AI Inference Providers
The model you pick matters less than who serves it. Compare current pricing, speed, free tiers, and native n8n nodes across every major inference provider, refreshed for the gpt-oss, Llama 4, and DeepSeek era.
Why Providers Matter
Most people building AI workflows in n8n obsess over which model to use. But the question that determines your actual costs, latency, and reliability is different: who is serving that model?
An AI inference provider is the cloud infrastructure between your n8n workflow and the model weights. It handles GPU (and increasingly custom-silicon) allocation, request queuing, batching, and rate limits. The same open model, say OpenAI's gpt-oss-120B, can cost roughly $0.09 per million tokens on the cheapest provider and $0.35 or more on the fastest one, and stream at under 200 tokens per second on one and over 1,500 on another.
2026 changed the math. Open-weight models now rival the proprietary flagships on everyday workflow tasks. OpenAI's Apache-licensed gpt-oss, Meta's Llama 4, DeepSeek V3 and R1, Alibaba's Qwen3, Moonshot's Kimi K2, and Z.ai's GLM are all served for cents by a dozen competing providers. Choosing the right one is now the single biggest lever you have on cost and speed.
Why not just use OpenAI or Anthropic directly? You can, and for their proprietary models (GPT-5, Claude) you often have to. But going direct locks you into one lab's models, pricing, and outages. Inference providers give you hundreds of models through a single API key, usually cheaper, and keep your workflows running when any one lab has an incident.
The good news for n8n: n8n now ships native Chat Model sub-nodes for OpenAI, Anthropic, Google Gemini, Groq, Mistral, DeepSeek, OpenRouter, and Ollama, plus an OpenAI-compatible option for any provider without a dedicated node. You rarely need a raw HTTP Request node anymore.
Every model named in this guide is live and serving traffic as of August 2026. Provider catalogs move fast: models get deprecated and replaced constantly, so we have dropped the retired names (Mixtral 8x7B, Gemma 2, Llama 3.1) and list only what is currently available. Always confirm on a provider's live models page before you build.
Prerequisite: This article assumes you have chosen your model or model family. If not, start with How to Choose the Right AI Model for n8n first, then come back here to pick the best provider for that model.
The 2026 Pricing Landscape
All prices are per million tokens. Input and output are billed separately by most providers. To compare fairly we anchor on gpt-oss-120B, OpenAI's Apache-2.0 open model that is hosted, live, by nearly every provider on this list, so it is the closest thing to a common benchmark. "Free tier" is what you get without paying.
| Provider | Free Tier | gpt-oss-120B (In / Out) | Also Serves | Speed |
|---|---|---|---|---|
| OpenRouter | 50–1,000 free req/day | routes from ≈$0.09 / $0.45 | GPT-5, Claude, Gemini + 400 open | Fast (routing) |
| DeepInfra | Signup credit | ≈$0.09 / $0.45 | Llama 4, DeepSeek, Qwen3, Kimi K2 | ~190 tok/s |
| Novita AI | Signup credit | ≈$0.10 / $0.50 | Open weights + image / video / audio | Fast |
| Groq | Daily free token limits | $0.15 / $0.60 | gpt-oss-20B ($0.075/$0.30), compound | ~500 tok/s, ~0.7s to first token |
| Cerebras | Free trial + PAYG | ≈$0.35 blended | Qwen3, Llama, gpt-oss | ~1,500–3,000 tok/s |
| Together AI | Small signup credit | $0.15 / $0.60 | 200+ open models, fine-tuning | Fast |
| Fireworks AI | Pay-as-you-go | ≈$0.15 / $0.60 | Open weights, tool calling | Very fast |
| Mistral (Direct) | Free tier (rate-limited) | own models | Large 3, Medium 3, Codestral | Fast (EU) |
| Hugging Face | Serverless free | varies | 1M+ models via Inference Providers | Varies |
| Ollama (Local) | Free forever | $0 | Any open weight on your hardware | Hardware-bound |
Read these as representative, not fixed. The Groq and Together numbers are their published gpt-oss-120B rates; the ≈ values are the current low-to-mid market range and shift month to month as providers compete. Cerebras publishes a blended per-token rate rather than split input/output. Always confirm on the provider's own live pricing page before committing high-volume workloads.
Reading these prices: $0.15/M input tokens means processing 1 million input tokens costs 15 cents. A typical n8n AI step processes 500–2,000 tokens. At $0.15/M, running 10,000 executions of ~1,000 tokens each costs about $0.0015 total, a fraction of a cent. In practice your VPS or n8n Cloud bill almost always dwarfs the inference cost.
