A corrected model for GPT/Codex and Claude upstream account cost, sustainable retail multipliers, margin thresholds, and supply risk.
· 12 min read
Loading article…
Find models, providers, pricing pages, and practical guides.
The important correction is conceptual: 0.05×–0.10× describes the researched upstream cost of GPT/Codex account capacity. It is not the customer-facing price.
Keep 0.05×–0.10× in the internal supply ledger. The retail multiplier must be higher and must absorb account failures, idle capacity, rate limits, refunds, payments, and support. Claude also needs its own supply pool and cost model.
GPT / Codex internal cost
0.05×–0.10×
A procurement assumption that still requires token-log validation.
Claude provisional cost
0.08×–0.15×
An inferred planning band, not a verified acquisition price.
account supply cost c
= upstream account and direct acquisition cost
/ official API-equivalent value successfully delivered
If an account costs $100 and delivers $2,000 of usage valued at the official API rate card, then c = 100 / 2,000 = 0.05×. This is an internal cost—not a five-percent official discount and not the price charged to a customer.
Store supply cost, retail multiplier, direct-loss rate, and route type separately. A single generic multiplier cannot distinguish procurement economics from the price a buyer sees.
| Model family | Internal planning cost | Public market reference | Retail planning range* |
|---|---|---|---|
| GPT / Codex | 0.05×–0.10× | low-cost account-pool retail around 0.029×–0.086× | 0.10×–0.22× |
| Claude / Claude Code | provisional 0.08×–0.15× | constrained routes around 0.086×–0.15×; other groups up to 0.171×–0.286× | 0.16×–0.33× |
* Derived from 10%–15% direct losses and a 30%–40% contribution-margin target; not a final product promise.
The GPT band comes from the procurement research supplied for this project. Public prices can test whether it is structurally plausible, but they cannot prove the real cost. The Claude band is a provisional inference from first-party subscription structure and verified public retail samples; replace it after 7–14 days of workload logs.
OpenAI currently prices Plus at $20, with Pro tiers at $100 / 5× and $200 / 20× the Plus usage allowance.
| Account | Monthly fee | Relative allowance | Cost per 1× allowance |
|---|---|---|---|
| ChatGPT Plus | $20 | 1× | $20 |
| ChatGPT Pro | $100 | 5× | $20 |
| ChatGPT Pro | $200 | 20× | $10 |
The top tier therefore has roughly half the cost per relative allowance when fully utilized. That supports a two-pool hypothesis—roughly 0.05× for a highly utilized pool and 0.10× for a less efficient or higher-quality pool—but does not verify it.
Codex consumption varies with the model, context, reasoning, tools, caching, five-hour windows, and weekly limits. Its credit rate card maps tokens to API-equivalent value, but subscription allowances are not fixed API-dollar balances.
| Account | Monthly cost | Delivery needed for 0.10× | Delivery needed for 0.05× |
|---|---|---|---|
| Plus | $20 | $200 / month | $400 / month |
| Pro 5× | $100 | $1,000 / month | $2,000 / month |
| Pro 20× | $200 | $2,000 / month | $4,000 / month |
These are acceptance thresholds, not OpenAI guarantees. A $200 account that delivers $2,500 of API-equivalent usage actually costs 0.08×, regardless of the budget assumption.
Claude has a similar subscription ladder: Pro is $20, while Max 5× is $100 and Max 20× is $200. The relative allowance economics look similar, but Claude and Claude Code share usage and are subject to session and weekly limits.
Public relay pricing also runs higher than GPT in comparable low-cost groups. API Fast Link publicly normalizes to about 0.029× for Codex and 0.086× for Claude Max; APIKEY.FUN shows about 0.10× for Claude Lite; Pateway shows a Claude example around 0.15×. These are retail observations, not known procurement costs.
Until real logs exist, a useful budget split is:
0.08×–0.12× for a highly utilized Max 20× or constrained Claude Code pool;0.12×–0.15× for a more stable, broader, or less efficiently utilized pool;0.15×, raise retail, restrict expensive models, or change supply.Public pages reveal selling prices, not acquisition cost, failure rate, or pool utilization. Do not present 0.08×–0.15× as an industry procurement fact.
Let c be supply cost, s retail price, q direct losses, and g target contribution margin:
contribution margin = 1 - q - c / s
minimum retail s = c / (1 - q - g)
With 10% reserved for direct losses:
Supply cost c | Retail for 30% margin | Retail for 40% margin |
|---|---|---|
0.05× | 0.083× | 0.10× |
0.08× | 0.133× | 0.16× |
0.10× | 0.167× | 0.20× |
0.12× | 0.20× | 0.24× |
0.15× | 0.25× | 0.30× |
Doubling cost produces about 40% contribution margin after a 10% loss reserve. At 15% losses and the same margin target, retail needs to be about 2.22× cost. That implies a practical floor of roughly 0.10×–0.22× for GPT and 0.16×–0.33× for Claude.
Avoid using Pro and Plus as synonyms for the 0.10× and 0.05× cost pools. A clearer product structure is model family plus route quality:
| Customer route | Supply assumption | Product boundary | Illustrative retail |
|---|---|---|---|
| GPT Economy | low-cost pool | lower priority, concurrency, and rolling limits | 0.10×–0.12× |
| GPT Stable | higher-cost or lower-utilization pool | priority scheduling and clearer SLA | 0.20×–0.22× |
| Claude Code Economy | constrained, high-utilization Max pool | Claude Code only and strict weekly cap | 0.16×–0.20× |
| Claude Stable | broader and more stable supply | model-specific floors and independent limits | 0.24×–0.33× |
These are financial-model examples, not finalized prices. Set production retail from the conservative P90 cost of each pool, then disclose supported models, concurrency, rolling limits, expiry, client restrictions, peak priority, refunds, and compensation.
Log account tier, model, input, cached input, cache writes, output, official API-equivalent value, retries, rate limits, success, and account-failure cost. At period end:
real supply cost
= (account acquisition + failures + direct operations)
/ official API-equivalent value successfully delivered
Compute P50 and P90 separately for GPT and Claude, by account tier and model family. Observe at least one complete weekly-limit cycle; two to four weeks is safer.
ChatGPT subscriptions and the OpenAI API are separate products, as are Claude subscriptions and the Claude API. Both consumer terms restrict credential sharing, resale, and unauthorized automated access. An account-pool route must not be marketed as an “official API discount,” and it should not share the same reliability promise as an authorized API route.
Keep GPT 0.05×–0.10× internal, model retail at roughly 2.0×–2.22× verified cost, and build Claude as an independent 0.08×–0.15× provisional cost pool with a 0.16×–0.33× initial retail floor. Replace every planning band with measured P90 economics as soon as operational logs exist.
Primary references include OpenAI Pro tiers, Codex pricing, the Codex rate card, Claude pricing, the Claude Max plan, Claude Code plan usage, and Anthropic API pricing. Market bands come from browser-verified public pages in the Sub2API ecosystem.
Continue with the competitor multiplier distribution for GPT/Codex and Claude.