Cribl announced StreamAI on Tuesday: an enterprise AI gateway with a built-in model router, positioned as the company's answer to the runaway token bills that arrive once agents start coordinating multi-step work with minimal human supervision. The headline hook: inference is free for customers who let StreamAI's automatic routing pick the model from its benchmarked lineup — no more surprise token bills.

The launch, announced at CriblCon, extends the company's core business — routing massive volumes of telemetry data — into routing AI prompts. StreamAI decides which model, proprietary or open-source, each request should hit, while enforcing budgets and spending cutoffs in real time.

The research behind the router#

The router is built on Cribl's SecIT Bench research, which put 20 AI models through 30 real-world IT and security investigations — the same messy incident data, from security breaches to service outages to performance problems, handed to every model. The result is the whole sales pitch in two numbers: a 17% spread in diagnostic accuracy across the models, against a 20x range in investigation spend. The most expensive model is not automatically the best fit for every workflow.

That finding is what makes the router more than a cost heuristic. When an AI-powered enterprise application reaches its cost limit, StreamAI automatically falls back to an alternative model to keep usage running without disruption, and built-in circuit breakers let IT teams cap token consumption before applications blow through quotas. "AI is not one-size-fits-all. The right model depends on the job, the context, and the economics of the request," said Cribl co-founder and CEO Clint Sharp. "Customers shouldn't have to send every prompt to the most expensive model or build their AI strategy around a single provider."

Dark-themed Cribl cluster metrics dashboard with connection and throughput charts
Image: Cribl (cribl.io)

Free inference, with a catch#

The free-inference offer comes with a clear structure, spelled out by Nikhil Mungel, Cribl's head of AI R&D, in comments this week: "Inference is free when StreamAI chooses the model." Requests sent directly to a specific model are billed at the provider's published token rates, with no Cribl markup — and the cost of every call is logged so customers can verify the savings themselves.

It is an aggressive pricing move in a year when AI gateways have become table stakes. Vendors from SaaS, DevSecOps, observability, and SecOps have all launched similar offerings in 2026. A Cribl customer quoted by TechTarget put the appeal bluntly: free inference when the router picks the model, and provider rates with no markup when you pick, "removes the worry that the gateway eats the savings."

Governance baked into the routing layer#

StreamAI also folds security and compliance controls directly into the routing layer. It builds on Cribl-Privacy, the custom telemetry-focused model behind Cribl Guard's background detection, designed for the high-volume, semi-structured data found in production telemetry. The gateway applies bidirectional sensitive-data redaction — shielding proprietary information, credentials, and sensitive data in both outgoing prompts and incoming responses — blocks malicious requests, controls what models can access, and records every model call and routing decision as normalized, audit-ready telemetry.

"It provides a single control point," Mungel said of the design. For enterprises scaling from a handful of pilots to fleets of agents across business units, that single choke point is where policy, visibility, and cost control converge — and where a compromise would hurt most, which is why the audit trail matters.

Cribl Stream routing canvas showing data sources on the left connected to destinations on the right
Image: Cribl (cribl.io)

Why it matters#

ItemDetail
LaunchedSeptember 29, 2026, at CriblCon
What it isEnterprise AI gateway with an automatic model router
HookFree inference when StreamAI's routing picks the model
EvidenceSecIT Bench: 20 models, 30 real-world investigations; 17% accuracy spread, 20x cost spread
GovernanceCribl-Privacy redaction, audit-ready telemetry of every call
AvailabilityAvailable soon; early access through account teams

Cribl's entry matters because of who it is, not just what it shipped. The company — founded in 2018, remote-first with a San Francisco office, and trusted by over half of the Fortune 100 — already sits at the choke point of enterprise telemetry. An AI gateway from a company that already routes the data is a natural extension, and its router is informed by Cribl's own published benchmark research rather than static rules about which team sent the request or whether the monthly budget is spent.

The caveat, from Omdia analyst Torsten Volk: the savings thesis only holds if SecIT Bench is right. If the accuracy spread really is just 17% against a 20x cost spread, "there is a lot of savings potential, compared to a standard AI gateway that only decides based on static rules." But "how big of an impact this could make is now up to the accuracy of Cribl's SecIT Bench results" — independent validation will decide whether the router is a genuine cost engine or a marketing layer over commodity routing.

What to watch#

  1. The SecIT Bench methodology. The entire value proposition rests on that 17%-vs-20x finding. Watch for Cribl to publish the full methodology and for independent reproductions.
  2. Latency in real environments. A router adds a hop. One customer told TechTarget that StreamAI's latency, "which depends on each customer's environment," is "the first thing an ops team will ask" before deploying it.
  3. The free-inference economics. Cribl is absorbing inference costs on router-picked calls — watch how long the offer lasts and what "benchmarked AI models" turns out to cover.
  4. General availability. StreamAI is "available soon" with early access through account teams. Watch the rollout cadence — and which models land in the benchmarked lineup.

Sources#