The best model for every call.
Multi-vendor LLM routing by cost, tools, inference, and adversarial-challenge.
XSI-AIMS Compass is the model router for the XSI-AIMS™ standard. Register your provider keys once; declare per-call requirements; Compass picks the cheapest model that meets each call's requirements.
Model choice shouldn't be hardcoded.
Cost and capability profiles vary by call. A code-completion call has different requirements than a contract-review call. Agents should declare what each call needs, not which model handles it. XSI-AIMS Compass is the router that makes that possible.
A hardcoded flagship pays the top of the spread.
A single agent is a rounding error. A fleet is a budget line. Every workflow that works recruits more, and token spend compounds with them — at whatever rate the code hardcoded back when the fleet was a prototype.
The price axis is wide. As of June 11, 2026, published input rates across the current model lineups of the three major vendors run from about $0.10 to $10 per million tokens — a 100× spread. Output rates spread wider, at 125×. The dominant 2026 pattern is one flagship model for every call, so the 40-token classification ping and the 150,000-token contract review ride the same top-of-spread rate. A classification ping does not need a 1M-token context window. An extraction call with a JSON schema does not need frontier reasoning.
The problem is not that frontier models are expensive. It is that fleets buy frontier capability for calls that never declared a need for it — because nothing in the stack asks the call what it needs.
Three steps. One router.
Provider keys.
Register your API keys with each vendor — Anthropic, OpenAI, Google, Mistral, locally-served, anything that speaks a known protocol.
Cost and capability.
Compass researches each registered model's cost-per-token, context window, tool-call support, JSON-mode, and other capability dimensions.
Cheapest that fits.
At call time, Compass picks the cheapest registered model that meets the call's declared requirements.
Least-cost routing, reborn.
Telecom carriers solved this problem class in the 1980s. Least-cost routing is the discipline of selecting, per call, the cheapest termination route that clears the call's quality bar — and carriers ran it for four decades, against per-minute spreads far narrower than the 100× the model market shows today. The board above is that discipline, applied to model calls. A route inventory becomes the registered models. A rate deck becomes the published per-token list prices. A quality floor becomes the call's declared requirements. Per-call selection becomes the router picking the cheapest registered model that meets the call's declared needs.
Per-call selection beats static assignment whenever three conditions hold — the price spread is wide, the traffic is heterogeneous, and prices move. Telecom met all three with narrow spreads. Model calls meet all three with a 100× spread, traffic ranging from 40-token pings to 150,000-token reviews, and a vendor field that shipped new models and new rates throughout the first half of 2026. Compass is the per-call switch for that field.
The spread is the margin.
Routing harvests the gap between what a call needs and what a hardcoded flagship charges. One scenario makes the structure visible — and the structure survives any reasonable change to the assumptions.
A midsize fleet runs ten million calls a month on one frontier model — about $225,000 a month, all-frontier. Route the 70% of calls whose declared requirements a mid-tier model already clears to that mid-tier model, and the same workload — same prompts — bills near $99,000.
56% off, from routing alone.
Before any caching, batch, or volume-discount lever — those stack on top. The whitepaper works the full scenarios and a sensitivity table.
Per-call requirements.
Compass routes by what your call needs, declared per call:
- Max cost per call (in cents or per-token)
- Required tool / function call support
- Inference time budget (max latency)
- JSON-mode capability
- Minimum context window
- Required modality (text / vision / audio)
- Adversarial-challenge round (for safety-critical calls)
- Vendor exclusions or pinned-model overrides
Two opinions for the important calls.
For high-stakes outputs, Compass routes the call through a primary model and a different-vendor challenge model in parallel. Disagreement surfaces to the agent before it commits. Maps directly to XSI-AIMS's blind-spot monitoring principle from the spec.
Use it when the cost of a wrong answer exceeds the cost of two inferences.
When a wrong answer costs more than two inferences.
Cost routing has an obvious failure mode — race to the cheapest model, accumulate quiet failures, lose more in rework than the routing saved. The requirements floor is the first guard: a call never routes below its declared needs, so the savings come from calls that were over-served, not from calls that get under-served. The adversarial-challenge round is the second, and it spends more on purpose. The rule prices itself — declare a challenge wherever a wrong answer costs more than two inferences. Wire transfers, contract commitments, regulatory filings, medical or legal drafting all sit comfortably above a two-inference premium.
Gate before commit, not after — because the published evidence says uncaught errors do not wash out. The Shor benchmark (arXiv:2605.22505, Yonsei University, May 2026) found 94.4% of non-prompt harness errors — memory, tool, and workflow updates — persisted into the final harness. Errors past the gate become the system. The cheap intervention point is the moment before commitment, which is exactly where the challenge round sits, and where XSI-AIMS's blind-spot monitoring principle places it.
What routing does not solve.
Naming the limits is part of the discipline. Routing cannot fix a bad prompt — a mis-specified call routes cheaply and fails cheaply, and quality of task specification is upstream of every mechanism here. Savings depend entirely on workload mix: a fleet whose every call genuinely needs frontier reasoning has a routable share near zero and sees single-digit savings, not 56%. And the requirements taxonomy is engineering work the customer owns — Compass routes on declarations. It sets the floor. It does not write the requirements.
Read the whitepaper.
The full economics, the sensitivity analysis, and the eight-dimension declarable-requirements taxonomy are in the whitepaper — The Economics of Requirements-Scored Model Routing. The list-price spread, the worked scenarios, the reliability arithmetic, and the limits, in one place. How Compass scores a declared requirements vector against the registered field stays XSI implementation IP — the deck is the margin, the same way a carrier never published its routing policy.
Commercial. Contact-gated.
XSI-AIMS Compass is licensed by XSI — reach out and we onboard your team directly.
Route a call by requirements.
// Compass is provider-agnostic — it routes across any OpenAI-compatible API; the three below are one example.
const compass = new Compass({
providers: {
anthropic: { apiKey: process.env.ANTHROPIC_API_KEY },
openai: { apiKey: process.env.OPENAI_API_KEY },
google: { apiKey: process.env.GOOGLE_API_KEY },
},
});
// Route a call by requirement, not by model name.
const result = await compass.call({
prompt,
requirements: {
maxCostCents: 5,
needsTools: true,
adversarialChallenge: true, // safety-critical
},
});Every routing decision auditable.
XSI-AIMS Compass registers as a model-routing instrument with the XSI-AIMS agent registry. Each routing decision — which model served the call, which lost — is recorded to the registry audit log. The registry exposes the decisions for conformance review without exposing the prompts or completions themselves.
Commercial, conformant.
XSI-AIMS Compass is a commercial XSI product implementing the open XSI-AIMS standard — XSI-AIMS defines the governance contract; Compass is one of the ways XSI runs it. Its conformance declaration lives in the public XSI-AIMS agent registry.