Free Models on OpenRouter: What's Available and Their Limits
OpenRouter's free model variants (slugs ending in :free) let you exercise the unified API at $0 inference cost, under strict platform rate limits and variable provider capacity.
Use them for learning, wiring, and light evals - not as a silent production backend for multi-turn agents.
All numeric limits below are verify at build against OpenRouter limits docs.
How to Use This List
- Confirm a model still offers a
:freevariant in the live catalog before coding against it. - Size prototypes to the requests/day cap, not marketing model quality.
- Read the privacy notes on free endpoints; many log or train on prompts.
- Promote winning prompts to paid slugs before load tests or user traffic.
Platform Rate Limits (Free Variants)
Limits apply to free model variants (IDs ending in :free), governed by lifetime credits purchased on the account:
| Credits purchased (all time) | Requests per minute | Requests per day |
|---|---|---|
| Less than 10 | 20 | 50 |
| At least 10 | 20 | 1000 |
Notes:
- RPM stays tight even after the daily bump.
- Creating extra accounts or keys does not bypass global free capacity governance.
- Cloudflare DDoS protection can still block extreme burst behavior.
- A negative credit balance can block requests, including free models.
- Exceeding caps returns 429 with rate-limit style errors.
What "Free" Usually Means
| Topic | Expectation |
|---|---|
| Inference price | $0 on the free variant |
| Quality | Often related to a paid sibling, not guaranteed identical hosting |
| Capacity | Best-effort; congestion and provider policy apply |
| Data policy | Frequently more permissive logging/training - read model cards |
| Production SLA | Not a substitute for paid endpoints |
| Catalog churn | Free hosts appear and disappear; pin and re-verify |
Finding Free Models
| Method | Use |
|---|---|
| Site filter | Models browser with max price = 0 |
| Slug suffix | Look for :free on the model id |
| Models API | GET /api/v1/models then filter id or pricing fields |
| Free models router | openrouter/free (and related cookbook routers) auto-picks a free model when available - verify at build |
# Conceptual filter - field names can evolve; verify at build
free_ids = [m["id"] for m in models if m["id"].endswith(":free")]Fit Matrix for Agent Work
| Workload | Free tier OK? | Why |
|---|---|---|
| SDK smoke test | Yes | Few requests |
| Prompt sketching | Yes | Interactive, low QPS |
| Offline eval of 20 cases | Maybe | Watch the 50/day ceiling |
| Multi-turn tool agent demo | Risky | Each turn burns a request |
| CI on every PR | Risky | Shared daily budget |
| Production user traffic | No | Caps + capacity + data policy |
| Load / soak tests | No | Use paid variants |
Agent Math (Rule of Thumb)
A ReAct-style agent with 1 model call per turn and max_turns=8 can burn 8 free requests for one task.
At 50 requests/day, that is roughly 6 full agent runs/day before hard failure - before retries and eval sweeps.
Budget free tier for integration proof, then switch slugs.
Operational Checklist
- Use a dedicated API key labeled
free-devso Activity filters stay clean. - Log every 429 with model slug and turn index.
- Prefer paid small models once the loop works.
- Do not put customer PII through free endpoints without legal review.
- Re-check
:freeavailability in CI weekly; skip tests if the variant vanishes. - If you need ~1000 free RPD, purchase the documented credit threshold (commonly 10 - verify at build) even if you stay on free models.
- Keep max concurrency at 1-2 on free tiers to respect 20 RPM.
Free vs Paid Sibling
| Dimension | :free | Paid slug |
|---|---|---|
| Cost | $0 inference | Per-token credits |
| Rate limits | Platform free caps | No free-tier RPD/RPM caps (upstream still applies) |
| Routing | Free endpoints only | Broader provider pool |
| Privacy | Often weaker | Better chance of ZDR/paid policies |
| Stability | Volatile | Better for SLOs |
FAQs
Are free models truly unlimited if I go slowly?
No. Daily and per-minute caps still apply. Going slowly only avoids RPM bursts.
Does buying $10 in credits make free models unlimited?
No. It typically raises daily free requests (e.g. to 1000) while RPM stays constrained - verify at build. Paid variants remove the free-variant platform request caps.
Why did a free call return 402?
Often insufficient or negative balance state, or key credit limits. Inspect GET /api/v1/key and account credits.
Can I run free models through BYOK?
BYOK is about your provider keys on supported providers. Free variants are a separate OpenRouter endpoint class. Do not assume they combine.
Is `openrouter/free` a real model?
It is a router-style slug that selects free models for you when configured that way. Confirm behavior and limitations in current docs before relying on it.
Do free models support tool calling?
Some do, some do not. Filter the catalog for tools capability and run a one-call probe before building an agent loop.
Will free endpoints train on my prompts?
Many free hosts disclose logging/training. Read the model page and privacy tags. Do not send secrets.
How do I upgrade a prototype to paid?
Change the slug (drop :free or pick a paid model), ensure credits, re-run evals, then raise traffic.
Are free rate limits per key or per account?
OpenRouter documents global governance for free usage; extra keys are not a bypass. Design as account-level caps.
What error should agents expect at the cap?
HTTP 429 with a rate limit exceeded style payload. Back off, alert, and stop the loop cleanly.
Related
- How OpenRouter's Credit and Billing Model Works - credits and 402s
- OpenRouter Fundamentals Basics - first free/paid calls
- OpenRouter vs Calling Provider APIs Directly - when free tier is enough
- Model Slugs and Naming Conventions on OpenRouter -
:freesuffix - Cost-Based Routing: Sending Cheap Tasks to Cheap Models - paid cost discipline
Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana
@solana/web3.js+@solana/spl-token.