Cost · July 6, 2026 · 6 min read

DeepInfra Pricing in 2026: What Every Model Actually Costs

DeepInfra costs in 2026, from our weekly verified price checks: per-model rates for Llama, DeepSeek, Qwen, gpt-oss and Whisper, and where it wins or loses.

DeepInfra is the host that keeps winning our weekly price checks: of the 88 model tables we track, it holds the cheapest verified endpoint for more open-weight LLMs than any other provider. This page collects what DeepInfra actually costs across the models we verify, where it is beatable, and the catches to know before moving volume.

Numbers below are from our latest verified check (see the live rankings for the current table per model; prices move monthly).

DeepInfra LLM pricing, per 1M output tokens

ModelDeepInfraCheapest rival we track
Mistral Nemo4 centsMistral official, 15 cents
Llama 3.1 8B5 centsNovita, 5 cents
Mistral Small 38 centsMistral official, 30 cents
gpt-oss-20b14 centsNovita, 15 cents
gpt-oss-120b17 centsNovita, 25 cents
Llama 4 Scout30 centsGroq, 34 cents
Qwen2.5 72B40 centsNovita, 40 cents
Llama 4 Maverick60 centsNovita and Together, 85 cents
DeepSeek-V389 centsNovita, 1.12 dollars
Qwen3 Coder 480B1.00 dollarNovita, 1.30 dollars
Kimi K22.00 dollarsNovita, 2.50 dollars
DeepSeek-R12.15 dollarsNebius, 2.40 dollars

Two patterns stand out. First, DeepInfra is frequently 50 to 75 percent below the model maker's own endpoint (Mistral Small 3 at 8 cents versus 30 official). Second, its margin over the next-cheapest host is usually small, which means you are not locked in: if DeepInfra changes prices, the switch away is cheap too.

Transcription: the quiet standout

DeepInfra serves Whisper large v3 at $0.00045 per minute in our latest check, the lowest rate we track anywhere, several times below Groq's $0.0019 and roughly 13x below OpenAI's hosted Whisper. If you transcribe at volume, this single row can dominate your savings. Compare the full table on the Whisper ranking page.

Where DeepInfra does not win

  • Frontier models: no GPT, Claude, or Gemini; it serves open weights only.
  • Peak speed: Groq and Cerebras deliver several times the tokens per second on the models they share. If latency is your constraint, cheap-per-token is not the metric.
  • A few tables: OpenRouter routes Llama 3.3 70B lower at 32 cents, and Hyperbolic leads Qwen3 235B.

Things to know before switching

  • The API is OpenAI-compatible; switching is a base URL and key change for most stacks.
  • Rate limits scale with prepaid balance; check your tier before pointing production at it.
  • Prices genuinely move: our last two checks caught DeepInfra cutting gpt-oss-120b from 19 to 17 cents and Mistral Small 3 from 20 to 8 cents, but also small increases elsewhere. Bookmark the live rankings rather than this snapshot.

Bottom line

DeepInfra is the default answer to "where do I run open-weight models cheaply" in 2026: most-frequent winner across our verified tables, dramatic savings on Whisper, OpenAI-compatible switching. Verify the current number for your exact model on the rankings, and if you are early stage, stack startup credits from the catalog before you spend cash anywhere.

Related reading: cheapest AI inference APIs, Novita AI pricing, LLM API pricing comparison.

Frequently asked questions

How much does DeepInfra cost?

Pay-as-you-go per token, per model: from 4 cents per 1M output tokens for small models (Mistral Nemo) to about 2.15 dollars for DeepSeek-R1 in our latest verified check, plus $0.00045 per minute for Whisper large v3 transcription. No subscription required.

Is DeepInfra cheaper than Together AI or Fireworks?

On most shared models in our tables, yes, often by 30 to 70 percent per output token. Together and Fireworks counter with broader enterprise features and startup credit programs, so the right answer can still differ per team.

Is DeepInfra reliable enough for production?

Many production teams run on it. Do your own soak test: rate limits scale with your prepaid balance, and speed is mid-pack rather than Groq-class. The switching cost is low either way since the API is OpenAI-compatible.

Does DeepInfra serve GPT or Claude?

No. DeepInfra hosts open-weight models only: Llama, DeepSeek, Qwen, Mistral, gpt-oss, Whisper and similar. For frontier models you buy from OpenAI, Anthropic, or Google directly, ideally with batch discounts.

Where do these numbers come from?

Perkstack re-checks every provider's posted pricing weekly across 88 model tables and pins the exact number with a dated verification. The live rankings page always has the current table.

Keep reading

Building on AI? Don't pay full price.

Perkstack tracks 200+ verified AI credits, free signup credits and startup grants, each with a step-by-step claim guide.