ARBR Review (2026): Pricing, Features & Honest Verdict

ARBR is an open-source LLM gateway and control plane: one OpenAI-compatible endpoint that routes, meters, governs and logs every model call across every provider you use. Best for platform teams running more than one model vendor. Price: $0, MIT licensed, self-hosted. Free plan: yes. Rating: 7.0/10.

What is ARBR?

ARBR sits between your applications and your model providers. You point any OpenAI-compatible client at it, change the base URL, and every request from that moment gets routing, budget enforcement, guardrails, authentication and structured logging without a line of application code changing. It ships as one Docker container on one port, with the gateway at /v1/*, the admin API at /api/* and the dashboard at /, backed by MongoDB.

The project is MIT licensed. We fetched the raw LICENSE file rather than trusting the badge, because vendors have been caught adding restrictive clauses under an open-source headline. ARBR’s is the unmodified MIT text with no supplementary terms, no anti-competition clause and no ban on hosted operation. You can run it, fork it and sell a service on top of it.

It comes from Prasanna Vaidya and Shubham Deshmukh, who build it under the Gyde name, with Vaibhav Domkundwar of the Domkundwar Foundation credited on the project. The GitHub repository project-arbr/arbr-control-plane was created on 17 June 2026 and was still receiving commits on the day we looked. That is a young project, and the numbers below say how young.

What Are ARBR’s Key Features?

One endpoint over 100 providers

Eight providers have native adapters written for them: OpenAI, Anthropic, Google Gemini, Amazon Bedrock, DeepSeek, Moonshot, xAI and Groq. Around thirty more, including Mistral, Cohere, Together, Fireworks, Perplexity, Cerebras, OpenRouter, Azure AI, Databricks and Vertex AI, are preconfigured and need only a key. Anything else that speaks the OpenAI shape works by pointing ARBR at its base URL. The endpoint covers streaming chat, embeddings, tool calling and a realtime voice proxy that relays a WebSocket session to OpenAI while still counting tokens.

Policy that runs before the model call, not after it

This is the part that separates ARBR from a logging tool. Budgets, rate limits, output guardrails, prompt-injection checks and a global kill switch all execute in the request path. A budget that stops spend is worth more than a dashboard that reports it a day later, and a one-toggle kill switch that halts all AI traffic with a custom message is exactly what you want during an incident. Authentication supports three modes: an admin key, OIDC, or a trusted header, so per-user attribution is available if you wire it up.

Evals that choose the cheapest model that still passes

ARBR can score a sample of live traffic with an LLM judge and then route to the cheapest model proven to clear the bar. It builds test datasets out of real traffic rather than asking you to write them. Paired with the routing rules and difficulty-aware selection, this is the mechanism behind the cost claims on the homepage. The hero dashboard shows 1.82M requests, $6,281 of spend and $4,392 of “realised savings,” but that is synthetic demo data, not a customer case study. Treat it as a screenshot, not evidence.

Deployment docs you can actually follow

The deployment page is the most convincing thing on the site. It names the port (4100), the state store (MongoDB) and the exact behaviour of the production compose overlay: with NODE_ENV=production it refuses to start without ARBR_ADMIN_KEY and ARBR_ENCRYPTION_KEY, disables demo seeding, binds to loopback so only your reverse proxy reaches it, and forces gateway API-key auth on. It gives the nginx block for SSE streaming with buffering off and a 300 second read timeout, and warns you to raise the ALB idle timeout to match. There is a step-by-step GCP walkthrough on a single e2-medium with certbot, and a gated deploy script that health-checks and rolls back on failure. Software that tells you how it fails is usually written by people who have watched it fail.

Does ARBR Ship Everything Its Homepage Claims?

Not quite, and the site disagrees with itself about it. The homepage headline reads “One control plane, six jobs” and lists Route, Govern, Observe, Evaluate, Deploy and Capture. Two of those six carry qualifiers in small print. Capture, which turns logged traffic into fine-tuning datasets, says “Not shipped yet.” Deploy, which covers canary rollouts and regression gates, is tagged “Roadmap.”

Now open the feature catalog at /features/. It opens with “Below is every feature in the codebase” and a counter reading 115 features. Under the heading “Rolling out & rolling back” it lists gradual canary rollout, automatic rollback on error rate or latency breach, one-click promote to 100%, manual rollback and an exportable evidence report. Five specific capabilities, described in the present tense, on a page that claims to enumerate shipped code, for the same capability the homepage calls a roadmap item.

