On this page
Gonka Broker resells inference from Gonka, a decentralized GPU network, as a normal USD-priced API: you get a key, a base URL and an OpenAI-style endpoint, and never touch crypto. Every account also gets a free monthly tier. We signed up, verified, and called every model to see what the free tier really gives you.
What you get for free
| Model ID | Context | Max output | Price per 1M tokens (in = out) | Free tokens per month |
|---|---|---|---|---|
deepseek-ai/DeepSeek-V4-Flash-0731 | 400K | 16,384 | $0.20 | 1,250,000 |
zai-org/GLM-5.3-Flash | 400K | 16,384 | $0.20 | 1,250,000 |
MiniMaxAI/MiniMax-M2.7 | 204.8K | 16,384 | $0.25 | 1,000,000 |
BAAI/bge-m3 (embeddings) | 8K | — | $0.01 | — |
Context, output and price come from GET https://proxy.gonkabroker.com/v1/models, which answers without a key. The last column is what our dashboard showed: the balance is displayed per model, and switching the model selector changed it from 1,000,000 (MiniMax) to 1,250,000 (GLM). So the "at least 1M tokens of any model" on the website is really a fixed dollar amount — about $0.25 — converted at each model's price. Input and output tokens cost the same, which matters for long answers: most providers charge several times more for output.
The allowance resets on the 1st of each month. We verified on September 29 and still got the full amount, with a reset date of October 1 — it isn't pro-rated.
The phone verification step
Signing up takes an email or a Google account and no card. But right after signup the dashboard shows a balance of 0, a banner saying "Verify your phone to keep your free monthly tokens", and any API key you create is marked BLOCKED — hovering over it says "Blocked due to insufficient balance". The verification dialog explains: "we only use the number to confirm you are a real person. One free tier per number."
- Which countries work isn't published. The web app only checks that the number is in international format (
+and 7–15 digits); whether a code is actually sent is decided on the server. A mainland China (+86) number was rejected in our test. - Once verified, the balance filled immediately and the same key switched to ACTIVE — no new key needed.
The free-tier rate limit: 6 requests per minute
The key page lists "600 RPM", but that is the paid limit. On the free tier, a few requests in quick succession returned HTTP 429:
{"error":{"message":"Free tier rate limit (6 RPM). Top up for full speed: https://app.gonkabroker.com/?topup=1"}}
The limit is per account, not per model — calls to different models count together. Six a minute is fine for a personal tool, a chatbot for yourself, or trying models out; it rules out agents that fire several calls per step, or anything with more than one user. Space requests at least 10 seconds apart and retry on 429.
How the three models behave
Measured September 30, 2026 with a direct connection from Asia, one streamed request per model asking for a ~300-word explanation. Latency from a US or European machine will differ; generation speed is mostly on the server side.
| Model | First token | Generation | Tool calling | Notes |
|---|---|---|---|---|
| DeepSeek V4 Flash | 2.5 s | ~30 tok/s | Works | Answers directly, no visible reasoning |
| MiniMax M2.7 | 1.3 s | ~22 tok/s | Listed as supported (not tested) | Reasons first; the reasoning comes back in a separate reasoning_content field |
| GLM-5.3-Flash | 51–77 s | fast once it starts | Works (one 502 before it did) | Nothing streamed for about a minute on all three runs, then the whole answer arrived in 6–15 s |
GLM-5.3-Flash's long silence looks like queueing on the network rather than slow generation: once tokens start, 1,500 arrive in seconds. It's usable for background jobs but not for a chat box. Its first tool-calling request failed with 502 "inference: no non-probe attempt finished" and the retry worked, so wrap calls in a retry. For interactive use, DeepSeek V4 Flash was the most predictable of the three.
For scale: all of our testing — about 20 requests, including several 300-word answers and three runs that used the full 1,500-token output limit — took 8,254 of the 1,000,000 MiniMax-equivalent tokens, under 1% of the month.
Use it with the OpenAI SDK
from openai import OpenAI
client = OpenAI(
api_key="gnk-prx-...", # from app.gonkabroker.com → API Keys
base_url="https://proxy.gonkabroker.com/v1",
)
resp = client.chat.completions.create(
model="deepseek-ai/DeepSeek-V4-Flash-0731",
messages=[{"role": "user", "content": "Explain rate limiting in two sentences."}],
)
print(resp.choices[0].message.content)
Or with the Anthropic Messages API
The same key works on an Anthropic-style endpoint. We sent a Messages request to https://proxy.gonkabroker.com/v1/messages with the key in the x-api-key header, and GLM-5.3-Flash answered with a thinking block followed by a text block, in Anthropic's format:
import anthropic
client = anthropic.Anthropic(
api_key="gnk-prx-...",
base_url="https://proxy.gonkabroker.com", # the SDK appends /v1/messages
)
msg = client.messages.create(
model="zai-org/GLM-5.3-Flash",
max_tokens=1024,
messages=[{"role": "user", "content": "Say OK."}],
)
print(msg.content[-1].text)
Gonka Broker advertises this for running Claude Code on open models. We tested the Messages endpoint itself, not a full Claude Code session — and with 6 requests a minute, an agent like Claude Code will hit the free-tier limit quickly.
Gonka Broker vs other free LLM APIs
| Gonka Broker | Groq | OpenRouter free models | |
|---|---|---|---|
| Card | No | No | No |
| Phone verification | Yes | No | No |
| Free allowance | ~$0.25/month (1–1.25M tokens) | 1,000 requests/day, 200K tokens/day per model | 50 requests/day (1,000 after a $10 top-up) |
| Rate limit | 6 RPM per account | 30 RPM per model | 20 RPM |
| Models | DeepSeek V4 Flash, GLM-5.3-Flash, MiniMax M2.7 | gpt-oss-120b/20b, Qwen 3.8 27B | 8 of 16 :free models answered on Sept 29 |
| Context | Up to 400K | 131K | Varies by model |
Where Gonka Broker fits: it's a way to call DeepSeek V4 Flash for free — DeepSeek's own API requires a top-up — and to get a 400K context window, which none of the other no-card options offer. If you just need a fast free model for a chat app or an agent, Groq gives you far more requests with no phone step.
FAQ
Is Gonka Broker's free tier really free?
Yes, with no card — but only after phone verification. Until you verify, the balance is 0 and your key is blocked. After verification you get roughly $0.25 of usage per month, reset on the 1st.
Why is my Gonka Broker API key "BLOCKED"?
Because the balance is zero. On a new account that almost always means the phone number isn't verified yet: verify it in the dashboard and the same key becomes active. On a verified account it means you've used the month's free tokens.
Can I verify with a Chinese (+86) phone number?
Not in our test on September 29, 2026 — the +86 number was rejected. Gonka Broker doesn't publish which countries are supported.
What happens when the free tokens run out?
Requests return an error with a top-up link until the tier resets on the 1st of the next month. If you have a paid balance, it's used automatically after the free tokens.
Which Gonka model should I use?
DeepSeek V4 Flash for anything interactive — it started answering in 2.5 s and supports tool calling. GLM-5.3-Flash only for background work, since it sat silent for about a minute before answering in our tests.
Verdict
Gonka Broker's free tier is real and easy to use once you're through the phone check: an OpenAI- and Anthropic-compatible endpoint, a 400K context window, and DeepSeek V4 Flash at no cost. Its limits are tight — about $0.25 a month and 6 requests a minute — so treat it as a place to try these three models or run a small personal tool, not a backend for other users. If you can't verify a phone number in a supported country, look at Groq instead.