Skip to content

The cross-harness, self-improving memory layer for AI agents. Join our community: https://discord.com/invite/zdNChS2fBu

fix(pricing): report only cache fallbacks used

Rudy Celeklid6614cc435b8
diff --git a/cli/beacon/internal/pricing/cache_fallback_test.go b/cli/beacon/internal/pricing/cache_fallback_test.go
new file mode 100644
index 00000000..0cc3d4be
--- /dev/null
+++ b/cli/beacon/internal/pricing/cache_fallback_test.go
@@ -0,0 +1,41 @@
+package pricing
+
+import "testing"
+
+func TestCacheWriteFallbackDescribesOnlyRatesUsed(t *testing.T) {
+ rates := Rates{RateSet: RateSet{Input: 2_000_000, Output: 3_000_000, CacheWrite1h: 5_000_000}}
+ for _, tc := range []struct {
+ name string
+ total, oneHour int64
+ wantCost Microdollars
+ wantFallback bool
+ }{
+ {"zero_counts", 0, 0, 0, false},
+ {"only_one_hour", 1000, 1000, 5000, false},
+ {"mixed_ttls", 1000, 600, 3800, true},
+ {"only_five_minutes", 1000, 0, 2000, true},
+ {"clamped_one_hour", 1000, 2000, 5000, false},
+ } {
+ t.Run(tc.name, func(t *testing.T) {
+ got := Cost(rates, Tokens{CacheCreation: tc.total, CacheCreation1h: tc.oneHour}, false)
+ if got.Cost != tc.wantCost || got.CacheWriteAtInputRate != tc.wantFallback || got.CacheWrite1hAtWriteRate {
+ t.Fatalf("estimate = %+v; want cost %d input fallback %v and no 1h fallback", got, tc.wantCost, tc.wantFallback)
+ }
+ })
+ }
+}
+
+func TestOneHourOnlyOverrideDoesNotReportInputFallback(t *testing.T) {
+ overrides, err := LoadOverrides([]byte(`{"schema":"beacon.pricing.overrides/v1","models":{"internal-model":{"input_usd_per_mtok":2,"output_usd_per_mtok":3,"cache_write_1h_usd_per_mtok":5}}}`), overridesCatalog(t))
+ if err != nil {
+ t.Fatal(err)
+ }
+ rates, ok := overrides.Rates("internal-model")
+ if !ok {
+ t.Fatal("override model missing")
+ }
+ got := Cost(rates, Tokens{CacheCreation: 1000, CacheCreation1h: 1000}, false)
+ if got.Cost != 5000 || got.CacheWriteAtInputRate || got.CacheWrite1hAtWriteRate {
+ t.Fatalf("override estimate = %+v; want explicit one-hour rate without fallback", got)
+ }
+}
diff --git a/cli/beacon/internal/pricing/cost.go b/cli/beacon/internal/pricing/cost.go
index c6ce4645..02642f15 100644
--- a/cli/beacon/internal/pricing/cost.go
+++ b/cli/beacon/internal/pricing/cost.go
@@ -100,7 +100,8 @@ func Cost(r Rates, t Tokens, requestScoped bool) Estimate {
cacheWrite := set.CacheWrite
if cacheWrite <= 0 {
cacheWrite = set.Input
- est.CacheWriteAtInputRate = t.CacheCreation > 0
+ est.CacheWriteAtInputRate = t.CacheCreation > t.CacheCreation1h ||
+ (t.CacheCreation1h > 0 && set.CacheWrite1h <= 0)
}
cacheWrite1h := set.CacheWrite1h
if cacheWrite1h <= 0 {
diff --git a/cli/beacon/internal/tokens/cache_fallback_test.go b/cli/beacon/internal/tokens/cache_fallback_test.go
new file mode 100644
index 00000000..5d696473
--- /dev/null
+++ b/cli/beacon/internal/tokens/cache_fallback_test.go
@@ -0,0 +1,55 @@
+package tokens
+
+import (
+ "encoding/json"
+ "reflect"
+ "testing"
+
+ "github.com/asymptote-labs/agent-beacon/cli/beacon/internal/pricing"
+)
+
+func TestOneHourOverrideFallbacksInReportJSON(t *testing.T) {
