BroadRouter

How inference providers price open models: 6 compared

By the BroadNet Inference team · 7 min read · Published · Updated

Most providers price open models per million tokens, with cheaper cached input and often 50% off batch jobs. Together AI, Fireworks and DeepInfra also rent GPUs by the hour. OpenRouter passes provider prices through and charges a fee on credit purchases. Groq covers a short model list. BroadNet adds flat monthly, weekly and hourly plans for DeepSeek V4.1 Flash.

“Price per million tokens” looks comparable across providers. It isn’t quite. Providers differ on which models they host, what cached input costs, whether batch jobs are discounted, whether you can rent GPUs, and what fees sit on top. This post compares six providers of open-weight models on how they price, then puts their published prices for the same models side by side. All competitor figures come from each provider’s official pages, checked 2026-11-14; the links are in the sources. Prices change often, so recheck before you commit.

What are the five ways open models are priced?

  1. Per token, serverless. Separate prices per 1M input and output tokens. Everyone in this list offers it.
  2. Cached input. A lower rate for repeated prompt prefixes. Discounts range from half price to a few percent of the input rate. Cached tokens explained.
  3. Batch. Lower prices for jobs that can wait hours.
  4. Dedicated GPUs. Rent hardware by the hour and run the model yourself or through the provider’s stack. Self-hosting vs an API: the break-even, or run it with your own numbers in the self-hosting vs API calculator.
  5. Flat plans. A fixed price for time, with no token meter. See flat-rate LLM plan.

How does each provider price open models?

Provider Model Per-token Cached input Batch Dedicated GPUs Flat plan
OpenRouter Router across many providers Provider’s price, “no markup” Per provider Per provider No No
Together AI Own serverless + dedicated Yes Listed for some models 50% off on selected models (Whisper Large v3 only, when checked) H100 from USD 5.49 per GPU hour No
Fireworks AI Own serverless + on-demand Yes, from USD 0.05 per 1M input 50% off by default, more on some models 50% off H100 USD 8.00 per hour No
Groq Own hardware Yes, short model list 50% off, on GPT-OSS models 50% off Not listed No
DeepInfra Own serverless + dedicated Yes, some models billed by execution time Listed for many models Not listed on pricing page H100 USD 2.90 per GPU hour No
BroadNet Inference Short model list; partner-hosted models with US inference partners Yes Listed for three of four chat models Not offered No Yes, DeepSeek V4.1 Flash

Notes from the pages:

  • OpenRouter says “there is no markup on inference pricing”. It charges a fee when you buy credits: 5.5% on the Standard plan (minimum USD 0.80, per its FAQ) and 8% on Business, or 5% for crypto. Free model variants are limited to 20 requests a minute and 50 a day, or 1,000 a day after buying at least USD 10 of credits. It is a router, so one model can have many prices. OpenRouter vs a flat-rate provider.
  • Together AI lists dedicated H100 at USD 5.49 per GPU hour on demand. Its batch docs say selected serverless models run at 50% off, with a completion window fixed at 24 hours, but the discounted-models table listed only openai/whisper-large-v3 when we checked. Other models run at standard rates.
  • Fireworks has Standard, Priority, Fast, Batch (50% off) and Spot (30% off on select models) modes. Cached input is billed at half price or less, with no write fee.
  • Groq lists per-token prices for a small set of production models, such as openai/gpt-oss-120b at USD 0.15 input and USD 0.60 output per 1M, and Whisper speech to text by the hour. Cached input is half price on its GPT-OSS models, and batch jobs run at 50% off within 24 hours to 7 days.
  • DeepInfra shows many models with a discounted rate next to the list price, for example 51% off DeepSeek-V4.1-Flash-1017 and 38% off GLM-5.3. The page does not say when these discounts end. It also offers Standard, Priority (1.5×) and Flex (0.8×) tiers.

What does the same model cost at each provider?

Published per-1M prices for models that BroadNet also serves. ”—” means the model was not on that provider’s pricing page.

DeepSeek V4.1 Flash (input / cached / output):

Provider Input Cached Output
Together AI USD 0.30 USD 0.006 USD 1.20
Fireworks AI USD 0.30 USD 0.006 USD 1.20
DeepInfra, list USD 0.20 USD 0.006 USD 0.60
DeepInfra, discounted (30% off) USD 0.14 USD 0.0042 USD 0.42
DeepInfra, DeepSeek-V4.1-Flash-1017, discounted (51% off) USD 0.098 USD 0.0029 USD 0.294
Groq — — —
BroadNet Inference $0.175 $0.0055 $0.52

DeepInfra’s dated -1017 variant is the same release DeepSeek’s own API serves as deepseek-flash; check which one your code calls. On OpenRouter, the model had 37 endpoints from 28 providers when we checked, from about USD 0.05 (Baidu) to USD 3 (Venice’s fastest tier) per 1M input tokens. Our DeepSeek V4.1 Flash price table lists every provider that publishes its own price for the model, with the source page and check date in each row. How wide the same-model price spread runs across the whole market is measured in our LLM API price index. For how these open-model prices compare with GPT, Claude and Gemini on the same workload, see open-weight vs closed model API cost.

