An hourly key is a prepaid API key that works for one hour at a fixed price, with no token meter. On BroadNet Inference it costs $0.50 for one hour of Gemma 4 VL 26B on one stream, capped at 120 requests and 300,000 tokens a minute. Below: who it suits, when it beats per-token billing, and how to use it.
What is an hourly LLM API key?
A prepaid API key for one hour of one model at a flat price. On BroadNet Inference:
| Detail | Hourly key |
|---|---|
| Price | $0.50 per hour |
| Model | Gemma 4 VL 26B (broadnet/gemma-4-26b-a4b) |
| Input | Text and images |
| Context | 32,768 tokens |
| Streams | One at a time |
| Requests a minute | 120 |
| Tokens a minute | 300,000 |
| Paid from | Account balance, including the $5 starting credit |
There is no token meter inside the hour. The limits that apply are concurrency and the rate limits above.
Who is the hourly key for?
- Live demos and talks. Hand a key to the demo laptop. Whatever happens on stage, the cost is fixed.
- Workshops and hackathons. Budget each hour in advance.
- Kiosks and event apps that run for a known window.
- Short batch jobs. Captioning or labelling a folder of images, classifying a list of tickets.
- Trying the API. The starting credit covers hourly keys, so you can test with no card.
It is not for a production app with many simultaneous users: one stream means one request at a time.
How does it compare with paying per token?
Gemma 4 VL 26B per token costs $0.10 per 1M input and $0.22 per 1M output. Dividing the hourly price by those rates gives the break-even point for one hour, at prices synced 2026-11-14:
| If the hour is mostly | Break-even volume in the hour | Ceiling in the hour |
|---|---|---|
| Input (reading images, documents) | ($0.50 ÷ $0.10) × 1M input tokens | 300,000 tokens a minute × 60 |
| Output (writing text) | ($0.50 ÷ $0.22) × 1M output tokens | 300,000 tokens a minute × 60 |
Below those volumes, per-token billing is cheaper. Above them, the hourly key is. The free flat-rate vs per-token calculator does the division for your own mix, and flat rate vs pay per token covers the general trade-off. A light demo that sends 50 short requests uses a tiny fraction of the break-even volume, so per token would cost less; you pay the flat price for certainty. A batch job that keeps the single stream busy with long image prompts can pass the break-even point.
Whether one stream reaches the input break-even in an hour depends on how quickly each request finishes. Measure one representative request on a per-token key, then multiply out.
Three example hours, priced both ways:
| Hour of work | Tokens in the hour | Per token | Hourly key |
|---|---|---|---|
| Conference demo: 60 questions of 1,000 tokens in, 300 out | 0.06M in, 0.018M out | A small fraction of $0.10 | $0.50 |
| Image labelling: 3,000 images at 1,000 tokens, 50-token labels | 3M in, 0.15M out | 3 × $0.10 plus 0.15 × $0.22 | $0.50 |
| Long-prompt batch: 300 requests of 20,000 tokens in, 500 out | 6M in, 0.15M out | 6 × $0.10 plus 0.15 × $0.22 | $0.50 |
The demo is cheaper per token; you would buy the key for the fixed cost, not the saving. Work out the other two rows at today’s prices: the labelling job stays below break-even, and the long-prompt batch passes it, provided one stream can finish 300 requests within the hour and stays under 300,000 tokens a minute; 6.15M tokens in 60 minutes is about 100,000 a minute, which fits.
How does it compare with the flat plans?
| Option | Price | Model | Concurrency | Best for |
|---|---|---|---|---|
| Hourly key | $0.50 an hour | Gemma 4 VL 26B | One stream | Demos, short batches, cheap vision |
| Meteor | $1 an hour | DeepSeek V4.1 Flash | One request per session | Testing an agent on the unlimited model |
| Orbit | $15 a week | DeepSeek V4.1 Flash | One request per session | A one-week project |
| Full Moon | $60 a month | DeepSeek V4.1 Flash | One request per session | Agents every day |
Running hourly keys around the clock for 30 days would cost $360.00, well above a Full Moon session at $60. For continuous use, a plan is the better buy. The plans also use a different model, with a 1,048,576-token context and reasoning. How plan sessions work. If your project lasts days rather than hours, see when a weekly plan fits a short project.
Why does a flat hourly price help with spend control?
A per-token key has no natural ceiling: a stuck loop or a leaked key keeps spending until something stops it. An hourly key’s worst case is its price. That makes it a good choice for keys that leave your servers, such as a key on a demo laptop or a workshop handout. Combine it with an IP allow-list where you can. More spend controls.
How do you use an hourly key?
The key is a normal OpenAI-compatible key. Point your client at https://inference.broadnet.ai/v1 and set the model:
curl https://inference.broadnet.ai/v1/chat/completions \
-H "Authorization: Bearer $BROADNET_HOURLY_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"broadnet/gemma-4-26b-a4b","messages":[{"role":"user","content":"Write a one-line caption for a product launch."}]}'
Practical tips:
- Queue requests. With one stream, send requests one after another. A second request while one is in flight can be refused; back off and retry. Errors and retries. To check that a batch fits 120 requests and 300,000 tokens a minute, use the rate limit calculator.
- Stay under 300,000 tokens a minute. Long image prompts add up fast.
- Keep prompts within 32,768 tokens, prompt and answer together.
- Have the work ready. Prepare inputs and scripts before you buy, so the hour goes on requests, not setup.
Where BroadNet Inference fits
The hourly key is BroadNet’s simplest way to cap the cost of a short job: one price, one model, one hour. It is a poor fit for many simultaneous users, for long documents beyond 32,768 tokens, or for tasks that need reasoning; use pay-as-you-go or a plan for those. See hourly key and model prices.

