DeepInfra
Serverless API for 200+ open-weight models on DeepInfra's own US hardware, priced per token.
DeepInfra runs 200+ models as serverless, OpenAI-compatible API endpoints on its own hardware in US data centers. The catalogue covers current open-weight families (DeepSeek V4.1, Kimi K3, GLM 5.3, Qwen, Llama, Gemma, Nemotron) plus embeddings, image generation, text-to-speech and speech recognition.
Pricing is per token and set per model, from $0.02 per 1M input tokens for small Llama models. Requests can run on a Priority tier at 1.5x the price or a Flex tier at 0.8x. Dedicated GPUs and clusters are also for rent.
Pricing: Per token usage
Data and privacy
Checked 17 November 2026- Data processed
- US only (DeepInfra's own GPUs in US data centers; no EU or other region) source
- DPA
- Read the DPA โ
- Sub-processors
- View the list โ
- US parent company
- Yes
Read from DeepInfra's own legal pages. "Not public" means it is not on their site; some providers share it on request.
Choosing an inference API right now?
One click, anonymous. Helps us know our visitors.
Thanks, that helps.
Choosing an inference API right now?
One click, anonymous. Helps us know our visitors.
Thanks, that helps.
DeepInfra prices by model
Per 1M tokens, read off DeepInfra's own pricing page on the date shown.
| Model | Input / 1M | Cached / 1M | Output / 1M | Checked | Notes |
|---|---|---|---|---|---|
| DeepSeek V4 Pro | $1.30 | $0.10 | $2.60 | 14 Nov 2026 | |
| DeepSeek V4.1 Flash | $0.14 | $0.0042 | $0.42 | 14 Nov 2026 | Promotional price, 30% off list $0.20 / $0.60 |
| DeepSeek V4.1 Flash 1017 | $0.098 | $0.003 | $0.294 | 17 Nov 2026 | Promotional price, 51% off. List $0.20 / $0.60 / cached $0.006. |
| GLM 5.3 | $0.5625 | $0.125 | $2.50 | 14 Nov 2026 | Promotional price, 38% off list $0.90 / $4.00 |
| GLM 5.3 Flash | $0.15 | $0.03 | $0.50 | 14 Nov 2026 | |
| Gemma 4 26B A4B | $0.07 | โ | $0.34 | 16 Nov 2026 | |
| Gemma 4 31B | $0.20 | โ | $0.40 | 16 Nov 2026 | |
| Kimi K3 | $2.85 | $0.285 | $14.25 | 14 Nov 2026 | Cache write is for 5-minute retention (1.25x input); 1-hour retention is $5.70 |
| Kimi K3.1 Pro | $1.90 | $0.19 | $9.50 | 17 Nov 2026 | Promotional price, 5% off. List $2.00 / $10.00 / cached $0.20. |
| Llama 3.3 70B Instruct | $0.10 | โ | $0.32 | 14 Nov 2026 | Llama-3.3-70B-Instruct-Turbo (FP8); the full-precision version was retired and redirects here |
| MiniMax M3 | $0.28 | $0.056 | $1.10 | 14 Nov 2026 | |
| Qwen3.8 27B | $0.15 | $0.038 | $1.875 | 14 Nov 2026 | Promotional price, 25% off list $0.20 / $2.50 |
| gpt-oss-120b | $0.037 | โ | $0.17 | 14 Nov 2026 | |
| gpt-oss-20b | $0.03 | โ | $0.14 | 14 Nov 2026 |
What is DeepInfra?
DeepInfra is a serverless inference platform for open-weight models. You call an OpenAI-compatible API and pay per token, with nothing to provision. It runs on its own hardware in US data centers. It raised a $107M Series B in May 2026, co-led by 500 Global and Georges Harik.
Models
The catalogue lists 200+ models. Text models include DeepSeek V4.1 Flash and V4 Pro, Kimi K3 and K3.1 Pro, GLM 5.3, Qwen, Llama 3 and 4, Gemma, NVIDIA Nemotron, Xiaomi MiMo and Mistral. The pricing page also lists Google Gemini and Anthropic Claude models. Beyond text there are embedding models, Flux image generation, text-to-speech and speech recognition.
Model pages state the quantization each model is served at, such as FP8 or FP4, which matters when comparing prices across hosts. Several popular models run at a limited-time discount, shown as a percentage off on the model page.
Pricing
Prices are per million tokens and set per model. The cheapest text model, Llama 3.1 8B Instruct Turbo, costs $0.02 in and $0.04 out; Kimi K3 costs $2.85 in and $14.25 out. Cached input is billed at a lower rate where the model supports caching.
Each request can pick a service tier: Standard at the base price, Priority at 1.5x for faster time-to-first-token under load, or Flex at 0.8x for asynchronous work that tolerates slower responses and occasional unavailability. Prices read on 14 November 2026 from deepinfra.com/pricing.
Dedicated GPUs
For custom or fine-tuned models, DeepInfra rents dedicated GPUs billed by the minute: H100 at $2.90 and B200 at $4.89 per GPU-hour. Larger DGX clusters are quoted on request. Its DeepCluster offer sells B300 capacity on five-year terms at $1.98 per GPU-hour.
Privacy and compliance
DeepInfra states a zero-retention policy for inputs and outputs. It holds SOC 2 and ISO 27001 certification.
Who it suits
Teams that want open-weight models behind a familiar API, priced per token, without running GPUs. Check the served quantization and whether a price is a promotion before building a budget on it.
DeepInfra Alternatives
Explore 173 products in the Inference APIs category. View all DeepInfra alternatives.
LibertAI
Decentralized, privacy-first inference API running open-source LLMs in trusted execution environments
Inceptron
EU-hosted LLM inference with an in-house compiler: serverless open-weight model APIs and dedicated GPU deployments
Fireworks AI
The production AI platform built for developers.
Groq
LPU-powered inference API for LLMs, speech, and vision models with usage-based pricing
BentoML
BentoML is the platform for software engineers to build AI products.
Compare
Work on DeepInfra? Feature it at the top of Inference APIs.
Is your product missing?