Skip to main content
Watch: Stop Building Apps. Build Tools for AI Agents

Open-weight API pricing - last verified

Where to run open-weight models, and what it costs

We read the pricing and model pages of 18 inference providers on October 18, 2026 and recorded every per-token price they list for 12 popular open-weight models: 80 offers, each value quoted from the provider's own page. Prices stay as listed, in USD per 1M tokens, and are only compared inside one unit. 4 prices are peak-hour or busy-hour, 3 carry a discount label, and 6 offers serve a shorter context than the model's native window.

Cheapest per modelWho serves whatLabels and readingsShorter contextNot covered

Models
12
Providers
18
Per-token offers
80
Quotes verified
561

Answer first

The cheapest listed input price for each model

The lowest input price any provider lists for each model, with every tie shown. It is the listed price only: DeepSeek's own rows are peak-hour prices, discount tags are flagged, and output, cached and batch prices are on each model page.
Cheapest listed input price per model: listed price, peak-hour where noted, discounts flagged. Computed only within USD per 1M input token rows.
ModelLowest listed inputUSD per 1M input tokensWhereFirst-party inputUSD per 1M input tokensOffers
DeepSeek V4 Pro$1.30$1.32DeepSeek, peak-hour price128 providers
DeepSeek V4.1 Flash$0.14
  • DeepInfraDeepSeek-V4.1-FlashCard shows 30% off - may be a temporary discount
$0.30DeepSeek, peak-hour price109 providers
Qwen3.8-27B$0.42$0.50Alibaba Cloud Model Studio55 providers
Kimi K3$2.85$3.00Moonshot AI99 providers
GLM-5.3$0.563
  • DeepInfraGLM-5.3Card shows 38% off - may be a temporary discount
$1.40Z.ai109 providers
MiniMax-M3$0.30$0.30MiniMax66 providers
gpt-oss-120b$0.05No first-party row1010 providers
gpt-oss-20b$0.07No first-party row44 providers
Llama 4 Maverick$0.25No first-party row33 providers
Gemma 4 31B$0.09No first-party row65 providers
Mistral Large 3$0.50$0.50Mistral AI22 providers
NVIDIA Nemotron 3.5 Super VL 120B-A12B$0.10No first-party row33 providers

Matrix

Which providers serve each model

Each cell is a listed input price in USD per 1M input tokens and links to that row on the model page. A cell with two prices means the provider lists two builds. A dash means not listed on the pages read.
Which providers serve each open-weight model, with each listed input price in USD per 1M input tokens linking to its row
ModelDeepSeekTogetherFireworksGroqCerebrasDeepInfraBasetenNovitaSambaNovaCloudflareBedrockVertex AIAzureAlibabaMoonshotZ.aiMiniMaxMistral
DeepSeek V4 Pro
  • $1.32first-partypeak-hour
----------
DeepSeek V4.1 Flash
  • $0.30first-partypeak-hour
---------
Qwen3.8-27B-------------
Kimi K3---------
GLM-5.3---------
MiniMax-M3-----------
  • $0.30first-party50% off, permanent
-
gpt-oss-120b--------
gpt-oss-20b--------------
Llama 4 Maverick---------------
Gemma 4 31B-------------
Mistral Large 3----------------
NVIDIA Nemotron 3.5 Super VL 120B-A12B---------------

Columns: DeepSeek = DeepSeek API; Together = Together AI; Fireworks = Fireworks AI; Groq = Groq; Cerebras = Cerebras; DeepInfra = DeepInfra; Baseten = Baseten; Novita = Novita AI; SambaNova = SambaNova Cloud; Cloudflare = Cloudflare Workers AI; Bedrock = Amazon Bedrock; Vertex AI = Google Vertex AI (Agent Platform partner models); Azure = Azure AI Foundry (Microsoft Foundry Models); Alibaba = Alibaba Cloud Model Studio; Moonshot = Moonshot AI (Kimi API); Z.ai = Z.ai; MiniMax = MiniMax; Mistral = Mistral AI.

Read carefully

Labels, discounts and inferred readings

Every label is visible text on the row it applies to. Inferred means our reading, not the page's words.
First-party
The model's own developer selling its API.
Peak-hour price
DeepSeek's listed price during its peak hours; off-peak is half, per its page. Alibaba's busy-hour rows work the same way.
Card shows N% off
A discount tag sits beside the price on the provider's catalog card. The number is the discounted one and may be temporary.
Inferred
Our reading, not the page's words: a name-only model match, a region taken from table order, or a context figure from an unlabelled tag.
Below native
The provider serves a shorter context than the model's native window in our local-models data.
Preview
The provider lists the model as a preview, for evaluation rather than production.

