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

Victor Tarasov — Fri, 25 Sep 2026 14:47:07 +0000

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"}]
  }'

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)

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.

AMA session. Topic: AI Training Beyond the Data Center

Victor Tarasov — Wed, 15 Oct 2025 09:54:42 +0000

Join us for an AMA session on Tuesday, October 21, at 9 AM PST / 6 PM CET with special guest - Egor Shulgin, co-creator of Gonka.

Topic: AI Training Beyond the Data Center: Breaking the Communication Barrier

Discover how algorithms that "communicate less" are making it possible to train massive AI models over the internet, overcoming the bottleneck of slow networks.

We will explore:
🔹 The move from centralized data centers to globally distributed training.
🔹 How low-communication frameworks use federated optimization to train billion-parameter models on standard internet connections.
🔹 The breakthrough results: matching data-center performance while reducing communication by up to 500x.

Click the event link below to set a reminder!
https://discord.gg/DyDxDsP3Pd?event=1427265849223544863

DEV Community: Victor Tarasov