An agent that runs all day sends the same large context again and again. On a per-token router, every one of those tokens is billed. On a flat-rate plan, none of them are. This post works out when that difference matters, using OpenRouter’s own published fees and limits. OpenRouter and BroadNet Inference are different kinds of service, and many teams use both.
BroadRouter is run by BroadNet. Facts about OpenRouter below are quoted from its own documentation, checked 2026-11-14; check the linked pages for current terms. If you want every kind of OpenRouter alternative side by side, including other routers and self-hosted gateways, read the OpenRouter alternatives buyer’s table.
How do OpenRouter and a flat-rate provider differ?
| OpenRouter | BroadNet Inference | |
|---|---|---|
| What it is | An LLM router across many providers | A provider serving its models on US hardware or through US inference partners, per its privacy statement |
| Models | Hundreds, from many labs | Six: four chat and two speech-to-text (catalogue) |
| Billing | Per token, provider prices passed through | Per token, per audio minute, or flat unlimited plans |
| Fees | 5.5% (Standard) or 8% (Business) when buying credits, 5% crypto; fee on BYOK above a free allowance | None stated beyond published prices |
| Flat unlimited plan | No | DeepSeek V4.1 Flash: $60/month, $15/week, $1/hour per session |
| Provider fallback | Yes, on by default | No; one provider |
| Data retention | Configurable; can restrict to ZDR endpoints | No content retained on any request |
| Arabic speech to text | Not its focus | BroadNet Transcribe Arabic, $0.0042/min |
| API format | OpenAI-compatible | OpenAI-compatible |
How do OpenRouter’s fees work?
OpenRouter’s FAQ says there is “no markup on inference pricing”. The cost is in two other places:
- Buying credits. 5.5% on the Standard plan (USD 0.80 minimum per purchase, per the FAQ) and 8% on Business; crypto payments carry 5%. Every token you buy through credits carries that fee.
- Bring your own key (BYOK). The pricing page lists USD 25,000 of list-price inference a month with no fee, and 5% of what the same model would cost on OpenRouter above that.
For light use these fees are small. For an agent that spends heavily every month, they are a fixed percentage on top of a bill that already grows with tokens. Compare total monthly cost for your volume in the LLM API cost calculator, and see what every provider charges for the plan model on its own pricing page in our DeepSeek V4.1 Flash price table. Market-wide medians and provider spread are in the LLM API price index.
What does an always-on agent cost on each?
Take BroadNet’s worked example of a full-time coding agent: 1,000M input tokens a month, 80% read from cache, and 50M output. At BroadNet’s own per-token prices for DeepSeek V4.1 Flash that is $65.40 a month; an always-on agent at three times the volume is $196.20. A flat session costs $60 either way. On a router, the same agent pays whatever the chosen provider charges per token, plus the credit fee. Cheaper providers lower the per-token line, so put your own measured tokens and your router’s price for the model into the flat-rate vs per-token calculator before you decide.
When is OpenRouter the better fit?
- You switch models often. Trying five frontier and open models in a week is what a router is for.
- You want automatic fallback across providers when one is down or rate-limited. OpenRouter’s
allow_fallbackssetting is on by default. - You want to sort by price or throughput across providers serving the same model.
- Your usage is light or bursty. Per-token billing is cheapest when you don’t use much.
When is a flat-rate provider the better fit?
- An agent runs most of the day on one model. That is where the token meter costs most; see the break-even math.
- You need a fixed monthly number for a budget, with no surprise from a retry loop. That is what a flat-rate plan is for.
- You want one provider’s terms on every request. BroadNet’s privacy statement says its own and hosted open-weight models run on hardware in the United States, with partner capacity as backup, and that partner-served models, such as DeepSeek V4.1 Flash, run under zero-data-retention terms with US inference partners.
- You need Arabic speech to text and Arabic-tested chat models on one key. See choosing an LLM for Arabic.
How do the rate limits compare?
OpenRouter documents several kinds of limit: limits on free model variants (IDs ending in :free), 402 errors while a credit balance is negative, and 429s passed on from upstream providers “when the provider serving your request is rate limiting or at capacity”. BroadNet’s pay-as-you-go keys allow 30 requests a minute and 4 in flight. Its plans have no token meter and run one request at a time per session. Neither is “unlimited” in every sense; what to check in any “unlimited” plan.
Which one keeps less of your data?
OpenRouter routes to many providers with different policies. Its provider routing settings let you narrow that: zdr: true routes only to zero data retention endpoints, and data_collection: "deny" excludes providers that may store your data. The default for data_collection is “allow”, so set it if it matters to you. On BroadNet Inference, no prompt or output content is retained on any request, and there is no setting to get wrong. The ZDR checklist lists what to ask either way.
How do you switch an agent between them?
Both speak the OpenAI API format, so moving a workload is a configuration change:
from openai import OpenAI
import os
client = OpenAI(
base_url="https://inference.broadnet.ai/v1", # was the router's base URL
api_key=os.environ["BROADNET_API_KEY"],
)
reply = client.chat.completions.create(
model="broadnet/deepseek-v4.1-flash", # BroadNet model ID
messages=[{"role": "user", "content": "Summarise this diff."}],
max_tokens=2000, # leave room for reasoning
)
Model IDs differ between services, and reasoning models need room in max_tokens. The migration checklist lists everything to check. The Python SDK guide covers setup.
What is a practical setup that uses both?
- Keep the router for experiments and rare models.
- Find the agent that runs every day on one model. Log a week of its token usage.
- Put that week into the break-even math. If a plan wins, buy a $1 Meteor hour and run the same work on it.
- Move that agent to plan sessions, one per agent running in parallel. How sessions work.
- Keep the router configured as a fallback for anything the plan doesn’t cover.
Where BroadNet Inference fits
BroadNet Inference is not a router and doesn’t try to be: six models, one provider, no fallback across providers. It fits the agent that runs all day on DeepSeek V4.1 Flash and needs a flat bill, plus Arabic speech to text. For broad model choice, OpenRouter or a self-hosted gateway remains the better tool. New BroadNet accounts start with $5 of credit and no card. That starting credit covers BroadNet Gemma 4 VL 26B and the two BroadNet Transcribe models; DeepSeek V4.1 Flash, GLM 5.3 and Qwen3.8 27B need a top-up first.
