“Unlimited” is the most stretched word in AI pricing. Some plans are unlimited messages that reset every few hours. Others are unlimited until a weekly ceiling. For a chat user that may never matter. For an agent that runs all day, it decides whether the plan works at all.
What kinds of “unlimited” plan are there?
Most offers fall into one of three shapes:
- Chat subscriptions. Unlimited or generous messages in an app, refilled on rolling windows. There is usually no API key, so your own agent can’t use it.
- “Unlimited tokens” with request caps. No token meter, but a cap on requests per day or per minute. A busy agent hits the request cap instead.
- Flat-rate by capacity. No meter and no cap on totals, but a fixed number of requests in flight. You buy more capacity to go faster.
Each can be honest. The trouble starts when the word “unlimited” hides which shape you are buying.
Which five limits should you look for?
| Limit | What it means | Why it hurts agents |
|---|---|---|
| Token meter | Usage is counted and billed or capped by tokens | Long contexts and retries burn through it |
| Rolling window | A budget that refills every few hours | Agents stall mid-task until the window resets |
| Weekly or monthly cap | A ceiling on total use per period | Heavy weeks hit it, often without warning |
| Rate limit | Requests per minute or tokens per minute | Bursty tool loops get 429 errors |
| Concurrency | Requests in flight at once | Parallel sub-agents queue |
A plan can be unlimited on one of these and limited on the rest. Read the terms with this table next to you.
What seven questions should you ask?
- Is there a token count anywhere? Even a “soft” one that slows you down?
- Is there a rolling window? Many chat subscriptions, including coding-assistant plans, refill on rolling windows of a few hours.
- Is there a weekly or monthly cap?
- How many requests can run at once?
- What are the requests-per-minute and tokens-per-minute limits?
- Which models are included? “Unlimited” often covers one model, not the catalogue.
- Is it an API key or a chat app? A subscription to a chat app is not an API key you can point your own tools at.
Where do you find the real limits?
Marketing pages rarely list them. Look in three places:
- The API reference, usually under “limits” or “rate limits”. This is where requests per minute and concurrency live.
- The terms of service and acceptable use policy. Fair-use clauses, sharing rules and what happens on heavy use.
- Response headers and error bodies. A 429 response often says which limit you hit. Log it.
If a limit is not written anywhere, ask support and keep the answer in writing.
How do BroadNet’s plans answer them?
| Question | Full Moon, Orbit and Meteor |
|---|---|
| Token meter | None |
| Rolling window | None |
| Weekly or monthly cap | None |
| Requests at once | One per session; buy one session per parallel agent |
| Model | DeepSeek V4.1 Flash, 1M-token context |
| API key | Yes, OpenAI-compatible at https://inference.broadnet.ai/v1 |
| Price per session | $60 a month, $15 a week, $1 an hour |
So the trade is clear: you give up parallelism inside one session, and in return there is no meter at all. If you need four requests at once, you buy four sessions. How sessions work and how to size them for parallel agents.
For comparison, BroadNet’s metered keys have different limits:
| Key type | Billing | Limits |
|---|---|---|
| Plan session | Flat, per session | One request at a time |
| Pay-as-you-go key | Per token or audio minute | 30 requests a minute, 4 in flight |
| Hourly key (Gemma 4 VL 26B) | $0.50 an hour | 120 requests a minute, one stream, 300,000 tokens a minute |
When a metered key hits a limit it returns 429; back off and retry. To see how limits translate into throughput for your workload, try the rate limit calculator.
How do other flat-rate LLM APIs compare?
