01
Overview
Fastly has launched a new control layer for production AI traffic. On September 21, 2026, the edge-cloud company announced AI Runtime Control (ARC), AI Firewall and expanded API Security capabilities aimed at organizations running AI applications and autonomous agents across multiple model providers.
ARC is not a model host. It sits between an application and the LLM providers or self-hosted model endpoints the application already uses. Requests go to ARC first; ARC applies routing and limits, substitutes the underlying provider credentials, forwards the request and records usage.
02
What is Fastly AI Runtime Control?
Fastly AI Runtime Control is an AI gateway and governance control plane. Its core idea is to replace direct provider calls with one managed layer where teams can see and control model traffic.
Fastly documents five especially important capabilities: virtual keys, usage monitoring, budget and rate limits, multi-provider failover, and optional security inspection through AI Firewall.
03
Virtual keys replace raw provider credentials
Applications authenticate to ARC with Fastly-issued virtual keys rather than holding the raw API keys for each upstream provider. A virtual key maps to an ordered provider list and carries the routing and limits applied to its requests.
Keys can be separated by application, environment, developer or another operational boundary. Fastly says a key can be refreshed or revoked without changing the underlying provider account.
ARC supports providers configured with customer credentials and can also route to a model endpoint the customer hosts. Fastly explicitly says ARC does not host or run the models itself.
04
Multi-provider failover
A virtual key can contain an ordered list of providers. ARC sends a request to the first target and, if that provider cannot serve it, can try the next target.
Fastly's own tutorial demonstrates Claude Code traffic routed to Anthropic with AWS Bedrock configured as a backup provider serving the same model. Each new request starts from the top of the configured list, so traffic returns to the preferred provider after it recovers.
This is provider failover, not evidence that arbitrary models are interchangeable. The backup needs to satisfy the workload and model requirements.
05
AI spending and rate controls
ARC records request volume, input and output token counts and usage by virtual key and model. Teams can use summary and request-level logging views to investigate which keys and models are driving usage.
Budget limits can be attached to individual virtual keys. Fastly documents both alert-only thresholds and Alert & Block limits that stop additional requests after the configured monthly dollar cap is reached.
Rate limits can use requests per minute or tokens per minute. Fastly notes that token-based rate limits are best-effort because the final token count is not known until a response completes.
These controls are particularly relevant to autonomous agents, where an unexpected loop or burst can consume provider quota or budget quickly.
06
What is Fastly AI Firewall?
AI Firewall is an optional, separately purchased security layer inside ARC. Fastly says it inspects AI requests and responses for threats such as prompt injection using its SmartParse detection system and additional guardrails.
The company describes the latency impact as near zero. That is a vendor claim rather than an independent benchmark, so latency-sensitive teams should test it with their own traffic before making production assumptions.
AI Firewall should also not be confused with Fastly's conventional web application firewall. The new product is designed for threats specific to LLM-powered applications.
07
API Security for AI agents
The September launch also expands Fastly's API Security positioning around autonomous agents. Fastly says API contracts can be enforced against agentic, agent-assisted and conventional traffic, allowing non-conforming requests to be observed or blocked.
That matters because agents can call enterprise APIs directly and may generate malformed or unsupported operations if they operate outside expected boundaries.
08
Is Fastly ARC generally available?
Fastly's September 21 launch announcement says AI Runtime Control, AI Firewall and API Security are now available. Some current ARC documentation still carries beta-release language, however. That lifecycle nuance should be checked against Fastly's latest documentation before a production commitment.
09
How much does Fastly AI Runtime Control cost?
Fastly's public product and launch pages reviewed for this article do not provide a simple standalone ARC price. The product page directs prospective users toward Fastly rather than publishing a universal ARC price table.
Do not infer ARC pricing from the upstream model spend ARC measures. Provider token charges and Fastly product charges are separate concepts.
10
Who should evaluate ARC?
ARC is most relevant to teams that use multiple LLM providers, need centralized AI cost controls, operate autonomous agents, want provider failover, or need to keep raw model-provider credentials out of individual applications.
A small project calling one model directly may not need another control plane. The operational value increases as provider count, agent autonomy, spend and security requirements grow.
11
What to test before adopting it
Before routing production traffic through ARC, teams should test provider compatibility, failover behavior, latency, logging/data-retention requirements, budget-limit behavior, rate limits, role requirements and the lifecycle status of the features they plan to use.
Security-sensitive deployments should separately validate AI Firewall against their own prompt-injection threat model instead of treating a vendor protection layer as a guarantee.
12
Bottom line
Fastly's September 21 release turns the company's edge platform toward a fast-growing infrastructure problem: AI applications increasingly call several models, agents can consume unpredictable amounts of quota, and model-facing traffic creates security and governance requirements that ordinary API integrations were not designed to handle.
AI Runtime Control centralizes those calls behind virtual keys, routing, monitoring, budgets, rate limits and failover. AI Firewall adds a separate LLM-specific security layer. The product is worth watching because it puts AI gateway, cost governance and prompt-injection defense into the same edge-control stack—but pricing and current lifecycle labels should be rechecked before publication or purchase.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.