BroadNet
The JournalProduct
Product

BroadRouter guide 2026: one OpenAI-compatible API for DeepSeek, GLM, Qwen, Gemma and Arabic speech

BroadRouter is BroadNet's new developer hub for BroadNet Inference, one OpenAI-compatible API for DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma and Arabic speech-to-text. This guide covers the models, when a flat-price session beats paying per token, what happens to your data and how to connect your code.

14 min read
Product · Issue 03
14 min read

BroadRouter is BroadNet's new developer hub for BroadNet Inference, an OpenAI-compatible LLM API that serves four chat models (DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B and BroadNet Gemma 4 VL 26B) and two speech-to-text models from one endpoint, https://inference.broadnet.ai/v1. You keep your OpenAI SDK, change the base URL, key and model ID, and pay either per token or a flat price per session for unlimited DeepSeek V4.1 Flash.

Facts checked at their sources on 16 November 2026: broadrouter.com and its llms.txt, BroadNet's published rate card (pricing.json, dated 14 November 2026), the BroadNet Inference page and its privacy statement. Prices change; the live rate card is the price of record.

This table is the short answer. The rest of the guide explains each row.

QuestionShort answer
What is BroadRouter?BroadNet's developer hub (broadrouter.com) for the BroadNet Inference API: models, prices, setup guides, free tools and open data
Is it a router between providers?No. You choose a model ID and call one provider, BroadNet Inference, directly
Which models?DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, BroadNet Gemma 4 VL 26B, BroadNet Transcribe Arabic and BroadNet Transcribe (14 languages)
Endpoints/v1/chat/completions and /v1/audio/transcriptions, OpenAI-compatible
Cheapest per tokenGemma: $0.10 per 1M input tokens and $0.22 per 1M output tokens
Flat priceUnlimited DeepSeek V4.1 Flash: $60 a month, $15 a week or $1 an hour per session
Free trial$5 starting credit, no card needed, for Gemma and both transcription models
DataPrompts, files, answers and transcripts are not retained or used for training

What is BroadRouter?

BroadRouter is the website where BroadNet explains its inference API to developers. It lists every model with its exact ID, context window and price, compares models side by side, and publishes setup guides for more than 30 tools, from the OpenAI SDKs to n8n and coding agents. It also runs free browser tools, such as an LLM cost calculator and token counters, and publishes open datasets on LLM prices and tokenizer efficiency under CC BY 4.0.

The API itself is BroadNet Inference. Accounts, keys and billing live at inference.broadnet.ai, and your code calls https://inference.broadnet.ai/v1 directly. Despite the name, BroadRouter does not route or forward your requests between providers and does not resell anyone else's API: you pick the model, and BroadNet serves it. You can read the overview on our BroadRouter page or visit broadrouter.com.

What does OpenAI-compatible mean, and why does it matter?

An OpenAI-compatible API accepts the same request and response format as the OpenAI chat completions and audio transcriptions endpoints. That format has become the common language of LLM tooling, so code and apps written for it can switch providers without a rewrite.

To move existing code to BroadNet Inference, you change three values:

  1. Base URL: https://inference.broadnet.ai/v1
  2. API key: a BroadNet inference key, sent only in the Authorization: Bearer header
  3. Model ID: for example broadnet/deepseek-v4.1-flash or broadnet/gemma-4-26b-a4b

Two limits are worth knowing before you migrate. BroadNet Inference serves chat completions and audio transcriptions, not the newer Responses API, so frameworks that default to it need that setting turned off. And the three reasoning models (DeepSeek V4.1 Flash, GLM 5.3 and Qwen3.8 27B) count their reasoning toward max_tokens, so set max_tokens to 1,000 or more or the answer can come back empty.

Which models does BroadRouter offer?

Six models run on the one endpoint. Prices below are in USD per 1 million tokens (input / output), or per minute of audio, from the rate card dated 14 November 2026.

