SU8 Codes
< Back to blog
Official article Updated 10 min read

Why the Same GPT Model Costs 7.5% of List Price at One API Provider and 95% at Another

The same GPT model can show 7.5%, 3.3%, and 95% of list price at different API providers. Here is how channel, package, and cache pricing stack up, plus 8 questions to ask any provider.

#pricing#api-relay#codex

When you compare API relays, you may see three quotes for the same GPT model: Provider A at 7.5% of list price on pay-as-you-go and 3.3% for a 30-day package, and Provider B at 95%. The gap looks impossible: either someone is losing money or someone is lying. There is a third explanation: the three numbers describe different things. This article breaks a quote into three layers: channel, pay-as-you-go versus a package, and cache / peak-off-peak pricing. It also gives you eight questions to ask any provider, including us.

Put the three quotes back into their layers

The landed rate is what you actually pay divided by the list price: 7.5% means 7.5% of list price, and 95% means 95%. The displayed discount may differ from the exact value; the exact value is on your bill. The API value is about 7.44%, as section 6 shows.

The three quotes can all be true. They are describing different things:

QuoteWhat the quote describesWhat you actually pay
7.5%Provider A, pay-as-you-go7.5% of list price, with channel quota and model limits
95%Provider B, pay-as-you-go95% of list price; ask about quota, protocol, and stability
3.3%Provider A, the lowest-discount tier of its 30-day package (SU750, $750)An upfront package fee, with a daily quota ceiling

Put them back into their layers: channel, billing mode, and cache / peak-off-peak pricing.

A concrete example: Provider A's 7.5% is pay-as-you-go; its 3.3% is a package rate with an upfront payment and a daily quota; Provider B's 95% is another provider's pay-as-you-go rate. You cannot ask A to match its 3.3% on pay-as-you-go; the buying methods differ. Nor is it fair to call pay-as-you-go expensive because the package is 3.3%. First check the buying method.

Layer one: the channel

A model can use more than one upstream channel: official direct, cloud marketplace, a channel built for coding agents, and so on. Each has different costs, quotas, rate limits, protocols, and model coverage, so the landed rate differs.

Low prices usually come with limits. Do not assume cheap means scam or a free lunch. Ask: which channel is this discount from? What are the quota, rate limit, protocol, and model coverage? Does switching channels require a client change?

Some channels fit Codex, some Claude Code, some only part of the model list. Ask whether the same key can switch channels. The more limits a channel has, the lower its rate can be. That is the first reason quotes differ.

Layer two: pay-as-you-go or a package

Pay-as-you-go usually means topping up and paying for what you use; whether the balance expires varies by provider. It suits uneven usage, trials, or occasional runs.

Here, a package means a one-time purchase for 30 days with a fixed daily quota; other providers may auto-renew. Stacking and rollover rules vary. Ask how overage is charged: some continue from your balance, some stop. A package may also have model-scope limits. The questions below apply to both.

A package discount is not free. You pay upfront, the daily quota has a ceiling, unused quota is lost when the 30-day period ends, and how overage is charged varies; ask first. To judge a package, estimate daily usage against the daily quota. Do not stop at the "lowest 3.3%" headline.

Package tierPriceDaily quotaRPM (requests per minute)
Low$42.47$7100
Mid$108.35$25200
High$750$187.5500

These three tiers come from the author's own public package page and are here only to show how to read a table; section 6 explains who we are and how we label them. The daily quota is not a dollar amount you can spend directly; the actual deduction depends on the quota multiplier (the provider's coefficient for converting package quota; ask support for the exact rule). The 3.3% figure is the $750 SU750 tier, not the cheapest Pro tier: the rate is lower (you pay a smaller share of list price), but you pay more upfront.

If you already use a provider or the official API, check last week's bill: daily spend, the peak day, and empty days. If you have not started, estimate the calls you expect per day. Compare your average daily usage with the daily quota. If your average is far below the quota, part of the package fee is sleeping. If it is close to or above, the package is worth calculating. Do not extrapolate from your highest day or reject a package based on your lowest.

Small, bursty, or short-term usage is more flexible on pay-as-you-go. A package can pay off when usage is steady and the daily quota is used.

Layer three: cache and peak/off-peak

A bill is more than an input price and an output price.

Take GPT-6.1 Sol: $2 per 1M input, $10 per 1M output, $0.1 per 1M cache reads, and $2.5 per 1M cache writes. Here, cache reads are cheaper than input; ask whether others price it the same way. Cache writes are separate.

Coding agents resend the same system prompt and code context, so a high cache hit rate lowers the bill; cache writes and long contexts push it up. The same million tokens can cost different amounts at different hit rates.