DeepSeek V4.1 Flash at DeepInfra: Card shows 30% off - may be a temporary discount

“30% off $0.004 cached, $0.14 in, $0.42 out / 1M”
Catalog card - deepinfra.com
“promotional pricing, 30% off already applied”
Model page - deepinfra.com

GLM-5.3 at DeepInfra: Card shows 38% off - may be a temporary discount

“38% off $0.125 cached, $0.563 in, $2.50 out / 1M”
Catalog card - deepinfra.com
“promotional pricing, 38% off already applied”
Model page - deepinfra.com

MiniMax-M3 at MiniMax: Page shows "Permanent 50% off" a struck-through list price

“MiniMax-M3 ≤ 512k input tokens Permanent 50% off $0.60 $0.30 / M tokens”
Pricing page - platform.minimax.io
  • Inferred: matched by name

    Kimi K3

    Amazon Bedrock: Kimi 3

    Amazon Bedrock lists this model as "Kimi 3" with no API model ID on the pricing page; it is matched to Kimi K3 by name only.

  • Inferred: matched by name

    MiniMax-M3

    Azure AI Foundry: FW MiniMax 3

    Azure lists this row as "FW MiniMax 3" under its Fireworks tab; it is matched to MiniMax-M3 by name only.

  • Inferred: region tab

    Qwen3.8-27B

    Alibaba Cloud Model Studio: qwen3.8-27b

    The International row is attributed to the page's first region tab by table order; the quoted row does not name the region tab.

  • Inferred from an unlabelled card tag

    DeepSeek V4.1 Flash

    DeepInfra: DeepSeek-V4.1-Flash

    Context inferred from an unlabelled card tag: the 1,048,576 figure is a header chip on the model page with no Context label.

  • Inferred from an unlabelled card tag

    GLM-5.3

    DeepInfra: GLM-5.3

    Context inferred from an unlabelled card tag: the 1,048,576 figure is a header chip on the model page with no Context label.

  • Inferred from an unlabelled tag

    DeepSeek V4 Pro

    DeepInfra: DeepSeek-V4-Pro

    Context inferred from an unlabelled tag: the quoted pricing row carries "1024k" with no Context label.

  • Inferred from an unlabelled tag

    Kimi K3

    DeepInfra: Kimi-K3

    Context inferred from an unlabelled tag: the quoted pricing row carries "1024k" with no Context label.

  • Inferred from an unlabelled tag

    Gemma 4 31B

    DeepInfra: gemma-4-31B-it

    Context inferred from an unlabelled tag: the quoted pricing row carries "256k" with no Context label.

  • Inferred from an unlabelled tag

    Gemma 4 31B

    DeepInfra: gemma-4-31B-it-turbo

    Context inferred from an unlabelled tag: the quoted pricing row carries "256k" with no Context label.

  • Inferred from an unlabelled tag

    NVIDIA Nemotron 3.5 Super VL 120B-A12B

    DeepInfra: NVIDIA-Nemotron-3.5-Super-VL

    Context inferred from an unlabelled tag: the quoted pricing row carries "256k" with no Context label.

Context

Offers that serve a shorter context than the model's native window

Native context comes from our local-models records (model cards and configs). A served figure clearly below it is flagged; written forms such as 1M against 1,048,576 are not.
Offers serving less than the model's native context
ModelProviderServed, as listedNative
Qwen3.8-27BGroq131,072262,144HF API (params, license, createdAt)
Qwen3.8-27BCerebras64k / 128k262,144HF API (params, license, createdAt)
MiniMax-M3Together AI524,2881,048,576Hugging Face model card
Gemma 4 31BSambaNova128k tokens262,144HF API
NVIDIA Nemotron 3.5 Super VL 120B-A12BDeepInfra256kinferred from an unlabelled tag1,048,576HF API (params, license, createdAt)
NVIDIA Nemotron 3.5 Super VL 120B-A12BTogether AI262,1441,048,576HF API (params, license, createdAt)

Vendor-stated

Speed claims, as the providers state them

Not measured by us and not comparable across providers. Shown because the pages state them beside the price.

Qwen3.8-27B: Groq states 450 tokens per second for qwen/qwen3.8-27b.

“Qwen/Qwen3.8-27B qwen/qwen3.8-27b 450 $0.80input$4.00output”
Vendor-stated - console.groq.com

Qwen3.8-27B: Cerebras states 1,850 tokens per second for Qwen 3.8 27B.

“Qwen 3.8 27B ~1,850 tokens/s $0.99/M tokens $1.49/M tokens”
Vendor-stated - cerebras.ai

gpt-oss-120b: Groq states 500 tokens per second for openai/gpt-oss-120b.