Several providers sell a flat monthly LLM API. Their limits sit in different places, and several are cheaper than Full Moon or allow more than one request at a time. Values below are from each official pricing page, checked 2026-11-14; prices in US dollars per month.
| Provider and plan | Price | Requests at once | Other limits | Models |
|---|---|---|---|---|
| ArliAI Core / Plus / Pro / Max | USD 15 / 20 / 30 / 160 | 1 / 1 / 2 / 6 | Context 32K / 128K / 256K / 512K; no published rate limit | Open models up to 552B |
| Infermatic Essential / Standard / Plus / Pro | USD 9 / 16 / 20 / 49 | 1 / 1 / 2 / 3 | 12 to 24 requests a minute; unquantified fair-use policy | 10 to 17 open models |
| Awan LLM Core / Plus / Pro / Max | USD 5 / 10 / 20 / 80 | “Parallel requests” on Pro and Max, no number | 20 to 200 requests a minute; daily request caps per model size below Max | Small, medium and large open models |
| Standard Compute Starter / Economy / Standard / Pro | USD 19 / 39 / 89 / 249 | One agent on Starter; parallel above, no number | No daily or weekly caps, but a monthly compute budget per plan; no overages, requests stop at the budget (Economy and up can top up at usage + 5%; checked 2026-11-15) | Closed and open models, including GLM, Kimi, DeepSeek and Qwen |
| BroadNet Full Moon | $60 | 1 per session | No meter, window or cap | DeepSeek V4.1 Flash, 1M context |
Where they beat BroadNet: ArliAI, Infermatic and Awan all start below $60, ArliAI and Infermatic sell two or more parallel requests on one plan (six on ArliAI Max), all four offer a choice of models, and Standard Compute includes closed frontier models. Where BroadNet differs: no published request-per-minute limit, daily cap or compute budget on the plan, and a 1,048,576-token context on every request (ArliAI’s largest plan stops at 512K; Standard Compute also lists 1M-context models, including DeepSeek V4.1 Flash, but inside its compute budget). On data, ArliAI advertises “Zero-Log” and Infermatic “no logs of prompts or outputs”, so compare each provider’s written retention terms with BroadNet’s zero data retention statement rather than the headline. Featherless, often listed here, sells a USD 25 unlimited Chat plan (32K context, 4 concurrent units) that its pricing page marks as not for app or API traffic or background automation, and its Developer plan is billed per token. Read each provider’s terms against the five limits above before you buy.
What is the physical ceiling of an unlimited session?
Even with no meter, a session can only produce tokens as fast as the model writes them. Typical speed is about 100 tokens per second (tokens per second). A session generating without a single pause for 30 days would produce about 259 million output tokens. Real agents spend time reading files, running tests and waiting, so the practical number is lower. That ceiling is per session: more sessions, more throughput.
What does an honest plan still limit?
No plan is a blank cheque:
- Speed. One session can only produce so much in a day, as above.
- Acceptable use. BroadNet’s plans follow its Acceptable Use Policy.
- Sharing. Each person uses their own account. You can build a plan into your product, but you cannot resell the key.
What are the pitfalls when buying an “unlimited” plan?
- Buying for the peak. Count the requests that overlap most of the time, not the worst minute, and queue the extra ones in your own code. BroadNet’s pages don’t say whether a request beyond your sessions waits or is refused.
- Assuming every model is included. Check the model list before you wire your tools to it.
- Testing with a chat window. Test with your real agent loop, because agents send far more and far longer requests than people do.
- Ignoring the exit. Check how renewal and cancellation work. On BroadNet, monthly and weekly renewal is managed in the console, and paid access continues to the end of the period.
When is “unlimited” the wrong choice?
If your monthly usage is small, a flat plan costs more than paying per token. In BroadNet’s worked example, a side project costs $14.23 a month on pay-as-you-go against $60 for Full Moon. See the break-even math or try your own numbers in the flat-rate vs per-token calculator.
Where BroadNet Inference fits
BroadNet’s plans are the third shape: no meter, no windows, no caps, and one request at a time per session. That suits agents that run for hours on one model. It doesn’t suit you if you need many different models on a flat price, or if your usage is light enough that per-token billing is cheaper. A $1 Meteor hour is the cheapest way to test it on a real task; the Aider guide shows a terminal agent set up in minutes.
