You already pay for three AI models
This picks the right one, and shows its work.
Connect the models you already pay for through one OpenAI-compatible endpoint. Agent Mystro reads every unpinned request, identifies the kind of work, and sends it to the model that performs best in blind testing. Those results take precedence over reputation. On a genuine tie, it picks the cheapest one.
A service by Epiphani Studio. Bring your own provider keys, pay your providers directly, and route your first 1 million tokens for free.
POST /v1/chat/completions { "messages": [{ "role": "user", "content": "Draft a note to a new supplier." }] } // no model named, so it routes
- 1Classifieddraft_cold asks for new written copy
- 2Evidencebattery b1 Gemini 5 · chain 3 · Claude 2 · GPT 0
- 3Your taste8 blind picks weight 0.57 yours
- 4Cache checkedno warm prefix elsewhere
Routed to Gemini. Confidence 0.71. The response includes every line above so you can check how it reached that decision.
A service by Epiphani Studio. Bring your own provider keys, pay your providers directly, and route your first 1 million tokens for free.
- 4routing defaults set by blind battery
- 2task types deliberately left unrouted
- 0.1xinput cost kept when a cache is warm
- 0%cut of your token spend
The habit
You are paying for three, and using one.
The best model for email differs from the best model for pages. On pages and interfaces, one model won eight of 10 blind tests. Yet usually, one model gets picked and used for everything.
Teams with live keys from Anthropic, OpenAI, and Google tend to settle into a routine within a week of connecting them. They choose one model for the front-end component, support reply, market research question, invoice arithmetic, and PDF extraction. That choice rarely comes from checking which model is best at all five. Switching by task takes conditional logic, an evaluation harness nobody has time to build, and model strings hardcoded in a dozen places.
Using the same model for everything carries a measurable penalty. A clean interface layout depends on strict structural syntax; cold outreach copy takes control over tone and pacing. Models differ at these jobs. When you send both to the same endpoint, you accept weaker output on every request that falls outside that provider's strengths.
In battery b2, one expert judge scored 10 briefs for pages and interfaces. Outputs were shuffled, stripped of clues to their source, and decoded against an offline key afterward. GPT won eight of 10. Gemini won one, and Claude won one. Battery b1 used the same method on 10 cold-drafting briefs. Gemini won five, the Claude-to-GPT chain won three, Claude won two, and GPT won none.
A model won eight of 10 interface briefs and none of 10 drafting briefs. If you keep it as your default for both, you accept a weaker answer whenever the task changes to drafting. Every default on this page names the run and sample size behind it, so you can judge the evidence yourself. You already pay for the keys and have the capacity to use these models. Choosing which one gets each request puts that capacity to work.
Battery b2, 10 briefs
Pages and interfaces
- GPT8
- Gemini1
- Claude1
Battery b1, 10 briefs
Cold drafting
- Gemini5
- Claude to GPT3
- Claude2
- GPT0
How it routes
How each request gets routed.
Every blind pick you make teaches Agent Mystro your taste and changes how it routes future requests. Over time, your choices determine the routing.
- 01Classify
Identify the kind of work
A deterministic scorer identifies the task: cold drafting, a reply, a page, code, live research. It runs in-process in under a millisecond, costs nothing per request, and returns every signal that fired.
- 02Route
Choose by test results
Each measured task type carries Bradley-Terry strengths built from blind head-to-heads; where no battery exists yet, the prior is weak and the response says so. The request goes to the strongest arm you hold a key for, with the confidence stated rather than implied.
- 03Learn
Learn from your picks
Ask for a taste check to compare two answers with their labels removed. Each pick adjusts the routing by one gradient step. After about six picks, your taste outweighs the population default. Pinning a model teaches it your preference too.