“GPT OSS 120B openai/gpt-oss-120b 500 $0.15input$0.60output”
Vendor-stated - console.groq.com

gpt-oss-120b: Cerebras states 3000 tokens per second for GPT OSS 120B.

“GPT OSS 120B ~3000 tokens/s $0.35/M tokens $0.75/M tokens”
Vendor-stated - cerebras.ai

gpt-oss-20b: Groq states 1000 tokens per second for openai/gpt-oss-20b.

“GPT OSS 20B openai/gpt-oss-20b 1000 $0.075input$0.30output”
Vendor-stated - console.groq.com

Coverage

What is not covered

Where a page did not list a model or a price, it says so here as not listed on the pages read, which is not evidence that none exists.

Not listed on the pages read

  • Amazon Bedrock: no Llama 4 Maverick price is listed on the pages read (no Llama 4 table in the pricing page's default view).
  • Amazon Bedrock: gpt-oss prices for US regions are not listed on the pages read; only the Asia Pacific (Sydney) table is in the default view.
  • DeepInfra: gpt-oss-120b, gpt-oss-20b, MiniMax-M3 and Llama 4 Maverick are not listed on the pages read.
  • Fireworks AI: only the headline-model price table is covered; models outside it fall under size-based pricing and are not listed on the pages read.
  • Together AI: gpt-oss-20b, Llama 4 Maverick and Qwen3.8-27B appear only in its fine-tuning table, not as serverless inference.
  • Cerebras: the pay-as-you-go Developer tier lists only GPT OSS 120B and Qwen 3.8 27B; other models are Dedicated Inference by quote.
  • Cloudflare Workers AI: no Kimi K3, DeepSeek V4.1 Flash, Gemma 4 31B, MiniMax-M3 or Nemotron 3.5 rows on the pages read.
  • Azure AI Foundry: no gpt-oss, Azure-direct Kimi K3, GLM-5.3 or Mistral Large 3 rows on the pages read.
  • Google Vertex AI: Gemma 4 is listed only as Gemma 4 26B, so no Gemma 4 31B row.

Batch discounts without per-model prices

  • Fireworks AI says batch inference is billed at 50% of serverless pricing, with no per-model batch figure.
  • Novita says batch inference is at an introductory 50% discount for supported models, with no per-model batch prices.
  • Together AI shows a Batch API price toggle but lists no per-model batch figures.
  • Amazon Bedrock says Flex tier and Batch pricing is at 50% discount to Standard tier pricing; per-model batch prices are listed only for gpt-oss, in the Asia Pacific (Sydney) table.
  • Mistral AI keeps Batch and Priority prices on separate tabs; the default Standard tab lists no batch price.

Providers checked and skipped

  • Hyperbolic

    hyperbolic.ai/pricing is a not-found page, and the Hyperbolic docs describe GPU rental products (On-Demand, Reserved, Private Cloud) with no per-token serverless model price list, so there is no per-token price to list.

  • Nebius AI Studio / Token Factory

    The Token Factory prices page (nebius.com/token-factory/prices) shows a "Log in to continue" wall, and the public Token Factory pages read had no per-model price table. No price is listed here rather than relying on a non-pricing page.

  • Hugging Face Inference Providers

    Hugging Face publishes no per-model rates of its own: its pricing doc says "Hugging Face charges you the same rates as the provider, with no additional fees." Prices are the underlying providers', which are covered directly.

Routers and marketplaces are out of scope; the providers they route to are covered directly.

FAQ

Where is each open-weight model cheapest to run through an API?

Lowest listed input price per model on October 18, 2026, in USD per 1M input tokens (listed price, peak-hour where noted, discounts flagged): DeepSeek V4 Pro $1.30 at DeepInfra (DeepSeek-V4-Pro) and DeepInfra (DeepSeek-V4-Pro-0813); DeepSeek V4.1 Flash $0.14 at DeepInfra (DeepSeek-V4.1-Flash), card shows 30% off; Qwen3.8-27B $0.42 at Novita AI (Qwen3.8 27B); Kimi K3 $2.85 at DeepInfra (Kimi-K3); GLM-5.3 $0.563 at DeepInfra (GLM-5.3), card shows 38% off; MiniMax-M3 $0.30 at MiniMax (MiniMax-M3 ≤ 512k input tokens), permanent 50% off, Together AI (MiniMax M3), Fireworks AI (MiniMax M3) and Novita AI (MiniMax M3); gpt-oss-120b $0.05 at Novita AI (OpenAI GPT OSS 120B); gpt-oss-20b $0.07 at Google Vertex AI (gpt-oss-20b); Llama 4 Maverick $0.25 at Azure AI Foundry (Llama4 Mavrick 17B Global); Gemma 4 31B $0.09 at DeepInfra (gemma-4-31B-it-turbo); Mistral Large 3 $0.50 at Mistral AI (Mistral Large 3) and Amazon Bedrock (Mistral Large 3); NVIDIA Nemotron 3.5 Super VL 120B-A12B $0.10 at DeepInfra (NVIDIA-Nemotron-3.5-Super-VL). Output prices differ and are listed on each model page.

