resellers also log prompts and reasoning traces — a separate risk from quota alone.
If you are already on official OAuth
- Confirm both quota bars (5-hour where active, weekly) in Settings or the Codex usage page before blaming fraud classifiers.
- Right-size context — Tibo's 1M config is opt-in; defaults exist because long context costs real tokens.
- Split chat from agentic work where you can — ChatGPT Work vs Codex documents which surfaces still route around the shared pool via Quick chat.
- Evaluate routing before you assume you need more subscription tier — enterprise teams are cutting burn with gateways and open-weight fallbacks, not bigger relay pools.
If you run a team
- One seat per human with OAuth — never a shared sub2api key in CI.
- Central API keys with budgets for shared automation.
- Model routers for repetitive lint/fix loops — the pattern in Databricks' cost-at-scale post and AT&T's August numbers.
What people are asking
"Is OpenCode safe if Tibo named it?"
Tibo listed OpenCode as an example of an OSS client that supports Sign in With ChatGPT — not an endorsement of every deployment. Safe means: you downloaded a client that implements OAuth to your account, not a hosted instance someone else operates with pooled credentials.
"Did Anthropic ban Boris for the same thing?"
Different company, related anxiety. Tibo vs Boris — Claude proxy suspension covered harness freedom and account classifiers on the Claude side. The lesson generalizes: upstream providers instrument for sharing patterns.
"Should I use OpenRouter Ox Alpha instead?"
Ox Alpha is a separate OpenRouter stealth model with its own terms — not a workaround for Codex subscription limits. If you need cheap agent tokens, evaluate it on its merits (Ox Alpha guide), not as a substitute for terms-compliant Codex access.
Related on explainx.ai
- ChatGPT Work vs Codex — complete guide — shared quota, mode split, Quick chat loopholes
- Codex as a platform — open agent harness — codex exec, SDK, app-server
- AT&T cut AI coding costs 56% with routing — LiteLLM tokenomics playbook
- AI token black market — sub2api and newapi — gray-market mechanics
- Limit Reset Day — Claude, Codex, Cursor — 5-hour vs weekly vs banked
- OpenAI Codex plugin for Claude Code — delegated Codex on your account
- Codex 15M users — Tibo reset tease — manual refill context
- Codex auto-review token usage controversy — hidden burn sources
Sources: Tibo Sottiaux on X (August 21, 2026) · OpenAI Codex pricing · OpenAI Developer Community — Codex rate limits
Subscription pooling tools, fraud classifier behavior, and Sol default context settings may change after publication. Verify current terms on openai.com and your account usage page before routing production traffic.