GLM 5.3 (every provider’s own price: GLM 5.3 price table):

Provider Input Cached Output
Together AI USD 1.40 USD 0.26 USD 4.40
Fireworks AI USD 1.40 USD 0.26 USD 4.40
DeepInfra, list USD 0.90 USD 0.20 USD 4.00
DeepInfra, discounted (38% off) USD 0.5625 USD 0.125 USD 2.50
BroadNet Inference $0.78 $0.173 $3.46

Qwen3.8 27B (every provider’s own price: Qwen3.8 27B price table):

Provider Input Cached Output
DeepInfra, list USD 0.20 USD 0.05 USD 2.50
DeepInfra, discounted (25% off) USD 0.15 USD 0.0375 USD 1.875
Groq (preview, evaluation only) USD 0.80 — USD 4.00
BroadNet Inference $0.175 $0.0435 $2.17

Gemma 4 26B A4B (BroadNet’s build is Gemma 4 VL 26B):

Provider Input Output
DeepInfra USD 0.07 USD 0.34
BroadNet Inference $0.10 $0.22

BroadNet is not the cheapest per token here. DeepInfra’s discounted rates are below BroadNet’s on every line for DeepSeek V4.1 Flash, GLM 5.3 and Qwen3.8 27B, and below it on Gemma input (BroadNet’s Gemma output is lower). DeepSeek’s own API halves all its prices off-peak, which puts its input below BroadNet’s then (details). BroadNet’s list prices are below Together’s and Fireworks’ on DeepSeek V4.1 Flash and GLM 5.3. Per-token price is one input; also check context length, data retention and rate limits, which these tables don’t show.

Across all 119 models that several hosts sell, not just these four, the dearest list price is a median 1.76 times the cheapest; see same model, different host. Host prices also move more often than makers’ prices: how often LLM API prices change.

What about speech to text?

Groq lists whisper-large-v3-turbo at USD 0.04 per hour of audio and whisper-large-v3 at USD 0.111 per hour. DeepInfra lists Voxtral Mini at USD 0.001 per minute. BroadNet’s multilingual model costs $0.002 a minute ($0.12 an hour), and its Arabic model $0.0042 a minute. For more providers, see speech-to-text API pricing compared. For Arabic, accuracy on your dialects matters more than the rate: how to choose an Arabic speech-to-text API.

How do you compare providers for your workload?

  1. Fix the model and version. “DeepSeek V4.1 Flash” and a dated variant can be priced differently.
  2. Measure your token mix: input, cached share and output, including reasoning. Estimate cost before launch.
  3. Price it with each provider’s three rates, not just the input price. For agents, the cached rate often decides. The LLM API cost calculator and prompt caching calculator do the arithmetic.
  4. Apply batch discounts only to work that can wait.
  5. Add fees: credit-purchase fees and tier multipliers.
  6. Check the limits: rate limits, concurrency and context length can rule a provider out at any price.
  7. Compare against a flat plan if one agent runs most of the day. The break-even math, or the flat-rate vs per-token calculator.

Where BroadNet Inference fits

BroadNet serves a short list of models, some partner-hosted with US inference partners, with per-token prices published in a price file. Its per-token rates are mid-table, not the lowest: if price per token is all that matters, DeepInfra’s discounted listings cost less. What the others here don’t sell is unlimited DeepSeek V4.1 Flash for $60 a month per session, plus $15 weekly and $1 hourly options. It is not a marketplace, has no batch discount and no dedicated GPUs. Its other edges are zero data retention on every request, Arabic speech to text and one OpenAI-compatible key for chat and audio. If you need dozens of model families or the lowest promotional rate for a bursty workload, a router or a larger host will serve you better.

Questions

Which inference provider is cheapest for open models?

No single provider, and not BroadNet per token. When we checked on 2026-11-14, DeepInfra's discounted rates were below BroadNet's on DeepSeek V4.1 Flash, GLM 5.3 and Qwen3.8 27B, and OpenRouter listed DeepSeek V4.1 Flash endpoints from about USD 0.05 to USD 3 per 1M input tokens. Compare the exact model, cached rate and batch discount.

Does OpenRouter mark up model prices?

OpenRouter says there is no markup on inference pricing. It charges a fee when you buy credits, 5.5% on its Standard plan (USD 0.80 minimum, per its FAQ) and 8% on Business, and 5% for crypto payments.

Is batch pricing worth it?

For work that can wait. Fireworks lists batch at 50% off, and Groq documents a 50% batch discount with a 24-hour to 7-day window. Together's batch docs list 50% off for selected models only (the table named just `openai/whisper-large-v3` when we checked), with a fixed 24-hour completion window.

When should I rent a dedicated GPU instead of paying per token?

When your traffic keeps the GPU busy most of the day. Hourly GPU prices are charged whether or not requests arrive, so low or spiky traffic is cheaper per token.

Does any provider offer a flat monthly price for an open model?

BroadNet Inference does, for DeepSeek V4.1 Flash. Full Moon costs $60 a month per session with no token meter, one request at a time.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity