Build1 publisher2 min readPublished
Fastly puts a dollar budget and a prompt scan in front of every model call
Fastly's AI Runtime Control is an endpoint you point an application at. It issues virtual keys against your own model provider account, and a key can carry a dollar budget, a rate limit and an ordered list of fallback models.
The Engineer · Build desk
What happened
- Fastly launched AI Runtime Control, AI Firewall and expanded API security, all immediately available, putting model access, prompt inspection and agent permissions into the request path it already operates.
- AI Runtime Control is a Fastly-managed endpoint between applications and public or self-hosted model providers, with customers keeping their own provider accounts and negotiated rates.
- Administrators issue virtual keys per application, developer or team, then meter token consumption, set dollar budgets and rate limits, revoke access, and search request and response logs.
- The third control applies Fastly's API security to agents, letting teams define API contracts and observe or block calls that use unsupported operations, carry malformed data or leave the service boundary.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost Customers keep paying their model provider directly, so Fastly's separately published rates for its AI products land on top of the token bill, and the control layer has to justify its own line item.
- capability Spend and rate limits attach to a key scoped to one team or one application, so a single consumer can be capped or cut off without disturbing anything else calling the same provider account.
- exposure Every model call now crosses infrastructure operated by the delivery vendor. That makes prompt and response content a Fastly data-handling question for whoever signs off on the security review.
- decision Fastly customers have to decide whether a request path they already pay for is worth more than a specialist AI gateway they would pick on published efficacy and latency numbers.
Fastly built its inspection product to check requests arriving at an application; it now checks the ones leaving, on their way to models and enterprise APIs, according to runtimewire.com [16]. Adoption means repointing the application at a Fastly-managed endpoint and calling it with a virtual key [2][3]. The provider account stays yours, at the rates you negotiated [3]. Fastly's documentation describes how those requests are routed and logged [4].
Because the endpoint sits in the path, a revoked key stops traffic before it reaches the provider [2][5].
A single key can also name an ordered list of models, and Fastly retries with the backup provider when the first one returns an error [6]. Taking cross-provider retry out of application code is good engineering. It transfers only if the caller can consume the second model's answer, so parsers for structured output and tool calls have to accept both. Fastly describes the trigger as an error from the first provider [6]. A provider that is slow but still answering stays a separate problem for the application to handle.
AI Firewall inspects the prompt before it reaches the selected model, using an updated version of Fastly's SmartParse detection engine plus custom guardrails aimed at prompt injection [7]. Fastly has not supplied an efficacy benchmark, a false-positive rate or a latency measurement for it [8]. Whether inline inspection is affordable comes down to the share of legitimate prompts it blocks and the milliseconds it adds before the first token. Both are properties of your own prompt traffic, so you measure them in staging.
One level up, the agent control enforces an API contract you define [11]. Coverage therefore tracks how current your definitions are, and a contract that omits a new legitimate operation will flag it.
"Organizations need control in production, at runtime, without delay, friction, or disruption," Kelly Shortridge, Fastly's chief product officer, said in the announcement [14][15]. The distribution under that pitch is large: Fastly says its network serves over five trillion requests per day as of March 31st, 2026, which is more than 57 million a second, and had 622 Tbps of capacity as of June 30th, 2026 [9][1][10]. runtimewire.com frames the commercial test as a question: do existing customers pay Fastly to consolidate AI routing and security, or do they add another specialist vendor [12].
What to watch
- Whether Fastly publishes a latency measurement or false-positive rate for AI Firewall inspection.
- Whether virtual-key failover extends beyond provider errors to timeouts and rate-limit responses.
- Whether Fastly discloses customer counts or usage for the AI products in a later announcement.