OpenRouter (The Aggregator)
If you only set up one inference provider in n8n, make it OpenRouter. It is not a model lab; it is a unified routing layer on top of virtually every major provider. A single API key reaches 400+ models: OpenAI (GPT-5), Anthropic (Claude), Google Gemini, Mistral, DeepSeek, Qwen3, Llama 4, Kimi K2, GLM, gpt-oss, and hundreds more.
n8n integration: n8n now ships a dedicated OpenRouter Chat Model node, so you just add your key and pick a model from the dropdown. (You can still use the OpenAI node pointed at base URL https://openrouter.ai/api/v1 if you prefer.) It works with the AI Agent, Basic LLM Chain, and every other AI node.
Price routing and fallback: for any model served by multiple providers, OpenRouter routes to the cheapest or fastest at real-time market prices, and if that provider errors mid-request it transparently retries on a backup. That is resilience you would otherwise build by hand with n8n error branches.
What it actually costs: OpenRouter does not mark up token rates, you pay the underlying provider's price. Its revenue comes from a 5.5% fee when you top up credits with a card (minimum $0.80 per purchase) and a small fee only on very high "bring your own key" volume, which is free up to $25,000 of list-price inference per month. For normal n8n usage the effective overhead is negligible.
Free tier: OpenRouter hosts 25+ :free model variants. Accounts with under $10 of credit get 50 free requests per day; adding $10 or more raises that to 1,000 free requests per day, at a 20 requests-per-minute cap. That covers a lot of testing before you pay a cent.
When NOT to use OpenRouter: when you need the absolute lowest latency (a direct Groq or Cerebras key shaves the routing hop), when you need guaranteed EU-only data processing (use Mistral Direct), or when you want fully local inference (use Ollama).
The Speed Tier: Groq & Cerebras
Two providers run custom silicon built purely for inference, and they are in a different league on speed. When a node sits on your workflow's critical path, or you are streaming a response back to a user, this is where you send it.
Groq runs its own LPU (Language Processing Unit) hardware and optimizes for time-to-first-token: responses start in around 0.7 seconds and stream at roughly 500 tokens per second, versus 60–80 tok/s on typical GPU inference. Groq keeps its catalog deliberately small and current. The live text models today are gpt-oss-120B ($0.15/$0.60), gpt-oss-20B ($0.075/$0.30), and the Groq Compound agentic systems (with built-in web search and code execution), alongside Whisper for speech. The older Mixtral and Gemma 2 models it used to host have been retired.
Cerebras runs wafer-scale (WSE) chips and optimizes for raw throughput: sustained rates of 1,500 to 3,000 tokens per second on open models like Qwen3, Llama, and gpt-oss. If you are pushing large batches or long generations, Cerebras finishes first. Pricing is a blended per-token rate (around $0.35/M on gpt-oss-120B), higher than the budget tier but worth it when speed is the product.
n8n integration: Groq has a dedicated Groq Chat Model node, just drop in your key. Cerebras is OpenAI-compatible: use the OpenAI node with base URL https://api.cerebras.ai/v1. Both work with the AI Agent and LLM Chain nodes.
Free tiers: Groq offers a genuinely usable free tier measured in daily token limits (enough to run, say, an entire classification pipeline on gpt-oss-20B at no cost). Cerebras offers a free trial plus pay-as-you-go.
Limitation: both serve open-weight models only, no Claude, GPT-5, or Gemini. Use them for the open models and route proprietary steps through OpenRouter or the lab's own node. When speed matters, reach for Groq for the fastest first token and Cerebras for the highest sustained throughput.
The Budget Tier: DeepInfra & Novita
In 2026 the price floor moved. Two providers consistently sit at the bottom of the market, and for high-volume n8n workflows the savings are real.
DeepInfra is the current cost leader with one of the broadest live catalogs. A snapshot of its rates shows how cheap open models have become:
| DeepInfra Model | Input $/M | Output $/M | Good For |
|---|---|---|---|
| Llama 4 Scout | $0.10 | $0.30 | Very long context |
| Llama 4 Maverick | $0.20 | $0.80 | Multimodal, general use |
| DeepSeek V4 Flash | $0.08 | $0.18 | Cheapest fast general model |
| DeepSeek R1 | $0.50 | $2.15 | Step-by-step reasoning |
| Kimi K2 | $0.45 | $2.25 | Long-horizon agents, coding |
DeepInfra also lets you trade price for latency with Standard, Priority (1.5x), and Flex (0.8x) service tiers. It is OpenAI-compatible: use the OpenAI node with base URL https://api.deepinfra.com/v1/openai.
