infernet::fabric ~/inference-control-plane try it live →
infernet@sg:~$ status --fabric
fabric build v3.1 — now shipping

The inference
control plane for
agent-scale volume.

Infernet routes, batches and orchestrates billions of tokens across every frontier and open model — with sub-second first-token latency, deterministic cost ceilings, and governance your security team will actually sign off on.

No agent left waiting. No token left unaccounted for.
fabric — live readout streaming
throughput1.94 M tok/s
p50 first-token312 ms
active routes4,120
spend vs ceilingon budget
policy = lowest-cost-under-latency
ceiling = { tokens: 2,000,000 · usd: 18.00 }
fallback = [ opus-4 → llama-4-405b ] · trace: on
FABRIC EVENT LOG
ok route policy compiled · 0 errors ok failover path armed · 3 regions ok batch window flushed · 512 reqs ok cost ceiling enforced · within budget ok provider health check · all green ok trace export · observability sink ok route policy compiled · 0 errors ok failover path armed · 3 regions ok batch window flushed · 512 reqs ok cost ceiling enforced · within budget ok provider health check · all green ok trace export · observability sink
$ infernet platform --list-modules

Everything between your agents and the models.

Agents don't fail because a model is slow. They fail because nothing sits between the two coordinating retries, budgets, routing and observability. Infernet is that missing control plane — production-grade, model-agnostic, and yours to govern.

00

Adaptive routing policy-driven

Route each request to the cheapest model that meets your latency and quality bar. Live health-checks reroute around degraded providers before your agents ever notice.

01

Continuous batching throughput engine

Coalesce thousands of concurrent agent calls into optimally packed batches. Squeeze 3–5× more throughput from the same token budget without touching your app code.

02

Deterministic cost ceilings finops native

Hard budget caps per route, tenant and workspace. When a runaway agent hits the ceiling, Infernet degrades gracefully instead of billing you a surprise.

03

Full-trace observability opentelemetry

Every token, every hop, every retry — captured. Replay any agent run, attribute spend to the line of code, and export traces to your existing stack.

04

Governed by design soc 2 · gdpr

SOC 2 controls, PII redaction at the edge, per-region residency, and audit logs that satisfy your compliance team. Bring your own keys, keep your own data.

05

One SDK, every model ts · python · go

Frontier, open-weight, or self-hosted — call them all through a single typed interface. Swap models with a config change, never a rewrite.

$ infernet route --interactive

Route a request right now.

This is the real routing logic, running in your browser. Pick a workload and a policy, hit run, and watch the fabric choose a model, batch the call, stream tokens, and account for every fraction of a cent — the same decision path a production request takes.

infernet://playground — sandbox interactive
Workload
Routing policy
Prompt
No sign-up, no key. Nothing leaves your browser.
# ready. pick a workload + policy, then ▶ route & run
routed to
first token
est. cost
$ infernet cases --by-industry

What the fabric does in production.

Representative deployments across the teams Infernet is built for. Figures reflect measured outcomes from the routing, batching and ceiling controls described above.

Developer tooling

A coding-agent platform tamed a 40-model sprawl

Their IDE agent fanned out to a dozen providers with no unified budget. Infernet collapsed it to one SDK, routing each completion to the cheapest model that cleared the quality bar and batching background lint calls.

Stack: TypeScript SDK · lowest-cost-under-latency policy · 90-day traces
−58%
inference spend
3.9×
throughput / $
Fintech · regulated

A payments team put a hard ceiling on runaway agents

An overnight reconciliation agent occasionally looped and burned budget by morning. Per-route dollar ceilings plus graceful downgrade capped exposure, while edge PII redaction kept card data in-region for audit.

Stack: Python SDK · per-tenant ceilings · region residency · SOC 2
$0
surprise overage
100%
requests in-region
AI-native SaaS

A support-automation SaaS held p50 under load spikes

Ticket surges pushed their single-provider setup past latency SLAs. Continuous batching plus health-check failover kept first-token latency flat through 6× peak traffic — no code change to the agent itself.

Stack: Go SDK · lowest-latency policy · multi-region failover
312ms
p50 held at 6× peak
99.98%
fabric uptime
$ infernet route --trace

A single hop that pays for itself.

  1. 01
    Point your agents at InfernetOne base URL replaces every model-provider SDK you're juggling today.
  2. 02
    Declare a routing policyLatency budget, cost ceiling, quality floor, fallback order — in config, not code.
  3. 03
    The fabric does the workBatching, caching, retries, failover and accounting happen inline, per request.
  4. 04
    Watch the ledgerReal-time spend, throughput and trace data — down to the individual agent step.
fig.2 — request pathinfernet/fabric
Your agents & services SDK
Infernet routing fabric POLICY
Batching · caching · failover INLINE
Frontier · open · self-hosted models ANY
$ infernet metrics --last 30d

The numbers, live from the fabric.

1.9M
tokens / sec sustained
312ms
p50 first-token
4.3×
throughput / dollar
99.98%
fabric uptime
$ infernet pricing --per throughput

Priced on throughput, not per seat.

You already pay the model providers directly. Infernet adds a thin, transparent fabric fee — and typically saves more than it costs through routing and batching.

planrateincludes
Builder $0/ month Up to 5M tokens/mo · all models, one SDK · basic routing · 7-day traces · community support start free
Scale ◆ most adopted 1.4%of routed spend Unlimited tokens · continuous batching · cost ceilings & alerts · 90-day traces + export · SOC 2 · priority support start building
Enterprise Custom Private fabric & region residency · BYO keys · self-hosted models · SSO/RBAC/audit · custom SLA & DPA · solutions engineering talk to us
You keep your own provider accounts. Infernet never marks up or resells model tokens.
$ infernet faq --engineers

The things engineers ask first.

Q1Is Infernet an AI middleman that resells model access?+
No. You keep your own provider accounts and keys — Infernet never marks up or resells model tokens. We sit in front of your models as a routing, batching and governance control plane. You pay the providers directly; we charge a transparent fabric fee for the orchestration layer.
Q2Which models can I route to?+
Every major frontier model, the leading open-weight families, and any endpoint you self-host. Because the SDK is model-agnostic, adding or swapping a model is a config change — your agent code never has to know which model answered.
Q3How do the cost ceilings actually work?+
You set hard token and dollar caps per route, tenant or workspace. As traffic approaches the ceiling, Infernet can throttle, downgrade to a cheaper model, or reject with a typed error — you choose the degradation strategy. No more surprise invoices from a looping agent.
Q4What does it take to migrate?+
Change one base URL and drop in the SDK. Most teams route their first production workload in an afternoon. Existing prompts, tools and agent frameworks keep working unchanged.
Q5Where does my data go?+
Nowhere you don't allow. PII redaction runs at the edge before requests leave your region, traces can be scrubbed of payloads, and Enterprise plans pin all processing to a residency region of your choice. We're SOC 2 aligned with a DPA available.

Give your agents an inference layer that keeps up.

Start free, route your first million tokens today, and watch the ledger balance itself. Talk to us when you're ready to scale.