GPT-6 Sol: Codex Subscription Limits vs API Billing

GPT-6 Sol usage in a ChatGPT plan, Codex credits and API-key billing are three different accounting questions. A remaining allowance percentage is not a dollar balance, and a token rate cannot tell you how many useful repository fixes your subscription will deliver.

This guide is for developers reconciling a Codex session with its payment route. Official documentation was checked September 25, 2026. The examples are transparent calculations, not a customer invoice, a measurement of included allowance or a promise about any account. Prices below are OpenAI’s direct rates, not Ofox offers.

Identify which account pays before comparing costs

Start with the authentication and billing route used by the session. A ChatGPT-signed-in session can draw on included usage and, when applicable, purchased credits. API-key use follows the API billing arrangement for that key. A third-party provider is another contract and must be audited against its own records.

Do not publish keys or raw authentication files in a billing report. Record identifiers that let your own team reconcile the session, and keep sensitive account details in the account’s protected records. The Sol/Luna access guide covers entry points; this article concentrates on the ledger after access is established.

What the subscription estimate does and does not mean

The current Codex pricing guide gives Sol estimates of 15–150 local messages per five hours on Plus, 70–700 on Pro 5x and 300–3,000 on Pro 20x. These are ranges, not reserved message counts. The guide says model choice, task complexity, context, reasoning, retrieval, tools and caching affect consumption. Local messages and cloud chats share allowance, and weekly limits may also apply.

A short edit and an extended repository task can therefore use very different amounts. Dividing a subscription price by the top of a range does not produce a reliable per-task price. Nor should you multiply the five-hour upper bound across an entire month: usage patterns and additional limits prevent that from being a guaranteed capacity calculation.

For your actual remaining allowance, use the usage dashboard or /status in Codex CLI. The estimate is useful for orientation; the account’s current limit and reset information is the operational evidence. Do not interpret a changed estimate as proof that your account received more capacity.

Credits and API dollars use different cache accounting

Under the checked Standard credit rate card, Sol charges 50 credits per million input tokens, 5 per million cached input tokens and 250 per million output tokens. Codex credit billing has no separate cache-write charge. The guide also warns that credit prices alone do not determine included subscription usage.

For the direct Sol API at no more than 272K input tokens, the rates are $2 ordinary input, $0.20 cached input, $2.50 cache writes and $10 output per million tokens. Above that threshold, input/cache rates double and output rates increase by 1.5 for the full request. See API pricing and the Sol rate explanation.

A hypothetical credit-billed usage record with 80K uncached input, 20K cached input and 10K output costs 4 + 0.1 + 2.5 = 6.6 credits. A hypothetical API-billed record with 80K ordinary input, 20K cache reads, zero cache writes and 10K output costs $0.16 + $0.004 + $0.10 = $0.264. These are independent examples using explicitly chosen categories, not a subscription conversion rate. The dollars needed to buy credits depend on plan or agreement.

If the API’s 80K uncached input instead consists of 60K ordinary input and 20K cache writes, with the other volumes unchanged, the example becomes $0.274. Billing every input token as ordinary input and then adding cache writes again would double-count a category. Never infer API cache-write volume from a credit summary that does not expose it.

Keep speed settings in the correct ledger

The checked Codex guide lists Fast mode at 2.5 times the Standard credit rate for GPT-6 Sol where available. The API model page lists Fast mode at twice the applicable API rates. Applying one multiplier to the other billing route produces a wrong comparison even before token use changes.

Similarly, changing reasoning effort can change usage without being a separate model. A longer tool loop can consume more input on later requests and more billed output while producing a short final patch. For effort selection, use the Sol high/xhigh guide and measure accepted work rather than visible answer length.

Reconcile one completed task before extrapolating

Save the billing route, model, speed, start/end time, token categories where available, retries and acceptance result. Match API requests to the API project; match credits to the relevant workspace. Keep included allowance changes separate from purchased-credit deductions. If a field is absent, mark it unavailable instead of estimating it from an unrelated rate card.

For a purchasing decision, compare a representative workload across the routes you are eligible to use. Include subscription fees allocated to that workload, purchased credits or API charges, and review time. If the subscription also serves unrelated work, state the allocation assumption. No public rate table can establish your personal break-even point without those usage and allocation records.