ModelModel IDInputContext (tokens)Price (USD)Best for
DeepSeek V4.1 Flashbroadnet/deepseek-v4.1-flashText and images1,048,576$0.175 / $0.52 (cached input $0.0055)Coding agents, long documents, flat-price unlimited use
GLM 5.3broadnet/glm-5.3Text1,048,576$0.78 / $3.46 (cached input $0.173)Tool calling and structured output over very long context
Qwen3.8 27Bbroadnet/qwen3.8-27bText and images262,144$0.175 / $2.17 (cached input $0.0435)Arabic text and images, including Gulf and Egyptian dialects
BroadNet Gemma 4 VL 26Bbroadnet/gemma-4-26b-a4bText and images32,768$0.10 / $0.22Lowest price per token, the free starting credit, hourly keys
BroadNet Transcribe Arabicbroadnet/cohere-transcribe-arabicArabic audion/a$0.0042 per minuteArabic voice notes, calls and dialects
BroadNet Transcribebroadnet/cohere-transcribeAudio in 14 languagesn/a$0.002 per minuteTranscription in 14 languages at the lowest rate

Vision here means the model reads images, such as a photo of an invoice; none of these models generates images. The starting credit covers Gemma and the two transcription models, while DeepSeek, GLM and Qwen need a top-up or a session plan first. For agents and tools, read the current list of IDs from BroadNet's machine-readable rate card instead of hard-coding them.

How much does BroadRouter cost?

There are three ways to pay, and you can mix them on one account.

  • Pay as you go: every model, billed per token or per audio minute from a prepaid balance. Cached input is billed at the lower cached rate and images count as input tokens. Each key allows 30 requests a minute and 4 requests in flight, an account can hold up to 10 live keys, and a key lasts one year.
  • Gemma hourly key: $0.50 for one hour of BroadNet Gemma 4 VL 26B with one stream, paid from your balance.
  • Unlimited DeepSeek V4.1 Flash sessions: a flat price per session with no token meter, sold by the month, week or hour.
PlanPrice per sessionLastsPaid by
Full Moon$60A month, renewsCard
Orbit$15A week; buy the next week before it ends to keep the same keyAccount balance
Meteor$1An hour from purchaseAccount balance

A founders offer, published in BroadNet's rate card, prices Full Moon at $45 a month per session for the first 3 months with a founders code at checkout, then $60. Plan keys are separate from pay-as-you-go keys and never draw on your balance.

Pay per token or unlimited sessions: which is cheaper?

It depends on how busy your agent is. A session runs one request at a time and has no token meter, so it pays off when a model works for most of the day. A tool that calls the model now and then is usually cheaper per token.

Two worked examples, using the published prices of 14 November 2026. Both are illustrations calculated by us, not measurements:

Workload in 30 daysPay-as-you-go cost on DeepSeek V4.1 FlashFull Moon sessionCheaper option
An agent generating around the clock at the typical 100 tokens a second: about 259 million output tokensAbout $135 for output tokens alone, before any input$60Session
A coding assistant used a few hours a day: 40 million input tokens (80% cached) and 4 million output tokensAbout $3.66 ($1.40 uncached input + $0.18 cached input + $2.08 output)$60Pay as you go

To decide for your own workload:

  1. Run your real workload on pay as you go for a few days and read the token counts from the usage field of each response.
  2. Count how many requests your agents send at the same moment, including sub-agents. Each one needs its own session.
  3. Multiply your monthly tokens by the per-token prices and compare with $60 per session.
  4. Before you commit to a month, buy a $1 Meteor hour and run your agent loop on it.

Sessions cover DeepSeek V4.1 Flash only. For GLM, Qwen, Gemma or transcription, use a pay-as-you-go key. Sharing or reselling a plan key is not allowed, and access follows BroadNet's acceptable use policy.

How do you send your first request?

  1. Sign in at inference.broadnet.ai with Google or an email link. New accounts start with $5 of credit, and no card is needed.
  2. Create an inference key. It is shown once, so store it in an environment variable such as BROADNET_API_KEY.
  3. In your OpenAI SDK, set the base URL to https://inference.broadnet.ai/v1 and pass your key.
  4. Send a chat completion with model broadnet/gemma-4-26b-a4b, which runs on the starting credit.
  5. To transcribe audio, post a WAV or MP3 file of up to 26,214,400 bytes to /v1/audio/transcriptions with the model ID and a language field, for example language=ar.

