New BroadNet Inference accounts start with $5 of pay-as-you-go credit and no card. BroadNet’s Terms of Service (§7.2) set two different periods: the free sign-up credit gives one person or organisation 30 days to try the service, while funds you pay in later are prepaid credit valid for 12 months after purchase. Before a top-up it pays for Gemma 4 VL 26B, the two transcription models and hourly keys: tens of millions of Gemma tokens or about 1,190 minutes of Arabic audio. Below: what it unlocks, how far it goes, and how to turn it into a cost forecast.
What does the starting credit cover?
From BroadNet’s pricing.json and plans page:
- Pay-as-you-go requests on models open from sign-up. BroadNet says no card is needed to start.
- Hourly keys for Gemma 4 VL 26B, at $0.50 each.
- Not the unlimited plans. Full Moon, Orbit and Meteor are purchased separately.
Which models open before a top-up?
The price file marks some models paid_only. The docs define it: those models open once the account has a top-up that refunds and disputes have not fully reversed. Until then, the trial credit reaches the other models.
| Model | Open on the starting credit | Price |
|---|---|---|
| Gemma 4 VL 26B | Yes | $0.10 input, $0.22 output per 1M |
| BroadNet Transcribe Arabic | Yes | $0.0042 per audio minute |
| BroadNet Transcribe (14 languages) | Yes | $0.002 per audio minute |
| DeepSeek V4.1 Flash | After first top-up | $0.175 input, $0.52 output per 1M |
| GLM 5.3 | After first top-up | $0.78 input, $3.46 output per 1M |
| Qwen3.8 27B | After first top-up | $0.175 input, $2.17 output per 1M |
So Gemma 4 VL 26B is the only chat model you can call without a top-up. It reads text and images, has a 32,768-token context, and does not list reasoning, which keeps output short and cheap.
Our open dataset of LLM API free tiers and rate limits compares what 14 providers publish about free access, starting rate limits, card requirements and training on free-tier data, with the source and date for each value. For paid-tier limits as well, see LLM API rate limits by provider.
How far does $5 go?
The arithmetic is simple: divide the credit by the price per unit.
In Gemma tokens
input-only: $5 ÷ $0.10 per 1M
output-only: $5 ÷ $0.22 per 1M
At the prices published on 2026-11-14, that is about 50 million input tokens if you sent nothing but prompts, or about 23 million output tokens if you only generated. Real traffic is a mix, so expect somewhere in between: tens of millions of tokens.
What does that look like in practice? If a typical chat turn sends 2,000 tokens and gets 300 back, the credit covers about 18,000 turns. If you send receipts or screenshots, each image counts as input tokens, and a large image uses many of them, so measure your own images first; here is how.
In audio minutes
BroadNet Transcribe Arabic costs $0.0042 per minute of audio, so the starting credit covers about 1,190 minutes of audio: about that many one-minute voice notes. BroadNet’s docs do not say how partial minutes are billed, so test a few very short clips before you plan around them. The 14-language BroadNet Transcribe model costs $0.002 per minute, so the same credit goes further there. See how to transcribe WhatsApp voice notes.
In hourly keys
An hourly key costs $0.50 and gives one hour of Gemma 4 VL 26B with one stream, 120 requests per minute and 300,000 tokens per minute, with no token meter. The credit can buy several of them. Hourly keys suit a demo, a workshop or a one-off batch job; when an hourly key fits.
How should you spend free API credit?
The worst use of free credit is a pile of “hello” prompts. The best use is a small, honest copy of your real workload, logged.
- Pick your real task. Classification, extraction, a support reply, a transcript. Use your real prompts and real inputs.
- Log usage on every call. Record
prompt_tokens,completion_tokensand, for audio, the length of each file. The Python SDK guide covers the setup. - Run a day’s worth of traffic. A few hundred representative requests are enough to get stable averages.
- Forecast. Multiply the averages by your expected monthly volume and by each model’s price. The monthly cost method walks through it.
- Decide what to top up for. If Gemma is good enough, you already know your cost. If you need long context, reasoning or tools, the forecast tells you what DeepSeek V4.1 Flash, GLM 5.3 or Qwen3.8 27B will cost after a top-up.
What does a good first day on the credit look like?
Here is one way to use the credit in a single afternoon without wasting it:
- First request. Follow the Python SDK guide and send one short prompt to
broadnet/gemma-4-26b-a4bathttps://inference.broadnet.ai/v1. Confirm you can readchoices[0].message.contentand theusageblock. - Quality check. Run 50 real inputs from your product: support tickets, product photos, form fields. Read the answers yourself and note the failures.
- Token check. Sum the usage from those 50 calls and divide by 50. That is your tokens per request on Gemma.
- Audio check, if relevant. Transcribe ten real voice notes or call snippets with
languageset correctly, and compare the transcripts with what was said. The Arabic speech-to-text guide shows the request. - Write down the numbers. Tokens per request, minutes per file and failure rate are what you need for a forecast. You will likely have used only a small part of the credit.
How do you make the credit last longer?
- Set
max_tokens. A finite limit stops a runaway answer from eating the credit. - Ask for short, structured output. Output costs more per token than input on every BroadNet chat model. Check prompt sizes first with the free token counter.
- Shrink images. Resize and crop before you send; image tokens are input tokens.
- Keep prompts stable. On models with a cached price, repeated prefixes are cheaper. Gemma has no cached rate, so on Gemma just keep prompts short.
- Convert audio once. Send mono WAV at a sensible sample rate; BroadNet’s docs show a one-line
ffmpegcommand for WhatsApp.oggfiles.
What can’t the free credit tell you?
The starting credit is enough to test quality and measure tokens. It is not enough to test what happens at production volume: pay-as-you-go keys are limited to 30 requests a minute and 4 in flight, and the paid models behave differently from Gemma on long or reasoning-heavy tasks. If your plan is to run a coding agent all day, the useful test is a $1 Meteor hour on DeepSeek V4.1 Flash after a top-up, not more Gemma tokens. Compare the options on the pricing page.
How does it compare with other free LLM API tiers?
Free tiers differ in what they open, how long they last and what happens to your data. What the official pages said when we checked on 14 November 2026:
| Provider | What is free | Main catch |
|---|---|---|
| BroadNet Inference | $5 of pay-as-you-go credit, no card | Only Gemma 4 VL 26B, transcription and hourly keys before a top-up |
| Google Gemini API | Free input and output tokens on several Gemini models, including Gemini 3.8 Flash and Gemini 3.5 Flash-Lite | Google’s free plan says “Content used to improve our products” (paid-tier data terms apply in the EEA, UK and Switzerland); free-tier rate limits apply |
| OpenRouter | Free variants of some models (:free) |
20 requests a minute and 50 a day, or 1,000 a day once you have bought at least 10 credits (USD 10; OpenRouter credits are in US dollars) |
| Groq | A free plan | Limits are set per account and shown in the Groq console; the public page lists only paid-plan numbers |
Gemini’s free tier and OpenRouter’s :free variants cost nothing per token and have no dollar cap, so both suit experiments. The trade-offs are data use and rate limits, so check the Gemini 3.8 Flash price page and the OpenRouter alternative page before you build on either. BroadNet serves all models with zero data retention, including on the free credit.
