DEV Community

Victor Tarasov
Victor Tarasov

Posted on

Four OpenAI-compatible endpoints are giving away free tokens right now

#ai #api #llm #opensource

Disclosure: I work on Gonka. Four independent brokers are running the free-token offer described below, and I've tried to keep the technical details accurate enough to be useful even if you skip the offer entirely.

If you build with LLMs, you've probably hit the same wall I have: you want to try a long-context model on a real workload, and the only honest way to find out how it behaves is to throw a few million tokens at it. That's expensive to do on a whim.

Right now there are four independent endpoints handing out free tokens, all OpenAI-compatible, so trying them costs you a base URL change and nothing else. Here's what's actually on offer and how to wire it up.

The four endpoints

Broker Welcome bonus Base host Docs
DAHL 100M tokens inference.dahl.global /docs/
Gonka GG 1M tokens proxy.gonka.gg /docs
Gonka Router $20 in credits gonkarouter.io /docs
Gonka API 10M tokens gonka-api.org /dashboard

These are model tokens, not a cryptocurrency. There's no wallet anywhere in the flow — you sign up, you get a key.

Each broker is a separate business. They set their own rate limits, data policy and pricing once the bonus runs out, so read the one you pick rather than assuming they match.

Models

All four serve the same three models:

  • deepseek-ai/DeepSeek-V4-Flash-0731
  • zai-org/GLM-5.3-Flash
  • MiniMaxAI/MiniMax-M2.7

Context limits are worth checking per broker rather than trusting a spec sheet. On Gonka GG, the published limits are 400K context / 16K output for the two Flash models and 180K / 16K for MiniMax. The DeepSeek and GLM models are served in the 380–400K range depending on who you're pointed at — if your workload actually depends on the top of that window, measure it on your key before you design around it.

Wiring it up

Every one of these speaks the OpenAI chat-completions API, so it's a base URL and a key:

curl https://BROKER_HOST/v1/chat/completions \
  -H "Authorization: Bearer $KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-ai/DeepSeek-V4-Flash-0731",
    "messages": [{"role": "user", "content": "Say hi"}]
  }'
Enter fullscreen mode Exit fullscreen mode

Python, with the official SDK:

from openai import OpenAI

client = OpenAI(
    base_url="https://BROKER_HOST/v1",
    api_key="YOUR_KEY",
)

resp = client.chat.completions.create(
    model="zai-org/GLM-5.3-Flash",
    messages=[{"role": "user", "content": "Summarise this repo"}],
)
print(resp.choices[0].message.content)
Enter fullscreen mode Exit fullscreen mode

The same two values drop into anything that takes a custom OpenAI-compatible provider — LibreChat, Cline, Roo, opencode, LiteLLM, Continue. No code changes, just config.

What to spend it on

A bonus is only useful if you learn something from it. Some things worth more than a chat demo:

  • Long-context retrieval. Put 300K tokens of your own codebase or docs in and ask questions whose answers sit in the middle. Models degrade very differently across that window.
  • Agent loops. These burn tokens fast, which makes them the honest test. A 1M bonus disappears in one enthusiastic afternoon; a 100M one supports a real evaluation.
  • Structured output under load. Ask for JSON a few thousand times and count the malformed responses.

If you've got 100M tokens to spend, that's enough to run a proper eval rather than a vibe check, which is the thing free tiers usually can't buy you.

Where the compute comes from

Worth knowing what you're pointed at: these brokers front Gonka, a decentralized inference network. Instead of one company's GPUs, independent operators serve requests, and correctness is kept honest by re-running a random sample of tasks — somewhere between 1% and 10% — and scoring operators on reputation over time. It's a Go codebase built on a fork of the Cosmos SDK.

For your purposes as a caller none of that matters: it's an HTTPS endpoint that speaks OpenAI. But it's the reason the token grants can be this large.

Getting a key

All four are listed at aidrop.gnk.space with sign-up links. Pick whichever bonus fits what you want to measure — 1M for a quick look, 10M or 100M for an actual evaluation.

If you hit something odd with keys, limits, or a model behaving differently than documented, each broker has its own channel on the Gonka Discord and the people running them read it.

Happy to answer questions in the comments — including sceptical ones about the verification claims above.

Top comments (0)