The inference control plane for AI agents
Persistent agents re-send the same instructions, tools, and context on every call. Zumik measures that repeated work, turns the stable parts into reusable state, and routes each request through the cheapest reliable path. OpenAI-compatible /v1, native stateful /v2.
What Zumik does
Measure reuse first. Optimize the path the evidence supports.
Prompt caching is becoming a provider feature, not a product. Zumik is the control plane around it: stable state semantics, reproducible routing, honest QoS outcomes, purge evidence, and replay-backed escalation only when it pays off.
Workload analysis
Score how much reuse is actually possible before changing infrastructure, with opportunity and capture reported separately.
02State semantics
Give applications stable, tenant-safe, opaque handles for reusable inputs instead of re-sending them.
03Provider optimization
Use provider-native caching, batch tiers, service tiers, and routing keys across five first-class providers.
04Reproducibility
Pin model aliases, prompt-compiler versions, and snapshots so any routing decision can be explained and replayed.
05QoS accountability
Report whether latency and reliability targets were met: admitted, degraded, missed, rejected, or expired.
06Purge guarantees
Return profile-specific deletion evidence instead of vague retention claims.
Two surfaces, one engine
An exact compatibility path and a native stateful path.
/v1 exact OpenAI compatibility
MigrateResponses, Chat Completions, Embeddings, and Models keep upstream JSON shapes intact, with token-by-token streaming relayed live from every provider. Zumik hints ride on optional HTTP headers, and each response reports what routing saved you via Agent-Savings-Micros.
from openai import OpenAI
client = OpenAI(
base_url="https://api.zumik.ai/v1",
api_key="zk_live_...",
)
r = client.responses.create(
model="code.balanced",
input="Why did CI fail on main?",
)/v2 native stateful API
OptimizeArtifacts, bundles, sessions, branches, diagnostics, purge jobs, replay runs, and rich telemetry live here. This is where reuse stops being a hope and becomes a measured quantity.
curl https://api.zumik.ai/v2/artifacts \
-H "Authorization: Bearer zk_live_..." \
-d '{"artifact_type":"policy",
"content":"Run the linter before commit."}'
# => { "id": "art_01JY…" }Data & tools
Model pricing and caching data we measure ourselves
Every page below is built from the Zumik trace corpus, not a marketing deck. Pricing, caching economics, capture rates, and TTFT savings you can actually route on.
Model catalog
Pricing, context, cache discounts and reuse capture across the catalog.
OpenPrompt-cache benchmarks
Measured cache-capture and TTFT savings from the Zumik trace corpus.
OpenLeaderboards
Models ranked by reuse-adjusted cost, TTFT and agent fit.
OpenPricing calculator
Estimate monthly spend with and without prompt reuse.
OpenWorkload Reuse Score
Score reuse potential from the six components, live, in your browser.
OpenPrompt-layout linter
Flag the ordering mistakes that quietly kill cache reuse.
OpenWorkload trends
Aggregate reuse patterns across agent workloads we have measured.
OpenCompare
Model-vs-model and Zumik-vs-gateway comparisons.
OpenCapabilities
Built for agent workloads that repeat themselves.
Stateful sessions
Artifacts, bundles, branches and snapshots with opaque handles.
Workload diagnostics
Score reuse opportunity vs. realized capture before you migrate.
Start a diagnostic
Free workload scan to proven pilot, no payment method to begin.
Model aliases
Immutable alias releases keep routing reproducible and auditable.
Provider routing
Managed, BYOK and failover policy across five first-class providers.
Purge receipts
Delete revokes handles; purge jobs emit profile-specific evidence.
BYOC runtimes
Escalate to self-hosted KV orchestration only when replay proves it.
Zumik is not a new inference engine, a generic gateway clone, a model marketplace, or a GPU cloud. It is a control plane that makes repeated agent work cheaper, faster, and reproducible. When managed providers already capture the savings, it will tell you to stay put.
Frequently asked
Questions teams ask before changing their inference stack.
What is Zumik?
Zumik is a provider-first stateful inference control plane for persistent AI agents. It measures repeated inference work, preserves reusable state behind opaque handles, and routes each request through the cheapest reliable managed-provider, BYOK, or BYOC path.
How does Zumik reduce LLM inference cost?
It separates reuse opportunity from realized capture, keeps stable content at the front of requests so provider prompt caches hit, moves background work to batch tiers, and only recommends self-hosting when replay proves it beats managed providers.
Is Zumik OpenAI-compatible?
Yes. The /v1 surface mirrors OpenAI request and response bodies exactly for Responses, Chat Completions, Embeddings, and Models. Migration is a single base-URL change. Proprietary behavior rides on optional headers or the native /v2 API.
Which providers does Zumik support?
OpenAI, Anthropic, xAI, Google Gemini, and Fireworks AI are first-class for both managed and BYOK execution, with broader coverage through Bifrost where policy allows. OpenRouter is reserved as an explicit last-resort outage bridge.
How much does Zumik cost?
Zumik is pay-as-you-go on prepaid credits: load any amount from $5 and spend it per token at published per-model rates. There is no subscription and no free inference tier. On managed routing the per-model price lands under provider list; BYOK and BYOC pay a small control-plane fee per request.
Does Zumik support streaming?
Yes. stream: true on /v1/chat/completions relays tokens from the provider as they are produced, on all five first-class providers, so time to first token is the provider’s own. /v1/responses supports streaming and tool calls with the Responses SSE event protocol.
Do I need to self-host (BYOC)?
Usually not. Managed-provider caching, batch tiers, and alias routing capture most of the available savings for typical agent workloads. BYOC is an evidence-gated escalation for dedicated SLOs, concentrated hot-model volume, or strict isolation requirements.
Start by measuring your reuse
Point an OpenAI client at Zumik, run a diagnostic on real traffic, and decide what to optimize from evidence. Pay-as-you-go credits from $5, no subscription.