Re-verified 2026-09-20: the genuinely free, genuinely maintained AI model routers are now LiteLLM (self-hosted, MIT outside its enterprise/ directory, 59k ★ and commits landing this week) and Portkey's free tier (managed, with guardrails). Of the seven routers this page originally listed, two no longer sell a router at all and a third has not shipped a commit since August 2024 — the status table below shows how each was checked. ClawRouters publishes this page, so we have taken ourselves out of the ranking rather than rank ourselves first — our own free BYOK tier is described in the disclosure below, where you can weigh it knowing who wrote it.
If you're building AI-powered applications, you already know that model costs add up fast. A single Claude Opus 4 response can cost $0.15+, while the same task handled by Gemini 3 Flash might cost $0.001. The difference is 150x — and an AI model router helps you capture those savings automatically.
But do you need to pay for a router to save money? Not necessarily. In 2026, several excellent free options exist. This guide compares every free AI router worth considering, with honest assessments of what "free" actually means for each one.
Status check, re-verified 2026-09-20
This list was first published in March 2026. We re-checked every entry against its live site and repository on 2026-09-20, because "best free X" lists rot faster than anything else in this category — and four of the seven had changed materially.
| Router | Status on 2026-09-20 | Evidence |
|---|---|---|
| LiteLLM | ✅ Active, and the strongest free option | 59,272 ★, last commit 2026-09-21. License is MIT except the enterprise/ directory, which is separately licensed — not plain MIT as this page previously said |
| Portkey | ✅ Free tier still offered | Pricing page still advertises a "Free Forever" plan |
| RouteLLM | ⚠️ Effectively unmaintained | 5,521 ★, but last commit 2024-08-10 — over two years stale. Still fine to read as research; do not build on it expecting fixes |
| Martian | ❌ No longer a public router product | /pricing and /products/router both return 404; the homepage no longer contains the word "routing" anywhere. The company now presents itself as "Martian: Understanding Intelligence" (interpretability research) |
| Unify AI | ❌ No longer a public router product | /pricing and /chat return 404, the docs domain no longer resolves, and the homepage now reads "a research lab working on continual learning" |
| DIY with the OpenAI SDK | ✅ Still the honest baseline | Unchanged — it is code you write, so it cannot rot |
What that means if you are choosing today: the genuinely free, genuinely maintained options have narrowed to LiteLLM (self-hosted, most active project in the category by a wide margin) and Portkey's free tier (managed, with guardrails). Two of the seven original picks no longer sell a router at all, and a third has not shipped a commit since 2024. The detailed write-ups below are kept for reference, but read them against this table rather than instead of it.
Disclosure
ClawRouters publishes this page and is itself a router, so treat anything we say about ourselves as an advertisement rather than a finding — that is the honest framing, and it is why we removed ClawRouters from the ranked positions above rather than ranking ourselves first.
For the record, what we offer: a free BYOK tier — you add your own provider API keys (OpenAI, Anthropic, Google, DeepSeek and others) under Model Providers in the dashboard, we route across them, and we charge nothing for the routing layer; you pay each provider directly. Paid plans use our keys instead and can fall back to yours when their quota runs out. Plans are on the pricing page. Whether that beats self-hosting LiteLLM depends on whether you want to run infrastructure; we would rather you decide that with the vendor bias labelled than have us decide it for you at position #1.
How to re-check this yourself
None of the above needed special access, and it takes about ten minutes:
- Is the project alive?
https://api.github.com/repos/<owner>/<repo>returnspushed_atandarchived— a repo whose last push predates the article you are reading is the single fastest disqualifier. - Is the product still sold? Request the vendor's
/pricingdirectly. A 404 on a pricing page is a louder signal than any blog post. - Does the homepage still describe the thing? Fetch it and grep for the category word. Martian's homepage containing zero instances of "routing" settled that question in one command.
- Read the actual LICENSE file, not the GitHub sidebar badge. GitHub reported
NOASSERTIONfor LiteLLM; the file itself explains why — a carve-out forenterprise/.
What Makes an AI Router "Free"?
Before diving into options, let's clarify what "free" means in this context. There are three models:
- BYOK (Bring Your Own Keys) — You provide your own API keys from OpenAI, Anthropic, Google, etc. The router charges nothing for the routing layer itself. You only pay what you'd pay the providers directly.
- Open-source self-hosted — You run the router on your own infrastructure. The software is free, but you pay for compute (a small VPS typically costs $5-20/month).
- Free tier with limits — A managed service offers a free plan with usage caps. You get a taste, but will likely need to upgrade for production use.
The most genuinely "free" options are BYOK and open-source. Let's compare them all.
The free AI model routers, one by one
Six entries, renumbered after removing our own — see the disclosure above. Read each against the status table, which is newer than the write-up.
1. LiteLLM — Best Self-Hosted Open-Source Router
Cost: Free (open-source, MIT license) Type: Self-hosted proxy Routing: Rule-based, manual configuration
LiteLLM is the most popular open-source LLM proxy, providing a unified OpenAI-compatible interface across 100+ model providers. It's a Python-based proxy server that translates between different provider APIs.
