Cribl StreamAI Routes AI Requests by Cost and Accuracy With Free Auto-Routed Inference
Cribl announced StreamAI on September 29, 2026, adding an enterprise AI gateway and model router to its telemetry platform. The gateway can select a model for each request, enforce spending limits, redact sensitive data in prompts and responses, and record model calls and routing decisions as normalized telemetry.
The commercial model is unusual enough to matter for deployment planning. Cribl says inference is free when StreamAI automatically selects from its benchmarked models. Requests pinned to a specific model are charged at the provider's published token rate with no Cribl markup, according to Cribl comments reported by TechTarget. StreamAI is currently announced for early access and will be available soon; Cribl has not given a general-availability date.
Cribl built the router around its SecIT Bench evaluation of 20 AI models across 30 IT and security investigations. The company reports a 17% spread in diagnostic accuracy alongside a 20x range in investigation cost. These vendor-reported results establish the routing methodology and claimed economics for the tested IT and security workloads; they are not a general model-quality ranking.
What StreamAI controls
StreamAI sits between applications or agents and the models they call. Its announced controls cover routing, cost, security and observability.
| Capability | Announced behavior |
|---|---|
| Model routing | Selects a model based on workload fit, benchmark results and economics |
| Budget controls | Applies spending limits and circuit breakers to token consumption |
| Fallback | Can move requests to an alternative model when an application reaches its cost limit |
| Sensitive-data handling | Redacts sensitive information in outgoing prompts and incoming responses |
| Security controls | Screens interactions to protect data, restrict model access and reduce malicious requests |
| Auditability | Records model calls and routing decisions as normalized telemetry |
| Provider choice | Supports proprietary and open-source model options |
| Auto-routed inference | Cribl says inference is free when its router chooses the benchmarked model |
This architecture addresses a growing operational problem with agentic applications: one user action can produce a chain of model calls with very different cost and capability requirements. A gateway can apply a common policy before those calls reach model providers and retain a record of which model handled each request.
The benchmark behind the router
Cribl says SecIT Bench gives multiple models the same incident material, including security breaches, service outages and performance problems. Its initial test covered 20 models and 30 investigations.
The reported 17% accuracy spread versus a 20x cost spread is the basis for StreamAI's economic argument. If several models produce sufficiently useful results for a task, routing lower-value or easier work to a less expensive model can reduce aggregate inference spend while preserving higher-cost models for workloads where they provide a measurable advantage.
The benchmark is specifically oriented toward IT and security investigations. Teams evaluating StreamAI for coding, document processing, customer support or other domains should validate routing quality against their own workloads because the published SecIT Bench result does not establish equivalent accuracy relationships outside its tested task set.
Cost controls for agents
Agent workflows make inference budgeting harder because a task can branch, retry and invoke models repeatedly. StreamAI adds circuit breakers that can stop applications from exceeding configured quotas and can fall back to another model after a cost threshold is reached.
TechTarget reports an additional pricing detail from Cribl: automatically routed inference is free, while explicitly selecting a model uses that provider's published token rates without a Cribl markup. Cribl also records cost for each call so teams can audit the savings attributed to routing.
That makes the routing decision itself an observable event. For platform teams, the useful metrics extend beyond aggregate token consumption to model selection, per-call cost, policy decisions and workload outcomes.
Security and governance
StreamAI applies bidirectional redaction so sensitive values can be removed before a prompt leaves the environment and from a model response before it reaches downstream consumers. Cribl also says the gateway can restrict model access, block malicious requests and reduce unsafe responses or actions.
Every model call and routing decision is normalized into telemetry for audit and governance. That can give security and platform teams a common record for questions such as which provider received a request, why a router selected a model, how much the call cost and which policy was applied.
The controls build on Cribl's broader telemetry platform and Cribl-Privacy, the telemetry-focused model used by Cribl Guard for background detection. Production evaluation should test the redaction and policy layer against the organization's sensitive-data classes and agent permissions before assigning it application-security responsibilities.
Availability and deployment fit
Cribl describes StreamAI as coming soon and is directing interested customers to its account teams for early access. The company has not published a general-availability date in the launch announcement.
The strongest initial fit is likely to be organizations already operating high-volume AI workloads across multiple model providers, particularly IT and security teams whose workloads resemble SecIT Bench. The combination of model routing, per-application budgets, sensitive-data controls and audit telemetry can centralize policy currently implemented separately in each agent or application.
The key production test is whether the router's task-specific model selection preserves required accuracy and latency while reducing total inference cost. Cribl's published benchmark provides an initial hypothesis; workload-specific evaluation determines whether those economics carry into a given environment.
Sources
- Cribl-issued StreamAI announcement, September 29, 2026: https://www.globenewswire.com/news-release/2026/09/29/3371035/0/en/cribl-launches-streamai-with-free-inference-giving-enterprises-control-over-ai-costs.html
- TechTarget independent analysis and Cribl pricing comments, September 29, 2026: https://www.techtarget.com/it-infrastructure/news/366651477/Cribl-targets-SIEM-data-costs-with-new-Detect-tool
- CriblCon 2026 event documentation: https://criblcon.cribl.io/faqs/