Novita AI sits right next to DeepInfra on price and adds breadth beyond text: it serves LLMs plus image, video, and audio models through one API. If a workflow mixes an LLM step with, say, image generation or transcription, Novita can cover both from a single key. Its LLM endpoints are OpenAI-compatible too.
When to use the budget tier: high-volume, latency-tolerant steps, batch classification, enrichment, summarization, and tagging across thousands of records, where shaving cents per million tokens compounds. For a single user-facing response, the speed tier is usually worth the small premium.
Together AI & Fireworks
Between the rock-bottom budget tier and the custom-silicon speed tier sits the production tier: broad, reliable, feature-rich platforms that cost a little more but are built for teams running real workloads.
Together AI serves 200+ open models and pairs inference with fine-tuning on the same platform. Current live models include gpt-oss-120B ($0.15/$0.60), the DeepSeek V4 and R1 line, Qwen3, Kimi K2, GLM, and Llama 4. Two things make it stand out for n8n:
- Cheap embeddings for RAG. Together's embedding models run around $0.02 per million tokens, among the cheapest anywhere, which matters when you are indexing large document sets into a vector store.
- Fine-tuning as a service. Upload training data, fine-tune a base model, and serve it through the same endpoint, useful for domain-specific accuracy (support, legal, financial text) inside an n8n workflow.
Fireworks AI focuses on production inference with strong function calling and structured (JSON) output. For n8n AI Agents that call tools using open-weight models, Fireworks reduces schema-hallucination errors more reliably than most, so agent tool calls actually return valid JSON. Pricing sits in the same range as Together, pay-as-you-go with no upfront credit.
Worth knowing: Baseten. If you need a specific model on dedicated GPUs (guaranteed capacity, custom serving, per-second billing) rather than shared serverless, Baseten is the go-to. It is overkill for most n8n setups but the right call for latency-sensitive, high-throughput production.
n8n integration: both are OpenAI-compatible. Together's base URL is https://api.together.xyz/v1; Fireworks is https://api.fireworks.ai/inference/v1. Use the OpenAI node for either.
Mistral & Hugging Face
Mistral AI (Direct) is the European lab of record, and for teams with EU data-residency or GDPR requirements it is the cleanest choice: data is processed on EU infrastructure and n8n has a dedicated Mistral Cloud Chat Model node. Its current lineup is a tidy ladder rather than a sprawl:
| Mistral Model | Tier | Best For |
|---|---|---|
| Mistral Large 3 | Flagship (~$2 / $6) | Complex reasoning, coding, agents |
| Mistral Medium 3 | Balanced | Everyday general-purpose work |
| Mistral Small | Cost-effective | Classification, extraction, tagging |
| Magistral | Reasoning | Step-by-step, math-heavy tasks |
| Codestral | Code | Code generation and completion |
| Ministral 3B / 8B | Edge | Cheap, fast, on-device |
(The older Mistral Large 2 and Nemo have been superseded, though Nemo is still downloadable as an open-weight model.)
Hugging Face hosts the world's largest model repository (over a million checkpoints) and now fronts it with Inference Providers, a router that serves models through partner providers behind a single HF token. Reach for it when you need a niche or domain-specific model, a fine-tuned medical NLP model, a low-resource language, a specialized classifier, that no mainstream provider serves. The serverless free tier throttles under load; for production, Dedicated Inference Endpoints bill per GPU-hour and can get expensive.
Ollama (Local Inference)
Ollama is not a cloud provider. It is a local runtime that downloads and runs open-weight models directly on your hardware. Paired with a self-hosted n8n, it is the ultimate zero-cost, zero-data-leakage AI stack.
Privacy: zero data leaves your infrastructure, every token is generated on your hardware. For workflows touching financial records, PII, or proprietary IP, Ollama is the only architecturally airtight option.
n8n integration: n8n has a dedicated Ollama Chat Model node. Point the credential at your Ollama server (default http://localhost:11434) and it works natively with AI Agent orchestration.
What to run in 2026: the sweet spot for local is gpt-oss-20B (its MXFP4 quantization fits in roughly 16 GB and it is genuinely capable), with Qwen3, Llama 4 Scout, Gemma 3, and DeepSeek-R1 distills as strong alternatives. Your only cost is hardware:
| Hardware | Memory | Comfortable Models | Speed | Approx. Cost |
|---|---|---|---|---|
| CPU only | 16 GB RAM | gpt-oss-20B, Qwen3 8B (Q4) | ~5-8 tok/s | $0 |
| RTX 4060 Ti / 5060 Ti | 16 GB VRAM | gpt-oss-20B, 14-24B (Q4) | ~40 tok/s | ~$450 |
| RTX 5090 | 32 GB VRAM | 32B dense, 70B (Q4) | ~55 tok/s | ~$2,000 |
| Mac M4 Max | 128 GB unified | gpt-oss-120B, large MoE (Q4) | ~25 tok/s | ~$4,000+ |
Limitation: performance is entirely bound by your hardware. Without a dedicated GPU, inference is 10-100x slower than any cloud provider, and Ollama cannot run proprietary models like GPT-5 or Claude. Use it for privacy and zero marginal cost, not for peak speed.