Cache hits are not automatic. Context, changes to it, and cache retention all affect the hit rate. Keep the system prompt stable, watch the cache-token share in the bill, and remember that a low cache price does not help if your hit rate is near zero.

Some models have peak and off-peak prices. If the price page lists the windows and multipliers, use them; if it does not, do not assume an off-peak discount.

Ask: how are cache reads and writes priced? Is there peak/off-peak pricing? How is long context billed?

Eight questions to ask any provider

  1. Which channel is this discount from? What are the channel's quota, rate limit, protocol, and model coverage?
  2. Is it pay-as-you-go or a package? For a package: validity, daily quota, and whether unused quota rolls over?
  3. When the package quota runs out, does it continue automatically from your balance? What happens if the balance is insufficient?
  4. How are cache reads and cache writes priced? Is there peak/off-peak pricing? How is long context billed?
  5. How long is the billing detail retained? Can you check the model, channel, tokens, and cost per request?
  6. Are refunds, invoices, support, a status page, and availability data available?
  7. How does the model name relate to the official model? Is it an alias, a resale, or a self-hosted endpoint? Is the protocol compatible?
  8. Are there minimum top-up, minimum spend, hidden multipliers, or model-scope limits?

If a provider cannot answer, or only says "it's cheap, don't worry," that is a warning sign. A low price is fine, but the provider should explain where it comes from and what the limits are.

A transparent example: how SU8 labels the same layers

We apply the same framework to our own pricing. For GPT-6.1 Sol: Codex pay-as-you-go is 7.5% of list price (page value; the API's exact value is about 7.44%), Azure pay-as-you-go is 95%, the Codex SU750 package is 3.3%, and the public channel table lists Codex-Pro (another public channel) at 9%.

The model page lists channel, quota multiplier, and billing unit; the pricing page lists daily quota and RPM; when the package quota runs out, calls continue from your balance; the console keeps per-request detail for 31 days, with daily summaries and the wallet ledger kept long-term (subject to the terms of service). How to switch channels is in this doc.

What we do not have: a public status page or availability data; if you need a written SLA or a direct vendor contract, confirm first; packages are one-time purchases, not auto-renewals; refunds carry a 10% fee. These are inputs for your decision.

Ask us the same eight questions. The answers should be on the page and in the bill; if they are not, we have work to do.

When a package or a relay is not the right fit

Your usage is small or one-off: pay-as-you-go is simpler than committing to a 30-day package.

You need a direct contract with the model vendor, specific compliance commitments, or a written SLA: confirm what the relay can provide; do not assume it can meet those requirements.

You want a direct connection to the vendor and no relay in between: that is a reasonable choice.

You need a public status page or availability data to decide: check whether the provider has one; if not, choose another provider or go direct.

You are running a short-term project: do not lock a 30-day package around it; pay-as-you-go is cleaner.

If you are worried that a small relay might disappear: start small on pay-as-you-go, do not top up a large amount at once, and do not prepay for a year; confirm the refund path for your balance. No test can spot one; cap the loss from any single provider. This applies to any relay, including us.

Conclusion

Back to the opening question. 7.5%, 3.3%, and 95% can all be true, but they describe different layers. Next time you see a quote, ask the eight questions, then do your own usage math.

Start small on pay-as-you-go, check one real bill, and then decide whether a package makes sense. If you want to see how SU8 labels the same three layers, open the GPT-6.1 Sol model page; to compare other models, open the model list.

Prices change; the figures here are as of 2026-11-16. This article was written by the SU8 Codes team.

FAQ

Is 7.5% real? Why is it so cheap?

It can be real, but it is a pay-as-you-go rate for a specific channel. A low price usually comes with limits on quota, model coverage, protocol, or capacity. Ask what the limits are before you pay, using the eight questions above.

Is a package always cheaper than pay-as-you-go?

No. Divide your average daily usage by the daily quota. If you stay far below the quota for 30 days, the discount is mostly unused; also factor in the quota multiplier, model scope, and refund fee. A package pays off when usage is steady and the quota is used.

What is the difference between a relay and the official API?

A relay may switch channels, convert protocols, and pool quota. The model name, protocol, data path, and failure handling can all differ from the official API. Ask the eight questions; do not assume it is "the same as official."

Can I refund an unused package?

It depends on the provider. SU8 allows refunds with a 10% fee; other providers may differ. Ask before buying, especially for a large package.

How does caching save money?

Here, cache reads are cheaper than input. Coding agents resend the same prompt and context, so a high hit rate lowers the bill. Cache writes and long contexts push it up, so check all four prices.

Back to the opening: ask the eight questions, run your own usage numbers, and 7.5%, 3.3%, and 95% stop being contradictory.