Why Reality Router

Enlarge

Only Reality Router earns all four checkmarks. Every other router in the market gives up at least one of: openness, ownership, real per-call routing intelligence, or zero markup — and Reality Router is the only one built on conformal + Venn prediction, the branch of statistics that produces provably calibrated probabilities.

Who this is for

Most developers running AI agents in 2026 aren't using a router yet. If you're one of them, one of these three is probably true:

• You're paying full GPT-5 or Opus 5 prices for tasks a Haiku 4.5 or DeepSeek Flash would nail
• You're watching your Claude Max / ChatGPT Pro fast-request counter burn faster than expected
• You're manually toggling between models mid-session like it's still 2024

A router fixes all three. Same workflow, same tools, same code. Just automatic model selection per call and a real bill you can inspect.

There are about a dozen routers on the market. If you've looked, you've probably ended up at OpenRouter and stopped there — because that's the biggest name and the first thing that shows up. But if you look seriously, every existing router falls into one of three categories that all miss the point. Reality Router is in a category of one — the only router that is open-source, self-hosted, actually routing per call, and built on mathematics proven to be optimal for exactly this problem. The chart above is the map; the rest of this post is the tour.

The three categories, in detail

Category 1: API aggregators (they don't route — they just aggregate)

OpenRouter is what most people mean when they say "LLM router." It isn't one. OpenRouter is a payment layer — one API key that unlocks every model, one monthly bill, one wire protocol. Genuinely useful. If all you want is convenient BYOK across vendors, OpenRouter is exactly right.

But you still pick the model. Every call. OpenRouter never asks whether GPT-5 is overkill for this file summary, or whether Haiku would nail it for 1/20th the cost. It executes what you tell it to execute, adds a 5.5% markup, and moves on. The routing decision — the actual money-saving decision — is still on you.

Aggregators like DeepInfra and AI-Hub-Mix sit in the same category. Different vendors, same idea: unified billing, no routing intelligence.

Category 2: SaaS routing platforms (someone else's black box)

This is where RR's most sincere philosophical competitors live.

NotDiamond was one of the first products to seriously attempt "the router should pick the model, not the human." Their approach is a proprietary classifier that scores incoming prompts and routes based on whichever model their benchmarks say handles that prompt shape best. Solid product. But SaaS. Their classifier, their servers, their reasoning, their pricing.

Sapiom just raised a $50M total round for something more ambitious: a full agent platform (Router + Agent Studio + Runtime) with routing as one piece of the stack. Real routing intelligence, credible results — one customer reportedly cut inference costs 75% after moving. Same architecture concern though: your traffic goes to their servers, their proprietary logic decides, their meter runs. You get a great product and hand over both control and margin.

Martian and Requesty are in the same bucket — commercial routing services with proprietary scoring, sold as SaaS.

The pattern across all four: you outsource the routing intelligence, which means you also outsource the decision you most need to see. If Sapiom decides Sonnet 5 was the right model for a call and you disagree, there's no logic you can inspect, no threshold you can tune, no fork you can make.

Category 3: Self-hostable gateways (proxy without routing)

Portkey has a fine OSS tier — self-hostable, MIT-licensed, feature-rich for observability, retries, caching, rate limiting. Great AI gateway. What it doesn't do is pick a model for you. It routes based on rules you write — if this header, use that model. That's not routing intelligence; it's a proxy with a config file.

LiteLLM is the OSS library that a lot of these products (RR included) use as their wire-protocol adapter. Beautiful primitive: uniform interface across 100+ providers. It is not, and doesn't claim to be, a full router product. LiteLLM handles the "how" of talking to any model; the "which" is left to whoever builds on top of it.

Kong AI Gateway brings the enterprise API-gateway pattern to LLMs — great if you're a large org already running Kong. Same limitation: gateway plumbing without routing brain.

The pattern here: self-hosted and open (great), but no routing intelligence. You're back to writing rules yourself, and the moment your rules are wrong (which they will be, because you can't hand-tune per-call), you're overpaying again.