+ for _, tc := range []struct {
+ name, extraRate string
+ oneHour int64
+ wantCost float64
+ wantFallbacks []string
+ }{
+ {"explicit_one_hour", `,"cache_write_1h_usd_per_mtok":5`, 1000, 0.005, nil},
+ {"mixed_ttls", `,"cache_write_1h_usd_per_mtok":5`, 600, 0.0038, []string{"cache_write_at_input_rate"}},
+ {"both_rates_missing", "", 1000, 0.002, []string{"cache_write_1h_at_write_rate", "cache_write_at_input_rate"}},
+ } {
+ t.Run(tc.name, func(t *testing.T) {
+ catalog := pricing.Default()
+ overrides, err := pricing.LoadOverrides([]byte(`{"schema":"beacon.pricing.overrides/v1","models":{"internal-model":{"input_usd_per_mtok":2,"output_usd_per_mtok":3`+tc.extraRate+`}}}`), catalog)
+ if err != nil {
+ t.Fatal(err)
+ }
+ line := `{"timestamp":"2026-10-12T10:00:00Z","event":{"kind":"agent_runtime","action":"token.usage","category":"metric"},"harness":{"name":"claude_code","collection_method":"poll"},"session":{"id":"s1"},"model":"internal-model","gen_ai":{"usage":{"cache_creation":{"input_tokens":1000,"ephemeral_1h_input_tokens":`
+ subset, _ := json.Marshal(tc.oneHour)
+ report := Aggregate(decodeEvents(t, line+string(subset)+`}}}}`), Options{Pricer: pricing.NewPricer(catalog, overrides)})
+ encoded, err := json.Marshal(report)
+ if err != nil {
+ t.Fatal(err)
+ }
+ var wire struct {
+ Pricing struct {
+ Models []struct {
+ Fallbacks []string `json:"fallbacks"`
+ EstimatedCostUSD float64 `json:"estimated_cost_usd"`
+ } `json:"models"`
+ } `json:"pricing"`
+ }
+ if err := json.Unmarshal(encoded, &wire); err != nil {
+ t.Fatal(err)
+ }
+ if len(wire.Pricing.Models) != 1 {
+ t.Fatalf("report JSON = %s", encoded)
+ }
+ got := wire.Pricing.Models[0]
+ if got.EstimatedCostUSD != tc.wantCost || !reflect.DeepEqual(got.Fallbacks, tc.wantFallbacks) {
+ t.Fatalf("report pricing = %+v; want cost %v fallbacks %v", got, tc.wantCost, tc.wantFallbacks)
+ }
+ })
+ }
+}
diff --git a/docs/cli/pricing.mdx b/docs/cli/pricing.mdx
index c280adfe..4d96e4b2 100644
--- a/docs/cli/pricing.mdx
+++ b/docs/cli/pricing.mdx
@@ -112,8 +112,10 @@ confidential: Beacon does not put rates in any event, but the file itself is pla
- **A row replaces the catalog row whole.** Fields you leave out are unpublished, not
inherited from the catalog, and are priced the way the catalog prices a missing rate: cache
reads and cache writes at the row's input rate, one-hour writes at its five-minute write
- rate. Those fallbacks are named in the report's `pricing.models[].fallbacks` and by
- `beacon pricing show`.
+ rate. Reports name only fallbacks actually used for non-zero token counts in
+ `pricing.models[].fallbacks`; one-hour writes at an explicit one-hour rate do not use the
+ five-minute write fallback. `beacon pricing show` also lists unpublished rates and their
+ available fallbacks.
- **Bands** apply to every token of a single request whose prompt (input plus cache read plus
cache write) is above `above_tokens`, exactly as the catalog's bands do, and only to usage
Beacon knows is one request (see [How the estimate is computed](/cli/token-usage#how-the-estimate-is-computed)).