This picks the right one, and shows its work.

The habit

You are paying for three, and using one.

Teams holding live keys from Anthropic, OpenAI, and Google tend to settle into one routine within a week of wiring them up. One model gets chosen, and then everything goes to it: the front-end component, the support reply, the market research question, the invoice arithmetic, the PDF extraction. Not because anyone decided that model was best at all five. Because switching by task means conditional logic, an evaluation harness nobody has time to build, and model strings hardcoded in a dozen places.

That default carries a measurable penalty. A clean interface layout rewards strict structural syntax. Cold outreach copy rewards control over tone and pacing. These are different skills, and the models are not equally good at them. Send both to the same endpoint and the output drops wherever that provider happens to be weak, quietly, on every request, forever.

Blind testing shows the size of it. In battery b2, one expert judge scored 10 briefs for pages and interfaces with the outputs shuffled, stripped of provenance tells, and decoded against an offline key afterwards. GPT won eight of 10. Gemini won one, and Claude won one. Battery b1 ran the same method over 10 cold-drafting briefs: Gemini won five, the Claude-to-GPT chain won three, Claude won two, and GPT won none at all.

One judge over 10 briefs per battery is not a population study, and the limits section says so plainly. But a model that takes eight of 10 interface briefs and none of 10 drafting briefs is not a rounding error, and a static default means accepting the weaker answer every time the task moves. The keys are already paid for. The capacity is already there. What is missing is the decision about which one gets the work.

How it routes

Four moves, on every call.

Watch it decide

Pick a request. See where it goes, and why.

These are the real classifier signals, the real batteries, and the real routing rules. The outputs are illustrative, not live model calls.

The evidence

The defaults are measured, and here they are.

Every routing default came from a blind battery, outputs shuffled and stripped of provenance tells, judged blind, then decoded against an offline key: 10 briefs each, except battery b10's six documents. The machine-readable version lives at /capabilities.

Two of these say do nothing, and that is the point. Where blind testing found no separation we route on cost and tell you so, rather than inventing a preference we cannot defend.

Everyday moments

Six requests, and where each one lands.

How Agent Mystro compares

A learned router, next to the aggregators and the gateways.

Three different products get called the same thing. An aggregator sells catalogue breadth: hundreds of models behind one connection. A gateway sells operations: failover, response caching, guardrails, and log collection. A learned router sells the decision itself, which model gets which work. Agent Mystro is the third one, and it reaches three providers rather than hundreds.

So the honest routing advice comes first. If you need niche open weights across a broad catalogue, OpenRouter is the better tool. If your prompts may not leave your own infrastructure, LiteLLM self-hosts and runs air-gapped and Agent Mystro does not, because the population prior is learned across customers and an isolated instance would ship frozen. If you need automatic failover, response caching, guardrails, or redaction today, Portkey and LiteLLM ship all four and Cloudflare ships caching and rate limiting, and we ship none of them yet.

What is left is the part the comparison above shows nobody else matching. The routing keeps learning from your own blind picks, one online gradient step each, blended against the population prior by Bayesian shrinkage at kappa 6, so your taste and the population weigh equally at six picks and yours dominates after. Every decision comes back line by line: the task type, the signals that fired, the battery behind the default, how much of the call was your own vector, whether a warm cache was weighed, and the confidence. The defaults are published with the tallies that produced them. And the routing fee is not a share of your tokens: you keep paying the providers directly, at their prices, and we take none of it.

Every figure above was read from the vendor's own pricing or documentation page in July 2026 and links from the sentences below. Not Diamond publishes $0.05 per million tokens routed on pay as you go, with an Enterprise plan it does not price publicly. OpenRouter takes no markup on inference and charges 5.5% with a $0.80 minimum on credit purchases, plus 5% on bring-your-own-key use past the first 1 million requests a month. Portkey lists a free self-hosted open-source tier, a free Developer plan at 10,000 recorded logs a month, and Production at $49 a month for 100,000 logs plus $9 for each further 100,000. LiteLLM is free to self-host, reaches over 100 providers, runs air-gapped, and quotes Enterprise privately. Cloudflare AI Gateway gives its core features away, charges 5% on credits bought through its Unified Billing, and caps persistent logs at 100,000 on the free Workers plan and 10 million per gateway on paid. Martian, which built one of the first model routers on $9 million of funding, now describes itself on its own homepage as researchers working on understanding machine intelligence, and publishes neither a router product nor a price, so it is named here rather than scored. Capabilities change; corrections are welcome at hello@epiphani.studio.