What you get for free:
- OpenAI-compatible API for 100+ providers
- Provider fallbacks (try model A, fall back to model B)
- Rate limiting and spend tracking
- Load balancing across multiple API keys
- Virtual keys for team management
- Detailed logging and callbacks
Limitations:
- No intelligent routing — you must manually specify which model to use or set up static rules
- Requires self-hosting (Docker, VPS, or Kubernetes)
- Configuration is YAML-based and can get complex
- No built-in task classification or cost optimization logic
- Maintenance overhead (updates, monitoring, scaling)
Why it stands out: LiteLLM is battle-tested with a large community. If you need a unified API layer and are comfortable managing infrastructure, it's a rock-solid choice. See our detailed comparison of LiteLLM vs ClawRouters vs OpenRouter.
Best for: Engineering teams with DevOps capacity who need a self-hosted proxy with maximum customization.
2. Martian Model Router — ❌ no longer a public router product (checked 2026-09-20)
Cost: Free tier available Type: Managed, API-based Routing: ML-based adaptive routing
Martian's router uses machine learning to dynamically select models based on prompt characteristics. Their free tier provides access to adaptive routing with reasonable usage limits.
What you get for free:
- ML-powered model selection
- Automatic quality-cost optimization
- API access with usage caps
- Basic analytics
Limitations:
- Usage caps on the free tier
- Smaller model selection than ClawRouters or LiteLLM
- Less control over routing decisions
- Limited provider support
Best for: Developers who want to experiment with ML-based routing without commitment.
3. RouteLLM — ⚠️ unmaintained since 2024-08, still worth reading
Cost: Free (open-source, Apache 2.0) Type: Self-hosted framework Routing: ML classifiers (BERT, matrix factorization, causal LLM)
RouteLLM, developed by researchers at UC Berkeley and Anyscale, is a framework for training and serving LLM routers. It focuses specifically on the routing decision — choosing between a strong model and a weak model based on query characteristics.
What you get for free:
- Pre-trained routing models (BERT classifier, matrix factorization, causal LLM)
- Framework for training custom routers on your own data
- Integration with major providers via OpenAI-compatible API
- Research-grade evaluation tools
Limitations:
- Designed for two-model routing (strong vs. weak) — not multi-model
- Research-oriented, not production-hardened
- Requires ML expertise to train custom routers
- No managed hosting option
- Limited provider integrations out of the box
Best for: ML researchers and teams who want to train custom routing models on their specific data distribution.
4. Portkey AI Gateway — Free Tier with Guardrails
Cost: Free tier (10K requests/month) Type: Managed gateway Routing: Rule-based with conditional logic
Portkey offers a generous free tier of their AI gateway platform, which includes routing, guardrails, caching, and observability features. See our ClawRouters vs Portkey comparison for more details.
What you get for free:
- 10K requests/month
- Conditional routing (if/else logic based on metadata)
- Automatic retries and fallbacks
- Request caching
- Basic observability
Limitations:
- 10K request cap (roughly 333 requests/day)
- No AI-powered smart routing on free tier
- Enterprise features locked behind paid plans
- Primarily focused on enterprise compliance use cases
Best for: Teams evaluating AI gateways who need guardrails and compliance features alongside basic routing.
5. Unify AI — ❌ no longer a public router product (checked 2026-09-20)
Cost: Free credits on signup Type: Managed router Routing: Benchmark-based optimization
Unify provides a routing layer that selects models based on benchmark performance data. They offer free credits for new users to test their routing capabilities.
What you get for free:
- Signup credits for testing
- Benchmark-driven model selection
- Unified API across providers
Limitations:
- Free credits are limited and don't last for production use
- Routing is benchmark-based, not real-time adaptive
- Smaller community than LiteLLM or ClawRouters
Best for: Developers who want to quickly test benchmark-optimized routing.
6. DIY with OpenAI SDK + Custom Logic
Cost: Free (your code, your infrastructure) Type: Self-built Routing: Whatever you implement
For developers who want maximum control, building a simple router with the OpenAI SDK is straightforward. You can implement basic routing logic (keyword matching, message length thresholds, regex patterns) in under 100 lines of code.
What you get for free:
- Complete control over routing logic
- No dependencies on third-party services
- Zero overhead beyond your own code
Limitations:
- You build and maintain everything yourself
- No pre-built task classification
- No analytics, dashboards, or monitoring
- Scaling, failover, and load balancing are your problem
- Typically results in basic rule-based routing, not intelligent classification
Best for: Developers with very specific routing needs who enjoy building infrastructure.
