descrybe/apps/api/internal/aiaudit/cost_test.go

package aiaudit

import "testing"

// Luna list price, 2026-07-30: $0.20 / $0.02 cached / $1.20 per 1M tokens.
func TestCostMicros_lunaListPrice(t *testing.T) {
	t.Parallel()
	luna, ok := DefaultModelPrice("gpt-5.6-luna")
	if !ok {
		t.Fatal("gpt-5.6-luna must have a shipped price")
	}
	if luna.InputPerMTok != 200_000 || luna.CachedInputPerMTok != 20_000 || luna.OutputPerMTok != 1_200_000 {
		t.Fatalf("unexpected Luna price: %+v", luna)
	}

	// 1M input + 1M output = $0.20 + $1.20 = $1.40 = 1_400_000 micros.
	if got := CostMicros(luna, 1_000_000, 0, 1_000_000); got != 1_400_000 {
		t.Fatalf("got %d micros want 1400000", got)
	}
	// A realistic enhance call: 4k prompt, 800 completion.
	// 4000*0.2 + 800*1.2 per 1M = 800 + 960 = 1760 micros ($0.00176).
	if got := CostMicros(luna, 4000, 0, 800); got != 1760 {
		t.Fatalf("got %d micros want 1760", got)
	}
}

// Cached tokens are a SUBSET of prompt tokens, not extra ones. Billing them on top
// of full input would overstate every repeated system prompt.
func TestCostMicros_cachedTokensAreDiscountedNotAdded(t *testing.T) {
	t.Parallel()
	luna, _ := DefaultModelPrice("gpt-5.6-luna")

	full := CostMicros(luna, 1_000_000, 0, 0)              // 1M fresh input  = 200000
	allCached := CostMicros(luna, 1_000_000, 1_000_000, 0) // 1M cached  =  20000
	if full != 200_000 {
		t.Fatalf("fresh input = %d want 200000", full)
	}
	if allCached != 20_000 {
		t.Fatalf("fully cached input = %d want 20000", allCached)
	}
	if allCached >= full {
		t.Fatal("cached input must be cheaper than fresh input")
	}
	// Half cached: 500k*0.2 + 500k*0.02 per 1M = 100000 + 10000.
	if got := CostMicros(luna, 1_000_000, 500_000, 0); got != 110_000 {
		t.Fatalf("half cached = %d want 110000", got)
	}
	// A provider reporting more cached than prompt tokens must not go negative.
	if got := CostMicros(luna, 100, 900, 0); got < 0 {
		t.Fatalf("cached > prompt produced negative cost %d", got)
	}
}

// Small calls must not round to zero, or per-call spend vanishes at scale.
func TestCostMicros_smallCallsStillCost(t *testing.T) {
	t.Parallel()
	luna, _ := DefaultModelPrice("gpt-5.6-luna")
	if got := CostMicros(luna, 500, 0, 100); got <= 0 {
		t.Fatalf("small call cost %d, expected > 0", got)
	}
}

func TestCostMicros_unknownModelIsFree(t *testing.T) {
	t.Parallel()
	if _, ok := DefaultModelPrice("some-local-model"); ok {
		t.Fatal("unknown model should not have a shipped price")
	}
	// Zero price rather than a guessed rate: better to under-report than invent spend.
	if got := CostMicros(ModelPrice{}, 1_000_000, 0, 1_000_000); got != 0 {
		t.Fatalf("unpriced model cost %d want 0", got)
	}
}

func TestNormalizeModel(t *testing.T) {
	t.Parallel()
	cases := map[string]string{
		"gpt-5.6-luna":             "gpt-5.6-luna",
		"GPT-5.6-Luna":             "gpt-5.6-luna",
		"openai/gpt-5.6-luna":      "gpt-5.6-luna",
		"gpt-5.6-luna-2026-07-30":  "gpt-5.6-luna",
		"  openai/GPT-5.6-Terra  ": "gpt-5.6-terra",
		"mock-llm":                 "mock-llm",
	}
	for in, want := range cases {
		if got := NormalizeModel(in); got != want {
			t.Fatalf("NormalizeModel(%q) = %q want %q", in, got, want)
		}
	}
	// A dated snapshot must still price like its base model.
	if _, ok := DefaultModelPrice("gpt-5.6-luna-2026-07-30"); !ok {
		t.Fatal("dated snapshot should resolve to the base model price")
	}
}

// Three years, so quarter- and year-scale cost questions still have data.
func TestUsageRetentionCoversThreeYears(t *testing.T) {
	t.Parallel()
	if UsageRetentionDays < 3*365 {
		t.Fatalf("usage retention %d days is under 3 years", UsageRetentionDays)
	}
	if RetentionDays >= UsageRetentionDays {
		t.Fatal("prompt bodies must expire long before the cost rollup")
	}
}