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

Providers serve different builds and snapshots (dated releases, Fast and turbo builds), at different precisions (16 of 80 offers state a quantization such as FP8 or FP4), sometimes with a shorter context (6 offers), and some price by time of day or carry a promotional discount. Each of those is a separate row here, never merged.

Is this a speed or quality benchmark?

No. It records listed prices and what each provider states about the model it serves. Speed figures appear only where a provider states them, labelled Vendor-stated, and nothing here measures output quality or latency.

Which providers are not included?

Hyperbolic: hyperbolic.ai/pricing is a not-found page, and the Hyperbolic docs describe GPU rental products (On-Demand, Reserved, Private Cloud) with no per-token serverless model price list, so there is no per-token price to list. Nebius AI Studio / Token Factory: The Token Factory prices page (nebius.com/token-factory/prices) shows a "Log in to continue" wall, and the public Token Factory pages read had no per-model price table. No price is listed here rather than relying on a non-pricing page. Hugging Face Inference Providers: Hugging Face publishes no per-model rates of its own: its pricing doc says "Hugging Face charges you the same rates as the provider, with no additional fees." Prices are the underlying providers', which are covered directly. Routers and marketplaces are out of scope; the providers they route to are covered directly.

This is a directory of listed prices and what each provider states about the model it serves, not a benchmark. Prices change often; check the provider's page before you commit. Speed figures appear only where a provider states them, labelled Vendor-stated.

Last verified from each provider's own pricing and model pages; 561 quotes were string-matched against those pages that day. Parameters and native context come from our local-models records, not from provider pages.

Editorial changes to the verified research file, all disclosed here. No quote and no price was changed. (1) Notes that described how pages were read (page text, captured tabs, a render failure, a not-found status) were rewritten to describe the source page itself, and the Alibaba qwen3.8-27b note no longer names a region for the second Qwen3.8 table, which the page does not state beside it. (2) Every DeepInfra catalog card the file cites on deepinfra.com/models/text-generation was checked for a discount tag: the DeepSeek-V4.1-Flash card shows "30% off" and the GLM-5.3 card shows "38% off" beside the price, and each model page says the promotional discount is already applied. Those two prices carry a visible "Card shows N% off - may be a temporary discount" label; the numbers are unchanged. The DeepSeek-V4-Pro, DeepSeek-V4-Pro-0813, Kimi-K3, gemma-4-31B-it and NVIDIA-Nemotron-3.5-Super-VL cards show no discount tag, and gemma-4-31B-it-turbo has no catalog card on that page. (3) DeepInfra context values come from tags that carry no Context label in the quote (a header chip on the model page, or a bare "1024k" or "256k" in the pricing row), so they render as inferred. Quantization tags such as fp8 and fp4 are shown as stated because the quote is the tag. (4) Two offers are matched to a model by name only and render with an inferred tag beside the provider's own model name: Amazon Bedrock "Kimi 3" as Kimi K3, and Azure "FW MiniMax 3" as MiniMax-M3. (5) The Alibaba qwen3.8-27b price ($0.5 / $3) is attributed to a region by table order and renders as inferred; the second table's $0.424 / $1.696 is not shown as a row because the file carries no quote for it. (6) DeepSeek first-party prices are peak-hour prices and Alibaba's DeepSeek rows are busy-hour prices; both are labelled from the file's own quotes and notes, and the cheapest-price summary says so. (7) The four discount quotes in this editorial layer were string-matched against the same DeepInfra catalog and model pages on the verification day; they are additional to the 561 quotes counted in verification.

Newsletter

Know when open-weight API prices move

One email per week. New model releases, price changes and where to run them, and what they mean for what you build.

Keep exploring

Run open-weight models locally

The self-host alternative: downloads, memory needs and commands for each model.

Open-weight model licenses

Which of these models you may use commercially, and on what terms.

LLM API cost levers

Prompt caching, batch discounts and service tiers by provider.

Free LLM API tiers

Which providers have a standing free tier or credits.

OpenAI-compatible endpoints

Base URLs and limits for switching providers without rewriting code.

LLM gateways and routers

Route one model across several of these providers with fallbacks.