Free AI Router Comparison Table
| Router | Status 2026-09-20 | Type | Smart Routing | Models | Setup Time | Maintenance |
|---|---|---|---|---|---|---|
| LiteLLM | ✅ active | Self-hosted | Manual rules | 100+ | 30-60 min | Medium |
| Portkey Free | ✅ free tier live | Managed | Rule-based | 30+ | 10 minutes | None |
| DIY | ✅ cannot rot | Self-built | Custom | Any | Hours-days | High |
| RouteLLM | ⚠️ no commits since 2024-08 | Self-hosted | ML classifiers | 2 (strong/weak) | 1-2 hours | High |
| Martian | ❌ router product gone | — | — | — | — | — |
| Unify | ❌ router product gone | — | — | — | — | — |
ClawRouters is not listed: we publish this page. See the disclosure above.
How to Choose the Right Free AI Router
Choose LiteLLM if:
- You need to self-host for compliance or data sovereignty
- You want maximum customization of the proxy layer
- Your team has DevOps capacity to maintain infrastructure
- You need support for 100+ model providers
Choose RouteLLM if:
- You're doing research on LLM routing
- You want to train custom routing models
- You have ML expertise on your team
Choose Portkey Free if:
- You need guardrails and compliance features
- 10K requests/month is sufficient for your workload
- Enterprise compliance is a primary concern
If you want to try ClawRouters
The free tier is BYOK: sign up, add at least one provider key under Model Providers in the dashboard
(the Free plan cannot make calls until one is configured), then point any OpenAI-compatible client at the
gateway and set model to auto. Paid plans skip the key step and use ours. Current plans are on the
pricing page.
from openai import OpenAI
client = OpenAI(
base_url="https://www.clawrouters.com/api/v1",
api_key="cr_your_key_here",
)
response = client.chat.completions.create(
model="auto", # ClawRouters picks the model per request
messages=[{"role": "user", "content": "Explain quantum computing"}],
)
One behaviour worth knowing before you build on it, because it surprised us while testing for this update:
on plans where smart routing is the product, an explicit model name is rewritten to automatic routing.
Asking for gpt-5-mini came back served by gemini-3-flash. The response's model field always reports
what actually served the request, so check it rather than assuming you got what you asked for. If you need
a guaranteed specific model, that is a reason to use a different tool — LiteLLM, for instance.
Frequently Asked Questions
Which free router should I actually pick in 2026?
LiteLLM if you can host it, Portkey's free tier if you cannot. As of our 2026-09-20 re-check those are the two entries on this page that are both free and actively maintained. LiteLLM is the more capable and by far the most active project in the category; Portkey trades self-hosting for a managed free tier with guardrails. Avoid starting new work on RouteLLM — it is good reading, but it has not received a commit since August 2024.
Why doesn't ClawRouters rank itself on this page?
Because we publish it, and a vendor ranking itself first in its own "best free" list is an advertisement wearing the clothes of a comparison. We do have a free BYOK tier, and it is described in the disclosure above so you can weigh it knowing who wrote the page. If you want a free router you run yourself, LiteLLM is the strongest option on this list and saying so costs us less than being the kind of list nobody should trust.
Can a free AI router actually save me money?
Absolutely. The savings come from smarter model selection, not from the router itself being cheap. If you're currently sending every request to GPT-4o or Claude Opus, a router can identify that 60-80% of those requests could be handled by cheaper models (Gemini Flash, GPT-4o Mini, DeepSeek) at equal quality. Read our complete guide to reducing LLM API costs.
What's the difference between an AI router and an AI gateway?
An AI router focuses on intelligent model selection — picking the right model for each request. An AI gateway is broader infrastructure that includes authentication, rate limiting, caching, logging, and compliance features. Many products (including ClawRouters) combine both. See our detailed comparison of routers vs gateways.
Should I self-host or use a managed free router?
It depends on your constraints. Self-hosting (LiteLLM, RouteLLM) gives you maximum control and data sovereignty but requires infrastructure maintenance. Managed free tiers (Portkey) eliminate operational overhead but mean your requests pass through a third party. Given how the list above aged, self-hosting also has a durability argument: LiteLLM cannot decide to stop being a router, whereas two managed options on this page did exactly that. Read our self-hosted vs managed comparison.
How much can I save with a free AI router?
Typical savings range from 60-90% depending on your workload. A developer spending $500/month on Claude Opus for everything could drop to $50-200/month by routing simple tasks to cheaper models — without any loss in output quality for those tasks. The savings scale with volume: the more requests you make, the more a router helps. Check our AI token cost breakdown for 2026.
Can I use a free AI router with Cursor or Windsurf?
Yes. ClawRouters works as a drop-in replacement for any OpenAI-compatible tool. Just change the base URL and API key in Cursor, Windsurf, or other AI coding agents. Since these tools make hundreds of API calls per session, smart routing can cut your coding assistant costs dramatically — see our guide to reducing Cursor and Windsurf costs.
Going deeper: which workloads to offload to cheaper open-source models, how routing + caching interact with pricing and latency, the LLM pricing tracker behind our numbers, and the full model-vs-model comparison index if you are weighing two specific models.
Also useful when things break: the LLM gateway & integration error reference — exact error strings, causes and fixes for the integration layer.