GPT and Claude API access in Vietnam, without the card fight.

Quick answer: one base URL, one key, models by id.

Set the base URL to https://api.apisrouter.com/v1, authenticate with an APIsRouter key, and address models by their exact ids: claude-sonnet-4-6 and claude-haiku-4-5-20251001 for Claude, gpt-5.5 for the GPT flagship, deepseek-v4-flash for the volume tier. The endpoint speaks the standard OpenAI Chat Completions dialect, so the OpenAI SDK, LangChain, and every tool with a custom base URL field work unchanged. For Claude-first workflows there is a second surface: the gateway also exposes an Anthropic-compatible /v1/messages endpoint, which means the Anthropic SDK and Claude Code work by swapping the base URL and key. No card attaches at any point; the balance is topped up when you choose, and the free start lets you test before paying.

from openai import OpenAI

client = OpenAI(
    api_key="sk-APIsRouter-...",
    base_url="https://api.apisrouter.com/v1",
)

response = client.chat.completions.create(
    model="claude-sonnet-4-6",
    messages=[{"role": "user", "content": "Xin chào"}],
)

The actual vendor status for Vietnam, verified.

As of July 2026, Vietnam appears on OpenAI's supported countries list for the API and on Anthropic's supported regions list. Google's Gemini API is also available, including the card-free free tier in AI Studio. Older forum posts describing Vietnam as unsupported reflect an earlier state; both vendors expanded their lists over the past two years. Lists change, so check the vendors' official pages before making commitments, but today the door is open. What has not changed is the billing mechanics behind that door. Vendor API usage bills in USD against an international card network. There is no VND pricing, no local acquiring, and no wallet integration, and that is where most Vietnamese developer setups stall.

Where payment actually fails for Vietnamese developers.

None of this is specific to AI vendors; it is the general shape of Vietnamese cards meeting US online billing. But AI APIs are usage-billed, so the card is not a one-time hurdle: it stays attached, gets charged repeatedly, and every renewal is another chance to decline. A prepaid top-up balance removes the recurring failure mode along with the initial one.

• Domestic-scheme cards do not work: cards on the domestic Napas network are not usable at a US online biller; only internationally enabled Visa, Mastercard, or similar scheme cards have a chance.
• International online payments are often off by default: many Vietnamese banks require explicitly enabling cross-border online transactions per card, and some decline US merchant categories even then.
• Wallets are not accepted: MoMo, ZaloPay, and VNPay cover daily life in Vietnam and none of them attach to an OpenAI or Anthropic billing account.
• Prepaid and virtual cards are unreliable: OpenAI rejects prepaid cards outright, and virtual card acceptance varies without a documented rule.
• Generic decline messages: when the issuing bank blocks the charge, the vendor checkout reports a generic failure, so debugging means calling the bank, not reading an error code.

One endpoint in front of every vendor.

APIsRouter is a gateway: one base URL, one key, and per-request routing to the vendor that serves the requested model id. For a developer in Vietnam that collapses the payment problem to a single top-up and collapses the integration problem to a single dialect. It also makes vendor mixing trivial. Claude for the coding agent, DeepSeek for the bulk translation pass, Gemini for image inputs: each is a model string on the same key, and the console attributes tokens and spend to each id. Tools popular with Vietnamese teams work as-is wherever they accept a custom base URL, from Cline and Continue in VS Code to Chatbox, LobeChat, and Dify.

Latency from Vietnam, stated plainly.

The gateway serves from cloud regions in the EU and US. From Hanoi or Ho Chi Minh City, that adds a transcontinental round trip to every request, a real fraction of a second before the model even starts generating. For the workloads most Vietnamese developers run through an endpoint like this, coding agents, batch content pipelines, translation jobs, evaluation harnesses, that overhead is small relative to generation time, and streaming masks most of what remains. For a realtime consumer chat app serving Vietnamese users at scale, a regional vendor endpoint will be faster where billing is workable; measure time-to-first-token from your production region against your latency budget and let the number decide.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

Model picks that make sense from Vietnam.

Per-key usage in the console shows tokens and spend per model, so the honest way to choose between the volume tiers is a one-week A/B on your real traffic rather than anyone's benchmark table.

• claude-sonnet-4-6 for coding agents and quality-sensitive product features; it is the id most tools default to for good reason.
• claude-haiku-4-5-20251001 where Claude behavior matters but volume dominates: support drafts, tagging, extraction.
• gpt-5.5 as the frontier comparison point; A/B it against Sonnet on your own tasks before standardizing.
• deepseek-v4-flash for Vietnamese-English translation passes and any high-message workload where the per-token rate is the whole story.
• glm-5 as a second volume-tier option worth testing on Vietnamese-language output quality against DeepSeek.

Verify the endpoint before wiring it into anything.

Two commands settle whether the setup works: list the models to confirm the key and see the addressable ids, then send one completion with the exact id you plan to use. If those pass, every remaining problem is in your tool's configuration, not the gateway. The common first-run failures map cleanly: 401 means the key is not reaching the request, model-not-found means the id string does not match the listing, and a 404 on every call means the base URL lost its /v1 suffix. The console's per-request log then shows model, tokens, and spend, which turns the first day of testing into a realistic monthly cost estimate.

curl -s https://api.apisrouter.com/v1/models \
  -H "Authorization: Bearer $APISROUTER_API_KEY" | head -50

FAQ