What Reality Router does that nobody else does

1. The math actually says this is optimal

This one deserves to be said plainly, because it's what separates RR from every other routing platform in the space:

RealitySignal is built on conformal prediction and Venn predictors — the branch of mathematical statistics that produces provably calibrated probabilities under the weakest possible assumptions about your data. Vladimir Vovk (who invented both) proved in the early 2000s that Venn predictors are the only procedure that yields guaranteed well-calibrated multiprobability forecasts online, without requiring strong distributional assumptions.

Translation: given the goal of routing based on "how likely is this model to succeed at this specific call, right now," Venn predictors are the mathematical ceiling. There is no better method — not a smarter classifier, not a larger benchmark corpus, not a bigger training set. Vovk's proof holds regardless.

Every other router in this space — NotDiamond's classifier, Sapiom's benchmark scoring, Martian's heuristics, Requesty's semantic routing — is an approximation of what conformal + Venn prediction gives you natively. They're doing engineering. RR is doing statistics.

For a workload where the difference between the right and wrong model is 20-50× in cost, that mathematical guarantee compounds fast.

2. Full ownership — your box, your keys, your dashboard, your logs

Every SaaS router in categories 1 and 2 sees your prompts. That's a compliance issue for anyone in healthcare, finance, or legal — but it's also just a bad idea if your prompts embed proprietary knowledge, internal architecture, or customer data. RR runs on your box (localhost, home lab, your VPC), holds your provider keys itself, exposes a dashboard nobody else has read access to, and logs every call to a database you own.

The Portkeys and Kong gateways get you halfway here — self-hosted, yes — but they still don't route intelligently. RR is the only option that's both self-hosted and actually routing.

3. Zero markup

OpenRouter takes 5.5%. NotDiamond and Sapiom charge subscription plus per-call fees. Even the "free" tiers are freemium funnels for paid seats. RR takes 0%. You pay the underlying provider (OpenAI, Anthropic, DeepSeek, whoever) their published price and nothing more. No SaaS wrapper on your API bill because there is no SaaS.

Multiply zero markup by the volume of calls a serious agent workload generates, and the difference is real money — often a bigger line item than the routing savings themselves.

RR and Portkey / Kong: complementary, not competitive

If you already run Portkey, Kong AI Gateway, or another OSS gateway for observability, rate limiting, org-level policy, or unified API access — keep them. RR isn't trying to replace them. RR is the routing brain that plugs into your existing gateway stack.

Your App → Portkey/Kong (observability, policy) → RR (routing) → OpenAI/Anthropic/DeepSeek

Both Portkey and Kong let you configure custom OpenAI-compatible upstream providers. Point them at your RR instance and you get the best of both stacks: enterprise-grade gateway features from your existing tools, plus RealitySignal-driven model selection from RR. No fork, no rewrite, no migration.

Where the gateways stop is where RR starts. They proxy, cache, log, and enforce rules you write. They don't pick a model based on real-time calibrated probability. RR does. Together, you're covered end-to-end.

Also in the space, for completeness

Four more competitors named for the record, all falling into one of the three categories above:

• Martian — SaaS routing, benchmark-driven model selection
• Requesty — SaaS routing with semantic classification
• Kong AI Gateway — enterprise API gateway with LLM extensions
• Cursor's built-in router — a black box inside a product you're already paying for

The category we're in, alone

Reality Router isn't trying to be a better OpenRouter, or a cheaper Sapiom, or an OSS NotDiamond. It's a different species: an open-source, MIT-licensed, self-hosted router with real-time calibrated per-call routing intelligence — backed by the only mathematical framework that produces provably calibrated probabilities — and zero markup on top of your model costs.

That's a category with one entry. So far.

→ realityrouter.dev

Reality Router is open source and self-hosted. If your favorite router is missing from this post, open a GitHub issue and we'll place it on the matrix.