Originally published at webofmike.com on 2026-10-06. The demo repo and every command in it were run before publishing.
agentgateway v1.6.0 went GA on October 2. Two features are worth a hands-on look before anything else: a built-in LLM pricing catalog that gives you cost tracking with zero configuration, and per-key CEL rate limiting that buckets quota by any request attribute you pick. I ran both against a live gateway and a real Claude Sonnet 5 backend. Code and captured output are in themsquared/agw-16-hands-on.
Cost tracking used to mean declaring rates by hand
I wrote up agentgateway’s per-API-key budgets back in v1.5.0: a budgets list per key, a limit in tokens or USD, and a 429 before the provider ever sees the request. That post also covered the gap underneath it. Budgets only work once the gateway knows what a token costs, and in 1.5 that meant a modelCatalog block naming every model and its per-million-token input and output rate, maintained by hand.
v1.6.0 replaces that with a model catalog agentgateway ships and maintains itself. The demo’s entire LLM route config:
llm:
port: 4000
policies:
localRateLimit:
– type: requests
maxTokens: 3
tokensPerFill: 3
fillInterval: 60s
key: request.headers[“x-api-key”]
models:
– name: claude
provider: anthropic
params:
model: claude-sonnet-5
apiKey: $ANTHROPIC_API_KEY
Enter fullscreen mode
Exit fullscreen mode
No modelCatalog anywhere. The access log still produces full cost data because claude-sonnet-5 is a model the built-in catalog already knows:
http.status=200 gen_ai.usage.input_tokens=8 gen_ai.usage.output_tokens=14
agw.ai.usage.cost.total=0.000156
cost.total=0.000156 cost.rate.input=2 cost.rate.output=10
Enter fullscreen mode
Exit fullscreen mode
cost.rate.input and cost.rate.output are USD per million tokens, pulled straight from the catalog. The math checks out: (8 * 2 + 14 * 10) / 1,000,000 = 0.000156. Nothing in the config told agentgateway what this model costs; it already knew.
Getting those fields into the access log in the first place is a frontendPolicies.accessLog.add block mapping CEL expressions to log keys:
frontendPolicies:
accessLog:
add:
model.requested: llm.requestModel
model.served: llm.responseModel
tokens.input: llm.inputTokens
tokens.output: llm.outputTokens
cost.total: llm.cost.total
cost.rate.input: llm.costRates.input
cost.rate.output: llm.costRates.output
Enter fullscreen mode
Exit fullscreen mode
Those same CEL fields are what you’d read in a budgets policy from the 1.5 post, so this isn’t a separate feature bolted on. It’s the same cost-accounting plumbing, now backed by a catalog instead of a hand-maintained table.
Per-key rate limits need one rule, not one rule per key
The second feature in this run is localRateLimit, keyed by a CEL expression rather than a fixed value. The config above keys on request.headers[“x-api-key”], with a 3-request bucket that refills every 60 seconds. One rule, evaluated per request, buckets independently per header value:
# key A, requests 1-3: all 200
http.status=200 … cost.total=0.000156 …
http.status=200 …
http.status=200 …
# key A, request 4, same minute
http.status=429 error=”rate limit exceeded” reason=RateLimit
# key B, same minute, different header value
http.status=200 … cost.total=0.000176
Enter fullscreen mode
Exit fullscreen mode
Key B never saw key A’s limit. There’s no second localRateLimit entry for it, no restart to pick up a new key. Any CEL expression over the request works as the bucket key, so this generalizes past API keys to things like a JWT claim or a source IP.
The gotcha: a 503 still spends a token
One run of the demo hit a transient upstream failure partway through:
http.status=503 error=”upstream call failed: SendRequest: connection error: peer closed connection without sending TLS close_notify” reason=UpstreamFailure
Enter fullscreen mode
Exit fullscreen mode
x-ratelimit-remaining still dropped on that request. The call never reached Anthropic successfully, a 503 is the opposite of a billable response, but it still counted as one of the 3 admitted requests in the bucket. localRateLimit counts requests agentgateway admits, not requests that succeed upstream. If you’re sizing maxTokens close to real traffic, budget headroom for upstream flakiness, because a bad backend day eats your quota exactly like a good one.
What this run doesn’t cover
- No Kubernetes. v1.6.0’s AgentgatewayModel CRD, now on by default in the Helm chart, and K8s-native session affinity are cluster-side features this standalone run doesn’t exercise.
- One provider. The built-in catalog covers more than Anthropic; this demo only validates the provider with a key on hand.
- remoteRateLimit is a different policy, for quota shared across replicas. localRateLimit buckets live in the single proxy instance that created them, which is exactly what makes this demo’s single-container setup representative of the behavior.
Run it yourself
Requirements: Docker and an ANTHROPIC_API_KEY. Tested on macOS (Apple silicon) against cr.agentgateway.dev/agentgateway:v1.6.0.
git clone https://github.com/themsquared/agw-16-hands-on.git
cd agw-16-hands-on
export ANTHROPIC_API_KEY=sk-ant-…
./run-demo.sh
Enter fullscreen mode
Exit fullscreen mode
To check the config against the v1.6.0 schema without sending any traffic:
docker run –rm -v “$PWD/config/config.yaml:/config/config.yaml:ro” \
-e ANTHROPIC_API_KEY=dummy-for-validate \
cr.agentgateway.dev/agentgateway:v1.6.0 -f /config/config.yaml –validate-only
Enter fullscreen mode
Exit fullscreen mode
Tear down with docker rm -f agw16-demo.
What changed, concretely
Going from v1.5.0 to v1.6.0, the same cost-and-quota problem from the per-key budgets post now needs less from you: no catalog to maintain, and a quota rule that keys itself by request content instead of being written once per key. The gotcha is the same shape either version: a limiter counts what it admits, not what succeeds, and that is worth checking against your own traffic patterns before you pick a maxTokens value. Repo and full captured output: themsquared/agw-16-hands-on.
Frequently asked questions
Does agentgateway v1.6.0 need a model catalog configured for cost tracking?
No. v1.6.0 ships a built-in model catalog, so a route to a known model like claude-sonnet-5 produces llm.cost.total and the input/output cost rates automatically, with no modelCatalog block anywhere in the config. Earlier versions required declaring rates by hand for every model in use.
How does agentgateway’s per-key CEL rate limiting work?
A single localRateLimit policy keyed on an expression like request.headers[\
Does a failed upstream request still count against an agentgateway rate limit bucket?
Yes. A request that fails upstream, such as a dropped TLS connection, still consumes a token from the bucket. localRateLimit counts requests admitted at the gateway, not successful upstream responses, so a flaky backend can eat into a tight per-key quota without a single call completing.
Is agentgateway’s remoteRateLimit the same as localRateLimit?
No. localRateLimit buckets live in memory on the single gateway instance that created them, which is exactly the single-container setup this demo runs. remoteRateLimit is a separate policy that shares one quota across replicas, which matters once more than one gateway instance sits behind a load balancer and you need one limit enforced across all of them.
Canonical version, with machine-readable markdown at https://webofmike.com/agentgateway-v16-cost-tracking/index.md: https://webofmike.com/agentgateway-v16-cost-tracking/