← Volver a las guías

Codex API relay cache reads: lower your bill on 59API

Why your Codex bill keeps climbing

Codex sessions are token-hungry. Every turn sends context, and every reply adds output tokens. If you pay per request by real token usage, the number that matters is the rate attached to each token, not a flat monthly seat fee. 59API is an independent third-party AI API relay that bills per request by real token usage from a prepaid balance. There is no monthly fee, and your balance never expires, so the lever you actually control is the unit price of each token category.

Picture a hypothetical build: an agent loops over a codebase, sends similar context repeatedly, and asks for short structured answers. Without any cost discipline, similar context is re-read and re-paid for on every call. A cache read gives that repeated context its own price line, and on 59API that line is priced far below the input rate for every model listed.

Before you tune anything, check the three numbers that define your run: input, output and cache read. 59API publishes all three per 1 million tokens, billed in CNY with USD figures shown at ¥7.2 = $1. That means you can compare a real usage estimate against the numbers on your balance screen rather than against a marketing guess. The goal is not to change how your product behaves. The goal is to see which line of the bill is growing, then decide whether cache reads are carrying the weight you want them to carry.

What a cache read costs on 59API

Here is the short version of the rate card for models that work with Codex, the OpenAI SDKs, cURL and OpenAI-compatible clients such as Cursor, Cherry Studio, Cline and OpenCode. Each entry below is priced per 1 million tokens, and the cache read rate is the one to watch.

• GPT-6.1 Sol (gpt-6.1-sol): input $0.1847 (¥1.33), output $0.9264 (¥6.67), cache read $0.0375 (¥0.27).
• GPT-5.6 Sol (gpt-5.6-sol): input $0.3472 (¥2.5), output $1.94 (¥14), cache read $0.0278 (¥0.2).
• GPT-6 Astra (gpt-6-astra): input $0.6944 (¥5), output $3.47 (¥25), cache read $0.0694 (¥0.5).
• DeepSeek V4 Pro (deepseek-v4-pro): input $1.25 (¥9), output $3.75 (¥27), cache read $0.0417 (¥0.3).
• MiniMax M3 (MiniMax-M3): input $0.1389 (¥1), output $0.5556 (¥4), cache read $0.0278 (¥0.2).

Read those rows again and the pattern is obvious: the cache read rate is a fraction of the input rate on the same model. For a hypothetical workload that leans heavily on repeated context, that gap is where the savings live. Note that output tokens are billed on their own rate, so shortening your instructions to the model is a separate lever from cache reads. You need both to see a lower total.

Two worked examples, priced without cache

It helps to anchor the cache read rates against real totals. These examples use 59API pay-as-you-go pricing and are calculated without cache, so every token is billed at the full input or output rate.

For GPT-6.1 Sol, 5M input tokens plus 1M output tokens cost $1.85 (¥13.32) on 59API. At the official list price of $20 (about ¥144), 59API is about 9% of list price. That $1.85 is the ceiling of the uncached run: it is what the same shape of workload costs when nothing is served from cache.

For GPT-5.6 Sol, 2M input tokens plus 0.5M output tokens cost $1.67 (¥12) on 59API. At the official list price of $18 (about ¥129.6), 59API is about 9% of list price.

Now read those totals as a checklist rather than a headline. Step one: confirm which model your Codex workflow actually calls. Step two: measure how many tokens are fresh input and how many are output. Step three: look at how much of that input is repeated context that can be served as cache reads at $0.0278 (¥0.2) per 1 million tokens for GPT-5.6 Sol, or $0.0375 (¥0.27) for GPT-6.1 Sol. The uncached figures above are the baseline you are moving away from, not a floor you are stuck with.

Prepaid balance, top-ups and optional cards

59API keeps the money side simple. You top up ¥10, ¥50, ¥100, ¥200, ¥500 or ¥1000 by WeChat Pay or bank card, and invoices are available. Because you are charged per request by real token usage, the balance drains exactly as fast as your traffic dictates, and it never expires.

Monthly and weekly cards are optional, and they behave differently from your wallet. A card has its own key, can only call that family's models, and is paid only from the card. It never falls back to wallet balance, and a balance key never draws on a card. Unused quota does not roll over.

One example: G-PRO $1500 (GPT, 7 days) costs $40.56 (¥292), with quota $1500 and weekly cap $1500. Card quota and caps are metering units written in $, not official list-price dollars, so treat them as their own measure rather than converting them into a list-price comparison. If your priority is a predictable spend ceiling for a burst week, a card gives you one. If your priority is paying only for tokens you actually consume, the prepaid balance is the simpler route.

Connect Codex and start measuring cache reads

Setting up 59API takes one account, one API key and one API host: https://api.59api.com. For Codex, the OpenAI SDKs, cURL and OpenAI-compatible clients, use base_url https://api.59api.com/v1 and the chat endpoint https://api.59api.com/v1/chat/completions. These clients pair with GPT, DeepSeek and other non-Claude, non-Gemini models only.

Your checklist before the first run:

• Pick the model whose input, output and cache read rates match your workload.
• Top up a prepaid balance that comfortably covers your expected usage.
• Send your first request through https://api.59api.com/v1/chat/completions and confirm the response.
• Watch the balance after a repeated-context run and compare the cache read line with the uncached examples above.

That last step is where the theory becomes a number you can act on. If cache reads are doing the work you hoped, you will see it in the balance rather than in a spreadsheet. If they are not, you now know exactly which model and rate to test next.

One host, one key, per-request billing by real token usage, and a cache read rate that sits well below input. That is a short list, and it is the list worth trying. Start with a small prepaid top-up, point Codex at https://api.59api.com/v1, and let the first bill show you where the cache is working.

Docs