BroadRouter
Get started

Same model, different host: how much LLM API prices vary

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

A lot. Across 119 models that two or more independent hosts sell, the dearest list price is a median 1.76 times the cheapest (blended, 3:1), from 41 providers' pages checked 14 and 15 November 2026. 50 models have a gap of 2 times or more. Yet most hosts copy the maker: 57% of host prices equal the maker's own price exactly.

If you use an open-weight model through an API, the host you pick can change the bill more than the model does. We measured how much. BroadRouter reads every price from each provider’s own pricing page, with a source URL, a check date and a verbatim quote. For this post we took every model that at least two independent hosts sell and compared their standard list prices. BroadRouter is BroadNet’s website, and BroadNet sells four of these models; its prices are shown separately and not ranked.

The data is our LLM API provider prices dataset, which also feeds the LLM API price index and every per-model price page.

How much does the same model’s price vary between hosts?

Usually by well over half, and often by double. We compared 119 models, each with list prices from at least two providers. For each model we took the dearest standard price and divided it by the cheapest, on a blended price of 3 input tokens per output token:

Dearest ÷ cheapest Models
Exactly the same price 13
Under 1.25× 34 (including the 13)
2× or more 50
3× or more 28
5× or more 13
Median 1.76×

The spread grows with the number of sellers. For the 19 models with 10 or more providers the median gap is 2.02 times; for the 51 models with only 2 or 3 providers it is 1.28 times. A popular model has more hosts, and more of them try a different price.

The widest gap is Llama 3.1 8B Instruct: 16.7 times between DeepInfra and Cloudflare Workers AI. Among models with many hosts, Llama 3.3 70B Instruct (7.9 times, 14 providers), DeepSeek V4 Flash 0731 (7.3 times, 14 providers) and gpt-oss-120b (6.9 times, 20 providers) have the widest gaps.

Seven models with many hosts. List prices in USD per 1M tokens, input / output, checked 14 and 15 November 2026. Each model links to its full price page with every provider and source.

Model Providers with a list price Cheapest, in / out Most common price, in / out Dearest, in / out Dearest ÷ cheapest (blended)
GLM 5.3 20 Chutes: 1.225 / 3.85 1.40 / 4.40 (16 providers) SambaNova: 2.80 / 8.80 2.29×
gpt-oss-120b 20 CoreWeave: 0.03 / 0.17 0.15 / 0.60 (6 providers) Cerebras, Cloudflare: 0.35 / 0.75 6.92×
DeepSeek V4.1 Flash 14 CoreWeave: 0.13 / 0.65 0.30 / 1.20 (12 providers) 12 providers tied: 0.30 / 1.20 2.02×
Llama 3.3 70B Instruct 14 DeepInfra: 0.10 / 0.32 0.71 / 0.71 (2 providers) Venice: 0.70 / 2.80 7.90×
Qwen3.8 27B 13 Chutes: 0.039 / 1.25 0.45 / 3.20 (3 providers) Groq: 0.80 / 4.00 4.68×
Gemma 4 26B A4B 11 DeepInfra, Featherless: 0.07 / 0.34 0.13 / 0.40 (5 providers) Vertex AI: 0.15 / 0.60 1.91×
Kimi K3 18 Chutes: 1.85 / 9.50 3.00 / 15.00 (12 providers) Venice: 3.75 / 18.75 1.99×

What the table shows:

  • GLM 5.3 is a price-matching market. 16 of 20 providers, Z.ai included, list exactly USD 1.40 / USD 4.40. Chutes is below it; Venice, Amazon Bedrock (US cross-region) and SambaNova are above. Four further hosts list promotional prices below the list (DeepInfra, GMI Cloud, FriendliAI and an AkashML “Limited Trial”); we leave those out. DeepInfra’s own undiscounted list price (USD 0.90 / USD 4.00) is shown only next to its promotion, not charged, so it is not counted either.
  • gpt-oss-120b is the opposite. 20 providers and 6.9 times between the cheapest and dearest. Groq, Together AI, Fireworks AI and three others charge USD 0.15 / USD 0.60; CoreWeave (USD 0.03 / USD 0.17) and DeepInfra (USD 0.037 / USD 0.17) charge a fifth to a quarter of that on input and under a third on output.
  • DeepSeek V4.1 Flash has 12 providers at USD 0.30 / USD 1.20. CoreWeave and SiliconFlow list about half that on input.
  • Llama 3.3 70B shows how variants blur the comparison. DeepInfra’s cheapest row is its quantised Turbo variant, while Venice charges USD 0.70 / USD 2.80.

Do hosts undercut the model maker?

Mostly they match it. For the 29 models whose maker also sells them on its own API, we compared each host price with the maker’s:

Host price vs the maker’s own price Host prices
Exactly the same input and output price 143 (57%)
Lower (blended) 83
Higher (blended) 23

The maker was the cheapest, or tied for cheapest, on only 3 of the 29 models. So the maker’s list price works as a ceiling that most hosts copy, and some undercut.

Which hosts are cheapest and dearest most often?