Quickstart

Point any OpenAI client at it.

Bring your own provider keys. We charge for routing and never resell inference, so your token spend stays exactly where it already is.

# 1. Create an account. The key is shown once.
curl -X POST https://agentmystro.com/v1/account

# 2. Add a provider key. We verify it before storing it, encrypted.
curl -X POST https://agentmystro.com/v1/keys \
  -H "Authorization: Bearer $MYSTRO_KEY" \
  -d '{"provider":"anthropic","api_key":"sk-ant-..."}'

# 3. Call it like OpenAI. Leave the model out and let it route.
curl https://agentmystro.com/v1/chat/completions \
  -H "Authorization: Bearer $MYSTRO_KEY" \
  -d '{"messages":[{"role":"user","content":"Draft a note to a new supplier."}]}'

The response carries a mystro block: the task type it decided on, the signals that fired, the battery behind the default, how much of the call was your own taste, whether a cache was weighed, and how confident it actually was.

Pricing

Nineteen dollars a month, and we add nothing to your token bill.

You pay Anthropic, OpenAI, and Google directly, at their prices. We take no percentage of it. Routing can still change what you pay them, and /v1/gain reports it either way.

The free tier runs on population defaults only. That is the honest split: it proves the routing works on your traffic, and the part you pay for is the routing becoming yours.

Straight answers

The questions worth asking first.

Privacy

What happens to your prompts, and to your keys.

Your prompts pass through to your own provider account and are not retained. The request log holds identifiers, counts, and enum values, and there is no field in the log record that can carry your text, which is a property of the schema rather than a policy anyone has to remember. Cache affinity needs to know whether a repeated prefix is still warm somewhere, so it stores a 12-byte digest of that prefix and never the prefix itself.

Your provider keys are encrypted with Advanced Encryption Standard 256-bit Galois/Counter Mode ("AES-256-GCM"), under a key derived for your account alone by an HMAC (Hash-based Message Authentication Code)-based Extract-and-Expand Key Derivation Function ("HKDF") from a master secret that lives in Cloudflare Worker Secrets and never in the repository. A key is verified against the provider before it is stored, so a typo fails in the moment you paste it rather than inside your production traffic three days later. Delete the account and the keys go with it.

That design has a cost, and it is worth stating rather than discovering. Rotating the master secret makes every stored provider key undecryptable. That is deliberate, and it means a rotation asks every customer to add their keys again. We would rather tell you that now than have it arrive as a surprise on a Tuesday.

Model versions are never stored. The rungs (opus, flagship, pro, flash) resolve to concrete model ids at call time by asking each provider what exists, cached for 24 hours, and they degrade to a date-stamped pin only when that discovery fails. When it does, the response carries a model_drift note saying so, and GET /v1/routing/health shows per provider which answers were live and which were pinned. A router that quietly called a retired model would be worse than one that says it had to guess.

Agent Mystro keeps no copy of your prompts. Per-account, it keeps your encrypted provider keys, your taste vector, your blind picks, and your spend against the cap, alongside the request-log identifiers and cache digest described above. That is what makes the routing yours rather than ours, and together, it is the whole of what is held.

What this does not do yet

The limits, in the same breath as the claims.

• One judge, 10 briefs per battery, six documents for battery b10. That beats an unmeasured guess and it is not a population study.
• Code and general reasoning have no dedicated battery. Their priors are weak, and every response says so.
• Priors refresh when a battery is re-run, not continuously. Models ship monthly, so a prior can age between runs.
• No automatic provider failover, response caching, guardrails, or personally identifiable information (PII) redaction yet. Portkey and LiteLLM ship all four today, and Cloudflare ships caching and rate limiting.
• No self-hosted build, and that is a design consequence: the population prior is learned across customers, so an isolated instance would ship frozen. If prompts may not leave your infrastructure, LiteLLM is the honest answer.
• No System and Organization Controls 2 (SOC 2) report, no European Union (EU) data residency guarantee, no third-party security audit.
• Cost figures are estimates from published list prices, used to enforce your cap. Your provider dashboard remains the billing truth.