agentgateway v1.6.0 prices LLM calls from a catalog it ships itself

The logged cost checks out by hand. Eight input tokens at $2 per million plus 14 output tokens at $10 per million is $0.000156, the total agentgateway wrote to the access log ^[5][6][19]. Output tokens were $0.000140 of that, about 90 percent ^[20]. Those log lines come from one author's hands-on run against a live gateway and a real Claude Sonnet 5 backend, published with code and captured output in a demo repo ^[18].

In v1.5, the per-token rates had to sit in a modelCatalog block that someone kept current by hand ^[3]. In v1.6.0, generally available since October 2, they come from a catalog the project ships and maintains ^[1][4]. Getting them into the log still takes configuration. The author added a `frontendPolicies.accessLog.add` block mapping CEL expressions such as `llm.cost.total` and `llm.costRates.input` to log keys ^[7]. According to the post, a budgets policy reads those same fields, so the shipped catalog feeds the per-key USD budgets that arrived in v1.5.0 ^[8][2].

The demo config contains no budgets policy ^[9]. So the case that USD budgets now work without a rate table rests on that shared-field argument, with no overspend 429 observed. I think the argument holds. Before deleting a production table I would still want one budget tripped against catalog prices.

Coverage is the other open item. The post says the catalog covers more than Anthropic, but the run validated only claude-sonnet-5 ^[13][9]. It does not say how often the catalog changes between releases, or whether a hand-written modelCatalog entry overrides a shipped price. Retiring a table is therefore a per-model decision: a model whose log lines show populated `cost.rate.input` and `cost.rate.output` fields can lose its entry ^[6][7].

The rate limiter is good engineering. One localRateLimit rule keyed on `request.headers["x-api-key"]` gave key A three 200s and a 429 on its fourth request inside a minute ^[9][10]. Key B went through in that same minute, with no second rule and no restart ^[11][12]. Any CEL expression over the request can be the key, so a JWT claim or a source IP buckets the same way ^[12].

It counts admissions. In one run the upstream peer closed the connection without a TLS close_notify, the gateway returned 503, and `x-ratelimit-remaining` still dropped ^[14]. The provider produced nothing billable on that call, and the bucket charged for it anyway ^[14]. In a 3-request bucket, that one failure used a third of the key's minute ^[21]. The author wrote that "a bad backend day eats your quota exactly like a good one" ^[15]. The policy type is `requests`, so it caps calls, and spend control stays with budgets ^[9][2].

Buckets live in the proxy instance that created them. Quota shared across replicas is the job of a separate policy, remoteRateLimit ^[16]. Spread one key evenly across three replicas with this config and it gets up to nine requests a minute ^[22]. The run used a single container and no Kubernetes, so the AgentgatewayModel CRD, now on by default in the Helm chart, went untested ^[17].