OpenRouter Free Models (2026): Best Free APIs & Setup Guide

Short answer: OpenRouter is the fastest way to call many LLMs through one API key, including free or :free model routes when providers allow it. Budget 15 minutes to sign up, copy a key, and send your first request. Skill level: you can run curl or paste Python into a notebook.

We last verified the models page and a live API call on June 2, 2026. Free IDs change; treat the table below as a pattern, not a permanent list.

Last updated: June 2, 2026.

What you need

Quick comparison: free vs paid routing

For how those models compare on coding, see DeepSeek V4 vs ChatGPT vs Claude and the latest AI models hub.

Step 1: Create an account and key

1. Go to openrouter.ai and sign in.
2. Open Keys in the dashboard.
3. Click Create Key, name it (example: aitoolsradar-dev), copy once.
4. Export locally:

export OPENROUTER_API_KEY="sk-or-v1-xxxxxxxx"

Expected result: The keys page shows your new key with create date. Revoke any key you pasted into a screenshot by mistake.

Common mistake: Committing the key to GitHub. Use .env and add .env to .gitignore.

Step 2: Find free models on the catalog

1. Open Models in the dashboard (or visit the public models page).
2. Sort or filter by price ascending.
3. Look for :free suffixes or $0 input and output on the row you want.
4. Copy the model ID string exactly, including provider prefix.

Expected result: You have one slug ready for model in JSON, such as deepseek/deepseek-v4-flash:free or openrouter/free (auto-picks a free route) when offered.

Common mistake: Using an old blog slug after the provider renamed the checkpoint. The catalog wins over blog posts.

Step 3: Send a test request with curl

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -H "HTTP-Referer: https://aitoolsradar.org" \
  -H "X-Title: AI Tools Radar Test" \
  -d '{
    "model": "REPLACE_WITH_FREE_MODEL_ID",
    "messages": [
      {"role": "user", "content": "Reply with exactly: openrouter ok"}
    ]
  }'

Swap REPLACE_WITH_FREE_MODEL_ID for the slug you copied.

Expected result: JSON with choices[0].message.content containing your reply.

Common mistake: Forgetting the Authorization header. You get 401 with a short error body.

Step 4: Use the OpenAI Python SDK

Install once:

pip install openai

Script:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
)

resp = client.chat.completions.create(
    model="REPLACE_WITH_FREE_MODEL_ID",
    messages=[
        {"role": "user", "content": "Write a one-line Python hello world."}
    ],
    extra_headers={
        "HTTP-Referer": "https://aitoolsradar.org",
        "X-Title": "AI Tools Radar Script",
    },
)

print(resp.choices[0].message.content)

Expected result: Terminal prints code or text from the free model.

Common mistake: Pointing the default OpenAI client at api.openai.com while using an OpenRouter model slug. Base URL must be OpenRouter.

Step 5: Route drafts cheap, finals expensive

A pattern we use in June 2026:

1. Classify the job: draft vs final, public vs confidential.
2. Draft with a free or Flash model (deepseek-v4-flash, Llama free route, Qwen coder when listed at $0).
3. Final with gpt-5.5 or claude-opus-4-8 on OpenRouter or direct API.
4. Log model per request in your app logs for cost audits.

Example two-step Python shape:

DRAFT_MODEL = "REPLACE_WITH_FREE_OR_FLASH_ID"
FINAL_MODEL = "openai/gpt-5.5"  # verify slug on catalog

def draft(prompt: str) -> str:
    return chat(DRAFT_MODEL, prompt)

def finalize(draft_text: str, rubric: str) -> str:
    return chat(
        FINAL_MODEL,
        f"Improve this answer.\nRubric: {rubric}\n\n{draft_text}",
    )

Expected result: Lower monthly bill with acceptable quality on internal tools.

OpenRouter pricing table (June 2026 signals)

Verify live before you publish a budget doc.

OpenRouter also shows per-request fees and context pricing on each model card. Long prompts on 1M-context models can still cost real money even when per-token rates look tiny.

Use OpenRouter in Cursor (optional)

1. Open Cursor Settings → Models (wording may vary by version).
2. Add an OpenAI-compatible custom provider if available, or use the OpenRouter base URL field when present.
3. Base URL: https://openrouter.ai/api/v1
4. API key: your OPENROUTER_API_KEY
5. Model: paste the catalog slug.

Expected result: Inline chat uses the slug you entered.

Common mistake: Assuming every free model supports tool calling the way GPT-5.5 does. Run one tool-heavy prompt as a smoke test.

