DeepSeek API 402 "Insufficient Balance": whose balance, how to tell, and what to do

What the error looks like

HTTP/1.1 402 Payment Required
{"error": {"message": "Insufficient Balance", ...}}

# the same error as tools surface it:
AI_APICallError: Insufficient Balance                      (OpenCode, Vercel AI SDK apps)
openai.APIStatusError: Error code: 402 - {'error': {'message': 'Insufficient Balance', ...}}   (Python SDK)

The message is the only stable part of the body; the other fields differ between DeepSeek's OpenAI-format and Anthropic-format endpoints. The status is 402, which the OpenAI SDKs do not map to a named exception: Python raises a plain APIStatusError with status_code == 402, and the SDK does not retry it. Frameworks that classify errors by name often file it under "unknown" or "malformed request", which is how it ends up logged as something it is not.

Whose balance is empty

The tell for the reseller case is timing. If the error started for you at the same minute it started for everyone else, and the app's own paid models fail alongside the free ones, it is the reseller's account. On July 3, 2026 OpenCode's free DeepSeek V4 Flash returned Insufficient Balance for every user at once until a maintainer posted "should be working now"; in May 2026, 11 of the 12 OpenCode Go models did the same for sub-agent calls.

Fix 1: if it is your account, add funds

Sign in to the DeepSeek platform, open the balance page and top up. Requests start succeeding as soon as the payment posts. If you are not sure the key is yours, this check settles it:

curl -s https://api.deepseek.com/chat/completions -H "Authorization: Bearer $DEEPSEEK_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"deepseek-flash","messages":[{"role":"user","content":"Say ok"}],"max_tokens":5}' -w "\nHTTP %{http_code}\n"
# 402 Insufficient Balance -> add funds.   401 Authentication Fails -> the key is wrong.   Both are final; neither is fixed by retrying.

Two habits prevent the next one. Set a balance alert if your provider offers one, and treat 402 in code as a stop signal: no retry loop, no backoff, an alert to a human.

Fix 2: if it is a reseller's account, bring your own provider

You cannot top up someone else's DeepSeek account, and their "free" model will come back when they do. In the meantime, most of these tools accept a provider you control. OpenCode takes any OpenAI-compatible endpoint in opencode.json (our guide, with a config block that puts DeepSeek V4 on US GPUs next to Go); Claude Code, Codex and the others have equivalents on the integrations page. With your own provider, the only balance that can run out is yours, and you will know about it first.

Fix 3: fail over to the same model on another balance

DeepSeek V4 Flash and V4 Pro are open-weight models, so a second endpoint can serve the same model on a different account. Since 402 is final, the fallback should happen on the first one:

import os
from openai import OpenAI, APIStatusError

deepseek = OpenAI(base_url="https://api.deepseek.com", api_key=os.environ["DEEPSEEK_API_KEY"])
backup   = OpenAI(base_url="https://api.inferenceapis.com/v1", api_key=os.environ["INFERENCEAPIS_API_KEY"])

def chat(messages, **kw):
    try:
        return deepseek.chat.completions.create(model="deepseek-flash", messages=messages, **kw)
    except APIStatusError as e:
        if e.status_code != 402:
            raise
        # Out of prepaid balance at DeepSeek. Retrying will not help; the same open-weight model elsewhere will.
        return backup.chat.completions.create(model="deepseek-flash", messages=messages, **kw)

The ids deepseek-flash, deepseek-v4-pro, deepseek-chat and deepseek-reasoner are all accepted here, so the model string does not change. New accounts get $1 of starter credit once the email address is confirmed, which covers a few hundred Flash requests, and there is no subscription; when the balance runs out, our own 402 says so and links the billing page.

What it costs, side by side

DeepSeek's figures are from its pricing page on September 21, 2026 and change by time of day; ours are flat. On price, DeepSeek off-peak wins. What you are paying for elsewhere is a balance nobody else can drain, a US location, zero retention (/trust), and no concurrency cap (the 429 entry).

Frequently asked questions

I have an active OpenCode Go subscription. Why does a model say Insufficient Balance?

Your Go subscription is a quota inside OpenCode. The model behind it runs on OpenCode's own DeepSeek account, and when that account is empty the request fails regardless of your quota. The reports on GitHub (#27598, #28129, #35142) all show this shape: active subscription, 402 from the provider, fixed when OpenCode topped up.

Does DeepSeek have a free tier?

No. The API is prepaid. Products that offer "free DeepSeek" are paying DeepSeek themselves, which is why their free models are the first to return this error when their account runs low.

Will retrying with backoff eventually work?

Not until funds are added. 402 is not a capacity or rate error. Retry loops that treat it like a 429 burn time and, in agent frameworks, can kill sub-agents that would have succeeded on another provider.

Is the DeepSeek model on Inference APIs the same one?

Yes: the public DeepSeek V4 Flash (0731), V4.1 Flash and V4 Pro (0813) weights on Together AI GPUs in the United States. Thinking is on by default as at DeepSeek and off with reasoning_effort: "none".

Where Inference APIs fits

The fallback in fix 3 is what we sell: DeepSeek V4 Flash, V4.1 Flash and V4 Pro on GPUs in the United States, billed per token from a balance that is yours alone. The model ids you already send work unchanged, and a new account starts with $1 of credit. When it runs out you get our own 402 insufficient_balance, with a link to add funds, and nobody else's usage can empty it. We are not cheaper than DeepSeek off-peak; the table above shows both.

Related entries

• DeepSeek API 429 "Rate Limit Reached" and 503 "Server Overloaded": the concurrency cap behind both, and your options
• DeepSeek API 400 "Model Not Exist": deepseek-chat and deepseek-reasoner were retired
• OpenRouter 402: "Insufficient credits" and "Free BYOK request cap exceeded for this month"
• OpenCode Go 403 RegionError: "only available hosted in China and requires explicit opt in"