- 04Count the cost
Weigh routing costs
Switching models can discard a warm prompt cache. Before moving a request, we weigh that cost and let the cache decide only when quality is a coin flip. Ask
/v1/gainto see the effect on your spending, including any increase.
x-mystro-request-id: req_5c1e9a07b2d84f63a91e0c4d x-mystro-task-type: draft_cold x-mystro-arm: google:pro x-mystro-model: the id resolved at call time x-mystro-model-source: live x-mystro-personalisation: 0.570 x-mystro-confidence: 0.710
Watch it decide
Choose a request and follow its routing.
Choose a request to follow each step of the decision. This example uses illustrative outputs.
The request
The decision
The evidence
See the test results behind each default.
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.
| Task | Battery | Result | Default |
|---|---|---|---|
| Cold drafting | b1, 10 briefs | Gemini 5 · chain 3 · Claude 2 · GPT 0 | Gemini |
| Replies | b7, 10 briefs | chain 6 · Gemini 4 · solo arms 0 | Claude→GPT chain |
| Pages and interfaces | b2, 10 briefs | GPT 8 · Gemini 1 · Claude 1 | GPT |
| Live research | b8, 10 briefs | Claude 5 · GPT 3 · Gemini 2 | Claude |
| Finance arithmetic | b5, 10 briefs | Every arm 10/10 on accuracy | UNROUTEDgoes to cheapest |
| Document extraction | b10, 6 documents | GPT 30/30 · Gemini 30/30 · Claude 29/30 | UNROUTEDgoes to cheapest |
For two of these tasks, blind testing found no separation between the models. The request goes to the cheapest option, and the response explains that cost decided the routing.
Every default names the run that produced it.
One expert judge, 10 briefs per battery
Six documents for battery b10, outputs shuffled, provenance tells stripped, decoded against an offline key afterwards. Each response carries the sample size behind its prior, so you can weigh it yourself.
Priors refresh when a battery is re-run.
Every response names the run and the date it came from.
Cost figures come from published list prices
Spend caps use published list prices to calculate costs. Check your provider dashboard for the actual bill.
Creative work across models
Two models in concert beat either alone.
Model families come from different labs and datasets, giving each distinct habits and stylistic instincts. You can see those differences when you give them the same drafting brief, naming question, or ideation task. Working across families brings those perspectives into the same piece of work, producing combinations that no single model generates alone.
Across ten reply briefs, a two-model chain with Claude drafting and GPT editing won six of ten. Both solo arms took zero; no single frontier model won the battery. The judging was blind, with outputs shuffled and stripped of clues to their source, then decoded against an offline key.
The same two-model chain won three of ten briefs in the cold-drafting battery, competing against every solo frontier model in the field. Agent Mystro uses the requirements of each task to select and sequence model families. Your creative work draws on their different approaches and measured strengths, with the choice of models changing to suit the prompt.
To compare options directly, POST /v1/council sends two to five options to one seat per model family using your own keys. Each seat votes independently. You get a ranking with every seat's vote and reasons, the strongest dissent, and the cost of each seat, so you can examine the evidence behind the verdict.
The same council runs in the browser. Sign in, name a question and your options at the council playground, and watch the seats vote. A result worth showing publishes at a link of its own, opt-in and revocable, carrying the question, the votes and the verdict. What it cost stays your business.
Which subject line for the spring launch email?
- ASpring is here, and so are we
- BYour spring launch, ready on day one
- CThree things new this spring
- ClaudeB, A, CSays what the reader gets, in the first words.
- GPTB, C, AConcrete, and short enough for a phone inbox.
- GeminiA, B, CStrongest dissent: A carries the season without leaning on the product.
Ranking: B, A, C. Two seats of three put B first, and the response names the dissent beside it.