One of those two pages is wrong. We cannot tell which from the outside, which is the problem. If you are evaluating ARBR for the canary story specifically, ask before you deploy.

A smaller point in the same spirit: the headline “3,000+ models priced” is real but borrowed. The docs say plainly that ARBR pulls the full public LiteLLM catalog to keep model pricing and capability flags current. That is a sensible engineering choice and ARBR discloses it. It also means the pricing intelligence in your control plane is maintained by a competitor.

How Much Does ARBR Cost?

Nothing. ARBR is free, MIT licensed and self-hosted, and there is no licence tier, no seat count and no request cap. Your cost is a VM, a MongoDB instance and the hour it takes to put TLS in front of it.

A hosted option exists but is not for sale yet in any public sense. The homepage links “Don’t want to host it?” to app.projectarbr.com, which returns an HTTP 401 and a JSON authentication error. There is no pricing page, no waitlist form and no published rate anywhere on the site. If you need someone else to run this, you have no number to plan against.

Set that against the paid competition and the case makes itself. Helicone charges $79 a month for Pro and $799 a month for Team, plus usage on top of a 10,000 request free allowance, for observability that overlaps heavily with what ARBR logs. That is $9,588 a year at the Team tier against $0 in licence fees. Portkey’s Production plan is $49 a month for 100,000 recorded logs with $9 per additional 100,000, which works out at $0.00009 per log. LiteLLM, the closest philosophical match, is also free to self-host forever with an Enterprise tier priced by quote.

Who is ARBR Best For?

Use ARBR if you already run more than one model provider and have lost track of what each application is spending. Its combination of pre-flight budgets, per-application attribution and a kill switch answers a question most teams currently answer with a spreadsheet and a provider invoice. Use it if you need self-hosting for regulatory reasons, because the MIT licence puts no conditions on that at all. Use it if you have someone who is comfortable running a container behind nginx and keeping MongoDB alive.

Skip ARBR if you want a managed service today, because the hosted tier is a 401 with no price. Skip it if your organisation cannot tolerate a project that reached v0.3.0 three months ago. And think carefully if you cannot accept the architecture: every request from every application transits one instance backed by one MongoDB, the docs describe one standalone instance per organisation, and there is no clustering or high-availability guidance published. That is a single point of failure sitting in front of everything your product does with AI.

Best ARBR Alternatives

LiteLLM is the incumbent and the obvious comparison, not least because ARBR consumes LiteLLM’s model catalog. The open-source gateway is free to self-host forever with no licence fee, and Enterprise adds governance, security and support at a quoted price sized to your annual request capacity rather than per token. There is a 30-day Enterprise trial key emailed on request with no sales call. LiteLLM is far more widely deployed. ARBR’s answer is a tighter opinion about governance and a dashboard that comes in the box.

Portkey is the managed version of this idea. Free Forever covers 10,000 recorded logs a month with requests unaffected beyond the cap, Production is $49 a month for 100,000 logs with $9 per additional 100,000, and Enterprise starts above 10 million logs a month. Pick Portkey if you want someone else to run the gateway and can live with a per-log meter.

Helicone is the observability-first option. Hobby is free for 10,000 requests a month with 1 GB of storage and a single seat, Pro is $79 a month with unlimited seats, alerts and its query language, and Team is $799 a month with five organisations plus SOC 2 and HIPAA compliance. Both paid tiers add usage-based charges. Helicone is stronger on analysis and weaker on enforcement. ARBR blocks; Helicone reports.

Final Verdict: Is ARBR Worth It?

ARBR is a serious piece of engineering hiding behind a very small GitHub badge. The licence is clean MIT with no traps, the governance model runs before inference instead of after it, and the deployment documentation is more specific than most funded companies publish. For a team paying Helicone $799 a month or wiring budgets together by hand, spinning up one container to see whether this replaces both is an afternoon well spent.

The numbers do demand caution. Twenty-four stars, four forks, twenty open issues, six contributors of whom two account for nearly all the commits, three tagged releases and a v0.3.0 version number describe a project that is ten weeks old. Add a homepage and a feature catalog that disagree about whether canary rollouts exist, and a hosted tier that answers with a 401. Run it in a test environment, mirror real traffic through it, and decide from your own logs. Do not put it in front of production revenue this month.