When to skip OpenRouter

• Your legal team requires a direct DPA with OpenAI or Anthropic only.
• You need a fixed model version for twelve months with no catalog renames.
• You run high-volume production traffic where router markup matters at scale and direct contracts are cheaper.

When you skip, still read the latest AI models hub for capability context.

Troubleshooting

Prompt templates for free-model workflows

Summarize logs (cheap)

Summarize this log in five bullets: first error, likely cause, suggested fix.
Do not invent file names not present in the log.

[paste log]

Draft only disclaimer

Draft an answer. Mark uncertain claims with [verify].
I will run a second pass on a frontier model.

Free model patterns we see in June 2026

The catalog rotates. These patterns repeat even when exact slugs change:

How we pick a free route in practice

1. Open the models page and filter price ascending.
2. Copy three candidates with the same task (one coding prompt).
3. Log latency, refusal rate, and answer quality in a spreadsheet.
4. Promote the winner to .env as DRAFT_MODEL for two weeks.
5. Re-run step 1 every Monday; free routes disappear without email notice.

Node.js and TypeScript snippet

For small internal tools:

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://openrouter.ai/api/v1",
  apiKey: process.env.OPENROUTER_API_KEY!,
});

const model = process.env.DRAFT_MODEL ?? "REPLACE_WITH_FREE_MODEL_ID";

const completion = await client.chat.completions.create({
  model,
  messages: [{ role: "user", content: "List three risks of using free LLM APIs." }],
});

console.log(completion.choices[0]?.message?.content);

Expected result: Console prints a short list. Swap model at runtime for A/B tests.

Common mistake: Hardcoding a free slug in ten repos. Centralize DRAFT_MODEL in one secrets manager.

Security and key hygiene

• Rotate keys after any contractor offboarding or leaked screenshot.
• Scope keys per environment (dev, staging, prod) in the OpenRouter dashboard when available.
• Never send customer PII through free models to “save money.” Free tiers still leave your network.
• Log retention: OpenRouter may retain metadata per their policy. Read it before HIPAA or EU client work.
• Compare with direct OpenAI/Anthropic enterprise DPAs when legal asks “who is subprocessors?”

We are not lawyers. When in doubt, block the router and use an approved direct API.

Monitor spend before it surprises you

1. Open Activity or Usage in the OpenRouter dashboard weekly.
2. Set a credit alert if the UI offers it, or keep a manual $20 top-up cap for experiments.
3. Tag requests in your app logs: model, route, user_id, feature.
4. Graph tokens per feature. Free-to-paid promotion often happens when one feature 10x spikes tokens.
5. If spend jumps, check for an agent loop calling gpt-5.5 every thirty seconds before you blame the free tier.

Rule we use at AI Tools Radar: Free models for internal drafts only. Anything client-facing goes through a paid frontier route with a human spot-check.

OpenRouter vs Together vs Groq (one paragraph each)

Together is strong when you already host fine-tunes or want open models on their cloud. You lose the single catalog OpenRouter provides.

Groq wins on speed for supported Llama/Mixtral chips. Great for latency demos, not always the full model list OpenRouter carries.

Fireworks is another dev favorite for fast open-weight inference. Same story: compare catalog, not brand loyalty.

OpenRouter’s edge is one integration surface for experiments. Pick a second provider only when a specific model is not listed or enterprise pricing beats the router at your volume.

Variations

• Together, Groq, Fireworks: Same draft/final pattern without OpenRouter if your team already has credits there.
• Local Llama: Zero API cost if you have GPU time; slower setup.
• Direct DeepSeek API: Skip router markup when you only need one vendor; see DeepSeek comparison.
• LiteLLM proxy: Wrap OpenRouter behind LiteLLM if you want one internal gateway for ten microservices.

Verdict

Use OpenRouter free models when you are learning, prototyping, or shaving cost on non-critical drafts. Add paid credits when you ship agents, tool loops, or client-facing features. Pair with GPT-5.5 or Claude for finals when a free route drifts on code quality.

We keep a weekly eye on the catalog in our June Week 1 radar and bump this guide when major free routes change. Pair with GPT-5.5 for Excel (2026) for spreadsheet finals and Make Money with AI Tools (2026) if you sell API-assisted freelance work.

Changelog

• 2026-06-02: Fact-check. Updated verification date; refreshed June 2026 free-model examples (openrouter/free, deepseek-v4-flash:free, Gemma/Nemotron/OSS routes). Confirmed openai/gpt-5.5 on OpenRouter catalog.
• 2026-05-29: Initial publish. Documented account setup, curl and Python examples, pricing table patterns, Cursor notes, draft/final routing, eight FAQs.