Requests you make every day
Where six everyday requests get routed.
| Task type | Arm | Evidence |
|---|---|---|
| draft_cold | google:pro | b1 5 of 10 |
| draft_reply | chain:claude_then_gpt | b7 6 of 10 |
| page_gui | openai:flagship | b2 8 of 10 |
| research | anthropic:opus | b8 5 of 10 |
| math_finance | cheapest key held | b5 10 of 10 each |
| any, pinned | the model you name | x-mystro-pin |
Drafting a note to a new supplier
The scorer reads an imperative draft with no quoted thread and names the task cold drafting, in under a millisecond. The request goes to Gemini, because battery b1 put Gemini at five wins of 10, the Claude-to-GPT chain at three, Claude at two, and GPT at none. The response tells you which signals fired and how confident the call was.
Answering a thread already in flight
A quoted message changes the task. The classifier names it a reply, and the request goes to the Claude-to-GPT chain, which took six of 10 briefs in battery b7 against Gemini's four, with both solo arms at zero. Mirroring a sender is a different job from starting a conversation, and the blind picks say so.
Asking for a landing page
Markup, layout, and styling constraints read as a page or interface, and the request goes to GPT on the strength of battery b2, where it took eight of 10 briefs with Gemini and Claude on one each. This is the widest measured gap in the set, and it is the one that costs the most to get wrong by habit.
Finding an answer with current sources
A question that needs today's sources is live research, and it goes to Claude, which took five of 10 briefs in battery b8 against GPT's three and Gemini's two. Narrower than the interface result, and stated as narrow: the confidence figure in the response is lower to match.
Checking the arithmetic on an invoice
Battery b5 put every arm at 10 out of 10 on accuracy, so finance arithmetic is deliberately unrouted and the request goes to the cheapest arm you hold a key for. Where blind testing found no separation, inventing a preference would be dishonest and expensive. Document extraction sits in the same bucket after battery b10: GPT 30 of 30, Gemini 30 of 30, Claude 29 of 30.
Changing a routing choice
Send the
x-mystro-pinheader to use the model you name. That choice also teaches the router your preference, at about a third of the weight of a blind pick. A pin carries less weight because you know the model's identity when you choose it.
How Agent Mystro compares
Compare learned routing with aggregators and gateways.
Three types of product often get called the same thing. An aggregator connects you to hundreds of models. A gateway handles failover, response caching, guardrails, and log collection. A learned router chooses which model gets each request. Agent Mystro provides that routing across three providers.
| Capability | Agent Mystro | Not Diamond | OpenRouter | Portkey | LiteLLM | Cloudflare AI Gateway |
|---|---|---|---|---|---|---|
| Learns continuously from your own blind picks | Yes | Trains per tenant on uploaded datasets | No | No | No | No |
| Returns the routing decision line by line | Yes | No | No | Logs the call, not the decision | Logs the call, not the decision | Logs the call, not the decision |
| Publishes the blind evidence behind each default | Yes, with the tallies | No | Not applicable | Not applicable | Not applicable | Not applicable |
| Share taken of your token spend | 0%; $0.20 per million past 5 million on Pro | $0.05 per million tokens routed | 5% on bring-your-own-key use after the first 1 million requests each month | 0%, subscription | 0%, self-hosted | 5% on credits bought through unified billing |
| Providers reachable | Three | Multiple | Hundreds of models | Broad | Over 100 providers | Broad |
| Log retention and observability depth | Request ids, counts, and enum values | Not published | Dashboard | 100,000 recorded logs a month on Production | Yours, you host it | 10 million logs per gateway on Workers paid |
| Price to start | $0 a month, 1 million tokens routed; Pro $19 a month | Pay as you go, then Enterprise on request | Free to 1 million bring-your-own-key requests a month | Free Developer plan, 10,000 logs a month; Production $49 a month | Free to self-host; Enterprise quoted by sales | Free core features with a Cloudflare account |
The comparison above shows one distinction nobody else matches: routing that keeps learning from your own blind picks. Each pick takes one online gradient step, blended with the population prior using Bayesian shrinkage at kappa 6. At six picks, your taste and the population have equal weight; after that, yours has more. You can inspect each decision's task type, signals, test battery, personal weighting, cache consideration, and confidence. Published tallies support the defaults. You also keep paying providers directly at their prices. We take none of that token spend; routing has its own fee.
