Why a pay-per-token AI gateway for Indonesian devs
Published 2026-07-08 · Nexotao
If you build with AI from Indonesia, you already know the tax. It isn't the model price — it's everything wrapped around it.
You want to try Claude Opus or Qwen3 Coder. So you reach for a credit card that charges in dollars, sign up for a plan sized for someone else's usage, buy seats you don't need, and hope your spend doesn't quietly balloon. For a solo dev, a student, or a two-person team in Jakarta or Bandung, that's a lot of friction to answer one question: is this model any good for my app?
Nexotao removes the wrapper. One Rupiah balance. 34 Phase 1 models. Pay per token. No plans, no seats, no credit card.
What “pay-per-token” actually means
You top up a balance in Rupiah — starting at Rp 10.000— and every call draws down that balance by the exact number of tokens you used, input and output billed separately, at each model's live per-token rate with no markup. Every Rp 1 of balance funds Rp 1 of usage.There's no “credit” abstraction sitting between your balance and your usage, no monthly minimum, and nothing expires on a billing cycle. Service and payment-method fees are shown separately when you top up.
Top up the way you already pay for things: QRIS, or crypto (USDT / USDC). No international card required.
Why this fits Indonesian devs specifically
- Rupiah in, tokens out.Prices are shown in Rupiah, per 1M tokens, right on the catalog. You're not doing FX math in your head or eating card conversion fees.
- No credit card gate. QRIS is how the country pays. Top up in about a minute and start calling models.
- Right-sized for real usage. A weekend project and a production workload use the same balance. You pay for what you send, not for a tier you have to grow into.
- Keep your tools. Nexotao is OpenAI-compatible — and speaks the Anthropicformat too. Point your existing SDK (or Claude Code) at Nexotao's base URL, drop in a Nexotao key, and nothing else in your codebase changes.
One catalog, lightweight to frontier
The point of one balance is that you're never locked to one model. Reach for a lightweight workhorse when the task is simple, and a frontier model when it's hard — same key, same balance, same API.
| Model | Input / 1M | Output / 1M | Good for |
|---|---|---|---|
| Claude Opus 4.6 | Rp 99.000 | Rp 495.000 | Flagship reasoning, 1M context, vision, and tools |
| Claude Sonnet 4.6 | Rp 59.400 | Rp 297.000 | Balanced quality, speed, vision, and tools |
| Qwen3 Coder Next | Rp 9.000 | Rp 21.600 | Agentic software development |
| Mistral Large 3 (675B) | Rp 9.000 | Rp 27.000 | General-purpose agentic workloads |
| Nova Micro | Rp 630 | Rp 2.520 | Fast, economical agentic tasks |
| GPT-OSS 120B | Rp 2.700 | Rp 10.800 | Open-weight agentic reasoning |
| DeepSeek V3.2 | Rp 11.160 | Rp 33.300 | Chat and text generation |
| GPT-OSS 20B | Rp 1.260 | Rp 5.400 | Low-cost chat workloads |
Notice the spread: Nova Micro to Claude Opus 4.6 is the difference between a lightweight task and a decision. With one balance you can route lightweight-by-default and escalate only when a task earns it — the single biggest lever on an AI bill.
Migration is minutes, not a project
You don't rewrite anything. If your code already talks to OpenAI or Anthropic, you change two lines — the base URL and the key — and keep every SDK, framework, and tool you already use. We wrote the 5-minute version: the quickstart →
Start where you are
Frontier AI shouldn't require a US credit card and a plan built for someone else. Top up a Rupiah balance, pick a model, ship.
Create a key and top up in about a minute.