If you want to test before creating an account, BroadNet's browser playground runs Arabic speech-to-text, document reading and Gemma chat with no sign-up. When a request fails, check the HTTP status first: fix 400, 403, 413 and 422 errors before retrying, treat 401 as a key or balance problem and 402 as an empty balance, and retry 429 and 5xx errors with bounded exponential backoff.

Which tools work with BroadRouter?

Any tool that accepts a custom OpenAI-compatible base URL works. BroadRouter publishes step-by-step guides, with the exact fields to fill in, for these:

  • SDKs and frameworks: the OpenAI Python and Node.js SDKs, the Vercel AI SDK, LangChain, LlamaIndex, LiteLLM, Haystack, Semantic Kernel, Spring AI, openai-go, Instructor and DSPy.
  • Agent frameworks: the OpenAI Agents SDK, LangGraph, CrewAI, Pydantic AI, Mastra and smolagents, plus promptfoo for evaluations.
  • Automation and chat apps: n8n, Dify, Flowise, Langflow, Open WebUI, LibreChat and Jan.
  • Coding agents and editors: Cline, Continue, Kilo Code, Zed, Aider, opencode and goose.

The full list, with a guide for each, is on BroadRouter's integrations page. Coding agents are the natural fit for the unlimited DeepSeek V4.1 Flash sessions, because an agent resends its context on every step and a flat price removes the token meter.

Does BroadNet keep your prompts and files?

No. BroadNet Inference processes prompts, images, audio, answers and transcripts to serve the request and does not retain them, write them to logs or use them to train, fine-tune or evaluate models. BroadNet says it checks its gateway configuration every hour to confirm that requests are served without storing content.

  • Never retained: prompts, images and audio you send; answers and transcripts; request and response bodies; raw URLs and query strings; raw authentication values.
  • Kept: content-free records needed to run the service, such as request IDs, timings, HTTP status, token counts and cost, plus account, billing and security data.
  • Where it runs: BroadNet's own and hosted open-weight models run on BroadNet-owned or leased hardware in the United States. Partner-served models, such as DeepSeek V4.1 Flash, run with US inference partners under zero-data-retention terms.
  • Data residency: regional processing in MENA or Europe, and private endpoints placed in the region you need, are available on request.

Keys add a second layer of control. Inference keys can call model routes only, never account settings, and you can lock them to your own servers with IP allow-lists and revoke a leaked key at once. For the full field list, read BroadNet's inference privacy statement.

Why is BroadRouter a good fit for Arabic apps?

Arabic is where BroadNet Inference is most specialised. BroadNet Transcribe Arabic handles Gulf, Egyptian, Levantine, Maghrebi and Modern Standard Arabic and returns Arabic script, at $0.0042 per audio minute, or about $0.25 an hour. On the Open Universal Arabic ASR Leaderboard it was listed #4 of 39 systems on 15 November 2026; BroadNet reported it #2 of 37 in September 2026.

Model choice also changes what Arabic costs. Tokenizers split Arabic into more pieces than English, so the same sentence uses more tokens. BroadRouter's open tokenizer efficiency study, measured on the same 2,009 FLORES-200 sentences, found these ratios for Modern Standard Arabic:

ModelArabic tokens for every English token (same sentences)
Qwen3.8 27B1.31×
BroadNet Gemma 4 VL 26B1.47×
DeepSeek V4.1 Flash1.62×
GLM 5.31.77×

In practice, a customer-service app that reads Arabic WhatsApp voice notes can convert each .ogg file to WAV with ffmpeg, transcribe it with language=ar, then pass the text to Qwen3.8 27B for an answer. For the transcription side in depth, read our Arabic speech-to-text API guide; to receive those voice notes in the first place, see the WhatsApp Business API.

What are the limits?