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 and keep paying your providers for tokens. We charge for routing. Your providers bill you for inference.
# 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."}]}'
Look in the response's mystro block for the task type, the signals that fired, and the battery behind the default. It also reports how much your taste influenced the choice, whether a cache was weighed, and the routing confidence.
Pricing
Nineteen dollars a month, and no markup on your tokens.
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.
Prove the routing on your own traffic.
- 1 million tokens routed each month
- Population defaults from the blind batteries
- The full explainability payload
- Spend caps and cost estimates
Route by your own blind picks.
- 5 million tokens routed, then $0.20 per million
- Your blind picks teach the taste engine
- Override learning when you pin an arm
- Cache costs weighed and reported when switching
- The gain report at
/v1/gain
Use the free tier's population defaults to test routing on your own traffic. The paid tier learns from your choices and adjusts which model gets each request.
Straight answers
Questions about using Agent Mystro
Where do my prompts go?
Straight to your own provider account, using your own key. The request log records IDs, counts and enum values, while cache affinity records a 12-byte digest of a repeated prefix to track where the cache is warm.
What happens to my API keys?
Encrypted with AES-256-GCM under a per-customer key derived from a secret held in Worker Secrets, and verified against the provider before they are stored, so a typo fails immediately rather than inside your production traffic. Calling DELETE /v1/account removes your keys, picks, and routing state.
What if I disagree with where it routed something?
Pin an arm with the x-mystro-pin header and you get that arm. The disagreement is also recorded and moves the routing toward your choice, at about a third of the weight of a blind pick, because a pin is sighted and a blind pick is not.
Does routing between models break prompt caching?
Before switching models, we weigh the cache. If another model holds a live cache for the repeated part of your request and quality is close to a coin flip, the request stays with that cache. The response explains the choice. When one model wins clearly on quality, it gets the request, and the response tells you the cache was passed over.
How do I know the routing is worth paying for?
Ask /v1/gain to compare your traffic's cost with sending all of it to the strongest general-purpose model. The report includes spending by model and how often your blind picks agreed with the routing. Any extra cost from routing appears there too.
Which providers does it route between?
Anthropic, OpenAI and Google: the three families the blind batteries measure. Every arm you can route to has a battery behind it, and every response names the run.
Privacy
What happens to your prompts, and to your keys.
Your prompts go straight to your own provider account. The request log is the fixed record shown here: ids, counts and enum values defined by its schema. To check whether a repeated prefix is still warm somewhere, cache affinity records a 12-byte digest of that prefix.
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 held in Cloudflare Worker Secrets. 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. Calling DELETE /v1/account removes the keys, along with your picks and routing state.
The rungs (opus, flagship, pro, flash) resolve to concrete model ids at call time by asking each provider what exists, with results cached for 24 hours. When discovery fails, the router uses a date-stamped pin, the response carries a model_drift note saying so, and GET /v1/routing/health shows which answers were live and which were pinned for each provider.
For each account, Agent Mystro keeps your encrypted provider keys, your taste vector, your blind picks, and your spend against the cap, along with the request log and cache digests described above. That is what makes the routing yours.
{
"request_id": "req_5c1e9a07b2d84f63a91e0c4d",
"customer_id": "4f2a91c07b3e",
"ts": "2026-11-14T15:02:11Z",
"task_type": "draft_cold",
"arm": "google:pro",
"model": "the id resolved at call time",
"model_source": "live",
"streamed": false,
"status": 200,
"latency_ms": 1840,
"classifier_ms": 0,
"classifier_confidence": 1,
"routing_confidence": 0.71,
"personalisation": 0.57,
"prompt_tokens": 18,
"completion_tokens": 164,
"cost_micros": 1663
}
error_kind.Stop guessing which model to use.
Use your own keys through one endpoint and check each routing decision line by line.