Which to Pick
Do not marry one provider. Match the provider to the job, and let n8n use a different credential on each AI node:
OpenRouter
One key, 400+ models, price routing and automatic fallback. The best starting point, and the only provider that also reaches GPT-5, Claude, and Gemini.
DeepInfra / Novita
Rock-bottom per-token pricing for batch and enrichment steps running across thousands of records. Novita also covers image, video, and audio.
Groq / Cerebras
Groq for the lowest time-to-first-token on user-facing steps; Cerebras for the highest sustained throughput on big batches.
Ollama + Self-Hosted n8n
For workflows touching personal data, financial records, or proprietary IP. Zero data leaves your infrastructure.
Mistral (Direct)
EU-hosted, GDPR-friendly, native n8n node. The default for European businesses processing regulated data.
Together AI / Fireworks
Together for cheap embeddings and custom fine-tunes; Fireworks for reliable JSON and tool calling in AI Agents.
The recommended architecture: use OpenRouter as your primary key for instant access to everything through one integration, then add a DeepInfra key for cheap high-volume steps, a Groq key for anything speed-critical, and Ollama or Mistral where privacy or EU compliance demands it. Each n8n AI node picks the credential that fits.
FAQ
Are all the models in this guide still available?
Yes. Every model named here is live and serving traffic as of August 2026, and we deliberately removed the ones that have been retired (Mixtral 8x7B, Gemma 2, Llama 3.1). Catalogs do change often, so a model can be deprecated between updates. Before you build, open the provider's live models page and confirm the exact model ID.
Can I use multiple providers in the same n8n workflow?
Yes. Each AI node can use its own credential. A common pattern: run a batch classification step on DeepInfra (cheap), a user-facing reply on Groq (fast), a complex reasoning step on OpenRouter (GPT-5 or Claude), and a private step on Ollama. Each node is independent, so mix freely.
Which free model should I start with in n8n?
Three genuinely free options with no card required: gpt-oss / Llama on Groq (generous daily token limits), Gemini 2.5 Flash on Google AI Studio (native n8n Google Gemini node, large free tier), and OpenRouter's :free model variants (50-1,000 requests/day). Any is enough to build and test a real workflow at zero cost.
Is OpenRouter just a reseller adding markup?
No. OpenRouter does not mark up token rates, you pay the underlying provider's price, and it routes to the cheapest provider serving your model. Its fee is 5.5% when you buy credits with a card (min $0.80) plus a small charge only on very high bring-your-own-key volume. For typical n8n usage the overhead is negligible, and the routing plus automatic failover more than pay for it.
What happens if my provider goes down mid-workflow?
The execution fails at that node. Handle it with n8n's error handling: set the AI node's "On Error" to continue to a fallback branch, or catch it with an Error Trigger. If you route through OpenRouter, its built-in failover retries a backup provider automatically before your workflow ever sees an error.
How much does AI inference actually cost in n8n?
Very little for open models. A typical AI node uses 500-2,000 tokens. At $0.15/M (gpt-oss-120B), running 10,000 executions of ~1,000 tokens each costs under a dollar; on the budget tier (~$0.09/M) it is cheaper still. Even a proprietary flagship like Claude or GPT-5 at around $5/M input runs 100,000 executions for a few tens of dollars a month. Your hosting bill almost always exceeds inference.
Do these providers support embeddings for RAG?
Yes. Together AI, DeepInfra, and Fireworks all serve embedding models, and Together's are among the cheapest at around $0.02/M, which matters when indexing large document sets. Ollama runs local embedding models like nomic-embed-text for zero-cost, private RAG. Point your n8n vector store's Embeddings node at whichever you choose.
Can I run Ollama on the same VPS as n8n?
Only if the VPS has a GPU (rare and pricey) or you accept slow CPU-only inference. Most VPS plans are CPU-only. For usable local inference, run Ollama on a separate machine with a dedicated GPU and connect n8n to it over your network. gpt-oss-20B on a 16 GB GPU is a great starting point.