  • No uptime commitment on self-serve. BroadNet's terms say self-serve has no availability commitment and unlimited plans are delivered best effort under fair use. Service levels come with a written agreement for private endpoints or reserved capacity.
  • Rate limits per key. Pay as you go allows 30 requests a minute and 4 requests in flight per key; a session runs one request at a time.
  • No Responses API. Use chat completions.
  • No image generation. The vision models read images; they do not create them.
  • Audio file size. Up to 26,214,400 bytes per file, so split long recordings first.
  • Six models, not hundreds. The catalogue is deliberately small. If you need a model that is not listed, ask BroadNet about a private endpoint.

How does BroadRouter fit with the rest of BroadNet?

BroadRouter is the self-serve, developer-first door to BroadNet's AI. Teams that want ready-made agents rather than raw models can use the BroadNet AI enterprise platform, and teams that need the agents to reach customers can connect them to WhatsApp, SMS and the other channels on BroadNet's messaging network. For a broader view of agent platforms, see our enterprise agentic AI platform guide.

Frequently asked questions

Is BroadRouter free to use?

The BroadRouter website, its guides and its browser tools are free. The API is paid, but new accounts start with $5 of credit, with no card needed, for BroadNet Gemma 4 VL 26B, both transcription models and Gemma hourly keys.

Is BroadRouter the same as BroadNet Inference?

They work together. BroadRouter (broadrouter.com) is the developer hub with models, prices and guides; BroadNet Inference (inference.broadnet.ai) is the API, where you create your account and keys and send requests.

Can I use BroadRouter with the OpenAI Python SDK?

Yes. Install the official openai package, create the client with base_url set to https://inference.broadnet.ai/v1 and your BroadNet key, and pass a BroadNet model ID such as broadnet/deepseek-v4.1-flash.

What is the cheapest model on BroadRouter?

BroadNet Gemma 4 VL 26B, at $0.10 per million input tokens and $0.22 per million output tokens on the 14 November 2026 rate card. For transcription, BroadNet Transcribe costs $0.002 per audio minute.

Is the DeepSeek plan really unlimited?

Yes, in tokens: there is no token meter, no daily cap and no weekly limit. What a session limits is concurrency, one request at a time, and use stays subject to context limits and the acceptable use policy.

How many sessions do I need for several agents?

One per request that runs at the same moment. A single coding agent needs one session; an orchestrator with three sub-agents working at once needs four; ten parallel batch workers need ten.

Is my data used to train models?

No. BroadNet states that prompts, images, audio, answers and transcripts are not retained and are never used to train, fine-tune or evaluate models, including on partner-served models, which run under zero-data-retention terms.

Can BroadNet run a model in my region or on a private endpoint?

Yes. Processing is in the United States by default, and regional processing in MENA or Europe, private endpoints, reserved capacity and Arabic fine-tunes are available on request under a written agreement.

Ready to try it? Start from our BroadRouter overview, compare models on broadrouter.com, or talk to BroadNet about private endpoints and reserved capacity.

Was this useful?

How we write and review content

Written by

BroadNet Editorial Team

The Editorial Team publishes the BroadNet blog. Articles are written with BroadNet's regulatory affairs, product, engineering, industry research and operator solutions teams, and reviewed under our editorial standards.

The Journal, by email

One essay a month. Nothing else.

Subscribe and we'll send the next piece — engineering, regulation, or industry — the day it ships. No drip campaigns, no product pitches, no third-party sharing. Unsubscribe in one click.

We send one email a month. Unsubscribe anytime. See our privacy policy.

What you'll get
  • Long-form, not link bait
    Six to eleven minutes of focused writing on a single topic.
  • Written by operators
    Engineers, regulatory leads, and product writing in-house.
  • Region-aware
    MENA-grounded view on regulators, operators, and channel mix.
  • Once a month
    No drip campaign. One essay when there is one to share.
Stop reading. Start shipping.

Most of what we write we learned in production.

If this piece made you nod, we should probably talk. Bring the use case, the volumes, the markets — we'll come back with routing, pricing, and a reference customer you can call.

$ curl -X POST https://api.broadnet.me/sms/sendSms