← Voltar aos guias

Claude Opus 5.5 pay-as-you-go on 59API: model costs

The problem: a premium model for every small request

Everyday coding has a predictable shape. You ask for a diff, a unit test, a log explanation, a short script, a one-line refactor. The expensive mistake is treating every one of those calls as a premium task. The other mistake is subscribing to a monthly plan, paying for it in a month where you barely used it, and still wondering where the money went.

59API is built for the first approach. It is an independent third-party AI API relay: one account and one API key, one API host at https://api.59api.com, covering Claude, GPT, Gemini, DeepSeek and other models. You top up a prepaid balance, and every request is charged per request by real token usage. There is no monthly fee, and your balance never expires. A quiet week costs you nothing extra, and a heavy day costs exactly what you used.

Funding it is deliberately simple. Top-ups come in ¥10, ¥50, ¥100, ¥200, ¥500 and ¥1000, paid by WeChat Pay or bank card, and invoices are available if you need one. You are not buying a seat or a tier. You are buying tokens, and you choose the model for each task.

Pay-as-you-go discipline is therefore the whole game: match the model to the job, watch the token rate, and let a low-rate model handle routine traffic. The numbers below make that choice concrete.

Claude rates on 59API: a three-step ladder

All prices are per 1 million tokens. Billing is in CNY, and USD figures are shown at ¥7.2 = $1. For Claude work, three models cover most practical needs, and the gap between them is large enough to notice.

• Claude Opus 5.5 (claude-opus-5-5): input $1.53 (¥11), output $7.22 (¥52), cache read $0.0694 (¥0.5). Official list price reference: $4 / $20.
• Claude Sonnet 5.5 (claude-sonnet-5-5): input $1.11 (¥8), output $5.56 (¥40), cache read $0.0972 (¥0.7). Official list price reference: $2 / $10.
• Claude Haiku 4.5 (claude-haiku-4-5-20251001): input $0.3472 (¥2.5), output $1.67 (¥12), cache read $0.0278 (¥0.2). Official list price reference: $1 / $5.

Read those three rows as a routing policy rather than a ranking. Haiku is the economical default for short edits, quick explanations and routine test generation. Sonnet 5.5 is the middle ground for multi-file reasoning and refactors where quality still matters. Opus 5.5 is the deliberate choice for the hard, high-stakes prompt where a wrong answer costs you more than the tokens ever would.

Cache reads deserve attention too, because repeated context is normal in coding work. Opus 5.5 reads cached tokens at $0.0694 (¥0.5) per 1 million, and Haiku 4.5 at $0.0278 (¥0.2). If your workflow reuses a repository context, a stable system prompt or a long file set, cache read pricing can matter as much as the headline input rate.

What a typical day actually costs

Rates are easier to judge once they are attached to a workload. Here are three worked examples, all without cache, so you can see the input and output split clearly.

• Claude Opus 5.5, 2M input + 0.5M output tokens: $6.67 (¥48) on 59API. At official list price, the same workload is $18 (about ¥129.6), so 59API is about 37% of list price.
• Claude Sonnet 5.5, 3M input + 0.5M output tokens: $6.11 (¥44) on 59API. At official list price, $11 (about ¥79.2), so 59API is about 56% of list price.
• Claude Haiku 4.5, 5M input + 1M output tokens: $3.4 (¥24.5) on 59API. At official list price, $10 (about ¥72), so 59API is about 34% of list price.

These are hypothetical workloads, but the pattern is the one you will feel in practice: output tokens carry the highest rate, so a model that answers briefly and precisely is cheaper than one that writes essays. A five-million-token Haiku session still lands at $3.4 (¥24.5), a number you can budget rather than negotiate.

The same logic applies to non-Claude work. GPT-6.1 Sol (gpt-6.1-sol) charges $0.1847 (¥1.33) input and $0.9264 (¥6.67) output per 1 million tokens, with cache read at $0.0375 (¥0.27). For 5M input + 1M output tokens without cache, that is $1.85 (¥13.32) on 59API, against $20 (about ¥144) at official list price, so about 9% of list price. If a task does not need Claude, the gap between families is worth checking before you route it.

Non-Claude options for routine daily traffic

Some of your daily work does not need a Claude model at all. For code completion, script drafts, summarisation and bulk transformations, several non-Claude models on 59API sit at the low end of the rate sheet. Here is a short, practical shortlist, all per 1 million tokens:

• MiniMax M3 (MiniMax-M3): input $0.1389 (¥1), output $0.5556 (¥4), cache read $0.0278 (¥0.2).
• Qwen3.7 Plus (qwen3.7-plus): input $0.1389 (¥1), output $0.4167 (¥3), cache read $0.0417 (¥0.3).
• MiMo V2.6 Pro (MiMo-V2.6-Pro): input $0.2083 (¥1.5), output $0.4167 (¥3), cache read $0.002778 (¥0.02).
• DeepSeek V4 Flash (deepseek-v4-flash): input $0.2778 (¥2), output $1.11 (¥8), cache read $0.006944 (¥0.05).
• GPT-5.6 Sol (gpt-5.6-sol): input $0.3472 (¥2.5), output $1.94 (¥14), cache read $0.0278 (¥0.2).

Compare output rates directly, because that is where volume concentrates. Qwen3.7 Plus outputs at $0.4167 (¥3) and MiniMax M3 at $0.5556 (¥4), while Claude Haiku 4.5 outputs at $1.67 (¥12) and Claude Opus 5.5 at $7.22 (¥52). A sensible split is to send mechanical traffic to the low-rate models and reserve the Claude tier for prompts where its reasoning earns its place.

MiMo V2.6 Pro is also worth noting for repeated context: its cache read rate is $0.002778 (¥0.02) per 1 million tokens, the smallest figure in this group, which matters when the same large file set is read again and again.

Connecting your client to 59API

Choosing the economical model is only half the job; the other half is making sure the request reaches 59API through a client that supports that model family.

For Claude, use the genuine Claude Code or Claude Desktop client (Claude Code 2.1.285 or newer). Set ANTHROPIC_BASE_URL=https://api.59api.com and ANTHROPIC_AUTH_TOKEN=YOUR_59API_KEY. Claude models work only through those clients, never through the OpenAI SDK, cURL or other clients.

For GPT, DeepSeek and other non-Claude, non-Gemini models, Codex, OpenAI SDKs, cURL and OpenAI-compatible clients such as Cursor, Cherry Studio, Cline and OpenCode work with base_url https://api.59api.com/v1 and chat endpoint https://api.59api.com/v1/chat/completions.

Gemini is a separate path. Gemini models use Google's native protocol only, through Gemini CLI, and are not OpenAI-compatible; set GOOGLE_GEMINI_BASE_URL=https://api.59api.com and GEMINI_API_KEY=YOUR_59API_KEY. They do not work on /v1/chat/completions.

Cards are optional. A card has its own key, calls only 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, and quota and caps are card metering units written in $, not official list-price dollars. If you buy one, treat it as a separate budget line rather than a discount on your prepaid balance.

Start small: top up ¥10 or ¥50, point one client at 59API, and let the request log teach you which model earns its rate. That is a low-risk way to build a cost-aware habit.

Docs