Host Cheapest for (models) Dearest for (models)
DeepInfra 33 7
Featherless 22 26
GMI Cloud 22 14
SiliconFlow 19 3
Novita AI 12 13
CoreWeave 11 1
Chutes 9 5
Cloudflare Workers AI 5 11
Venice 4 17
Amazon Bedrock 4 13
Alibaba Cloud Model Studio 2 13
Microsoft Azure 0 14

Ties count for every tied host, and models where every provider charges the same are not counted as “dearest”. Featherless and GMI Cloud appear high in both columns: they host many models and price each one separately. Featherless’s per-token rates also need a Developer plan with credits from 50 USD a month. No host is cheapest for everything, so compare on the exact model you will use.

What else differs besides price?

The same weights do not guarantee the same service:

  • Precision. 237 of the 681 rows in this comparison state a precision; FP8 is the most common (177 rows); 23 state FP4 and 4 state INT4. Lower precision can change output quality. See quantization.
  • Context window, rate limits and throughput differ by host and plan.
  • Data terms. Retention and training use differ; see LLM API data retention compared.
  • Price stability. Some hosts reprice often. In our archive, DeepInfra repriced half of its long-listed models at least once, in both directions. See how often LLM API prices change and LLM API price increases.

A practical approach: shortlist two or three hosts at or below the common price, run the same 50 to 100 real prompts through each, and keep the runner-up as a tested fallback. How to choose an LLM API provider has a checklist.

Methodology

  • Prices. 41 third-party providers’ own pricing pages, each price with source URL, check date (14 or 15 November 2026) and a verbatim quote: the LLM API provider prices dataset and its CSV. Standard tier only; batch, flex, priority and free rows excluded. BroadNet’s own file is excluded.
  • Promotions excluded. 49 standard prices labelled promotional, “limited trial” or ”% off list” are left out, as on the dataset pages’ rankings.
  • Which models. A model is included when at least two independent hosts list it (not counting the Amazon Bedrock, Azure and Vertex AI clouds, which also resell closed models) and at least two providers in total have a list price. Closed models from OpenAI (except gpt-oss), Anthropic, Google Gemini, xAI and Amazon are excluded. 119 models qualify.
  • Spread. One price per provider (its cheapest list row for the model), blended at 3 input tokens per output token; dearest ÷ cheapest. Our script recomputes every figure from the JSON files; the price index’s published spread uses all models and keeps promotional rows, so its median differs.
  • History. Price movement over time comes from the separate LLM API price history dataset (CSV), read from Internet Archive captures. BroadNet is not in it, because the archive has no captures of its pricing page.

Where BroadNet Inference fits

BroadRouter is BroadNet’s own site, and BroadNet Inference sells four of the models in this comparison. Its prices come from its pricing.json and are not ranked above:

Model BroadNet input BroadNet output Context in the comparison above
GLM 5.3 $0.78 $3.46 Below the list prices above; promotional prices at DeepInfra, GMI Cloud and AkashML are lower on at least one side
DeepSeek V4.1 Flash $0.175 $0.52 CoreWeave and SiliconFlow list lower input, BroadNet’s output is lower than theirs; DeepInfra’s promotional price is lower on both
Qwen3.8 27B $0.175 $2.17 Chutes and Hyperbolic list lower blended prices, as do promotional prices at DeepInfra and AkashML
Gemma 4 VL 26B (BroadNet’s build of Gemma 4 26B A4B) $0.10 $0.22 DeepInfra and Featherless list lower input; BroadNet’s output is lower

BroadNet is not the cheapest host for every model. DeepSeek V4.1 Flash on BroadNet is partner-hosted, and self-serve accounts have no SLA. BroadNet also sells a flat plan for DeepSeek V4.1 Flash at $60 a month per session, one request at a time. Of the chat models, only Gemma runs on the $5 starting credit; the other three need a top-up.

Questions

Why does the same open model cost different amounts at different providers?

Hosts run the same weights on different hardware, at different precision, utilisation and margin, and some add regional or private-deployment premiums. In our data the dearest list price for a model is a median 1.76 times the cheapest, across 119 models.

Do hosts usually charge the model maker's price?

Often, yes. Where the maker sells the model on its own API, 143 of 249 host prices (57%) matched the maker's input and output price exactly; 83 were lower and 23 higher. For GLM 5.3, 15 hosts list exactly Z.ai's USD 1.40 input and USD 4.40 output.

Which provider is cheapest for open models?

No single one. DeepInfra had the lowest list price, alone or tied, for 33 of 119 models, Featherless and GMI Cloud for 22 each and SiliconFlow for 19. Featherless was also the dearest for 26 models, so it depends on the model.

Are promotional prices included?

No. 49 standard prices that the provider labels as promotional or a limited trial are left out of every figure here, because they can end without notice. Several are lower than every list price, for example DeepInfra's GLM 5.3 at 38% off.

Is a cheaper host the same model?

Same weights, not always the same service. Hosts can serve lower-precision weights (where stated, FP8 was most common), shorter context or different rate limits. DeepInfra's cheapest Llama 3.3 70B row is its quantised Turbo variant. Test output quality and limits on your own prompts.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity