# How Codex Hosted Billing Works: $129 Flat, No Markup ProxyLLM charges $129/month flat, no inference markup. OpenAI bills your ChatGPT plan separately. What each invoice covers, what the dashboard tracks, and the free tier's scope. *Published 2026-06-12 ยท https://proxyllm.ai/blog/how-codex-hosted-billing-works* Codex Hosted billing is two invoices and one rule. We charge $129 a month, flat, and that is the entire ProxyLLM bill: no inference markup, no per-token fee, nothing that scales with usage. OpenAI bills your ChatGPT plan directly, exactly as it did before you connected the account. The rule: we never sit inside anyone else's charge. This page itemizes both invoices, explains the savings number the dashboard computes, and draws the line around the free tier. ## The two invoices | Line item | Billed by | Typical amount | What it buys | | -------------------------- | --------- | ---------------------- | ----------------------------------------------------- | | ChatGPT plan(s) | OpenAI | $20-200/mo per account | The Codex capacity your workloads actually run on | | ProxyLLM Codex Hosted | Us | $129/mo | Container, endpoint, fallback lanes, logs, dashboard | | API-key overflow, optional | OpenAI | Metered, usually small | Requests served while your plan windows are exhausted | Nothing in that table flows through us except our own fee. If you ever stop using ProxyLLM, your ChatGPT subscription and your OpenAI key keep working with OpenAI as if we never existed. ## What the $129 covers The fee pays for the operations layer that turns a subscription into infrastructure: - An isolated container per connected ChatGPT account, running the official, unmodified Codex CLI. One account, one container, never pooled. - The OpenAI-compatible gateway: one base URL, request queueing, standard response shapes. - Fallback lane ordering: second account, then your API key, then back when a window resets. - The request log, per request and per lane, so you can always answer "what ran where". - The dashboard: observed usage against your plan window, and the API-equivalent value math below. The $129 is the entire ProxyLLM invoice: it does not move with requests, tokens, or how heavy your month was. We deliberately avoid percentage pricing because a gateway that earns a cut of your spend has the wrong incentive about your spend. ## What OpenAI bills Three possible charges, all direct between you and OpenAI: 1. **Your ChatGPT plan.** $20 for Plus, $100 for Pro 5x, $200 for Pro 20x, billed by OpenAI on OpenAI's page. 2. **Overflow on your API key**, if you configured one as a fallback lane. It bills at OpenAI's published per-token rates, which we pass through untouched; current numbers are in [the API pricing reference](/blog/openai-api-pricing-explained-2026). 3. **On-demand Codex credits**, if you choose to buy them inside ChatGPT. That is a manual OpenAI purchase, never something we trigger. ## What the dashboard shows: API-equivalent value A flat fee makes one question harder to answer from invoices alone: what was the work worth? The dashboard answers it per request. Every call served on the Codex lane gets priced at current OpenAI API rates for the same tokens, and that number is recorded next to the request as API-equivalent value. A worked month, at June 2026 GPT-5 rates ($1.25 input, $10 output per million tokens): - One typical request: 2,000 input plus 500 output tokens, about $0.0075 at API rates. - 90,000 such requests in a month: roughly $675 of API-equivalent value, inside a Plus plan's estimated ~$700 window. - What you paid: $20 (Plus, to OpenAI) + $129 (us). That spend against $675 of API-equivalent work is about 78 percent less than the metered bill for the same requests. Overflow requests served by your API key show their actual metered cost instead, so the two kinds of spend never blur. Five full scenarios with this arithmetic, from $250 to $14,000 a month, are in [the savings examples](/blog/codex-hosted-savings-examples). The capacity figures here are planning estimates derived from plan windows, never guarantees; OpenAI tunes limits over time. ## The free tier Starter costs $0 and is deliberately useful on its own: bring your own OpenAI or OpenRouter keys as a passthrough lane with no markup, and get the request log and dashboard on top. It exists so you can see your traffic clearly before deciding anything. What Starter does not include is the Codex lane. Subscription-backed capacity, the isolated container, and fallback ordering are the $129 feature, because they are the part that costs us real operations work. ## Where the fee earns out Honest arithmetic, both directions. Below roughly $150-250 of monthly API spend the token savings alone do not clear $129 yet, and at that spend the reason to switch changes rather than disappears: the meter comes off, so the calls you ration on a per-token bill (a critic pass on every request, thousands of synthetic test conversations, an eval sweep across prompt versions) become free to run. A low bill is usually that rationing, not small ambition. The free Starter tier shows the API-equivalent number on your own traffic if you want to watch it first, and the one genuine misfit is a compliance rule against routing inference through a third party. The crossover math, tier by tier, is in [OpenAI API vs ChatGPT subscription cost](/blog/openai-api-vs-chatgpt-subscription-cost). Above the crossover, the fee amortizes fast: at a $3,500 workload on Pro 5x, you pay $100 for the plan and a flat $129 to us, against a $3,500 metered bill. Pricing questions usually end as arithmetic questions. The [calculator](/calculator) takes your current monthly bill and returns the covering plan, the total with our fee included, and the difference.