Cost Management

Reading3 min

Updated5 Aug 2026

Sourcev1/ai-sdk/cost-management.mdx

The CostTracker monitors token usage and costs across all LLM operations, enforces budgets with alerts, and provides optimization recommendations.

Creating a Cost Tracker01

import sdk "github.com/xraph/ai-sdk"

tracker := sdk.NewCostTracker(logger, metrics, &sdk.CostTrackerOptions{
    RetentionPeriod: 30 * 24 * time.Hour, // Keep 30 days of records
    EnableCache:     true,
})

Recording Usage02

tracker.RecordUsage(ctx, sdk.UsageRecord{
    Timestamp:    time.Now(),
    Provider:     "openai",
    Model:        "gpt-4",
    Operation:    "chat",
    InputTokens:  150,
    OutputTokens: 300,
    TotalTokens:  450,
    Cost:         0.0195,
    CacheHit:     false,
    Metadata:     map[string]any{"agent": "research-assistant"},
})

Budget Management03

Set spending limits with alert thresholds:

tracker.SetBudget("daily", 10.0, 24*time.Hour, 0.8)
// Name: "daily"
// Limit: $10.00 per period
// Period: 24 hours
// AlertAt: 80% ($8.00)

tracker.SetBudget("monthly", 300.0, 30*24*time.Hour, 0.9)

Cost Insights04

Get analytics on spending:

insights := tracker.GetInsights()

fmt.Printf("Total cost: $%.2f\n", insights.TotalCost)
fmt.Printf("Total tokens: %d\n", insights.TotalTokens)
fmt.Printf("Average cost per request: $%.4f\n", insights.AverageCostPerRequest)
fmt.Printf("Most expensive model: %s\n", insights.MostExpensiveModel)
fmt.Printf("Cache hit rate: %.1f%%\n", insights.CacheHitRate*100)

Optimization Recommendations05

recommendations := tracker.GetOptimizationRecommendations()

for _, rec := range recommendations {
    fmt.Printf("[%s] %s\n", rec.Priority, rec.Description)
    fmt.Printf("  Potential savings: $%.2f/month\n", rec.EstimatedSavings)
}

Recommendations may include:

Switching to a cheaper model for simple tasks
Enabling caching for repeated queries
Reducing token usage with shorter prompts
Using batch operations

Model Pricing06

The SDK includes default pricing for common models:

var DefaultModelPricing = map[string]ModelPricing{
    "openai/gpt-4":          {InputPer1KTokens: 0.03, OutputPer1KTokens: 0.06},
    "openai/gpt-4-turbo":    {InputPer1KTokens: 0.01, OutputPer1KTokens: 0.03},
    "openai/gpt-3.5-turbo":  {InputPer1KTokens: 0.0005, OutputPer1KTokens: 0.0015},
    "anthropic/claude-3-opus": {InputPer1KTokens: 0.015, OutputPer1KTokens: 0.075},
    // ...
}

UsageRecord Fields07

Field	Type	Description
`Timestamp`	`time.Time`	When the usage occurred
`Provider`	`string`	LLM provider name
`Model`	`string`	Model used
`Operation`	`string`	Operation type (`chat`, `completion`, `embedding`)
`InputTokens`	`int`	Tokens in the prompt
`OutputTokens`	`int`	Tokens in the response
`TotalTokens`	`int`	Total tokens consumed
`Cost`	`float64`	Computed cost in USD
`CacheHit`	`bool`	Whether the result was cached
`Metadata`	`map[string]any`	Custom metadata

← Previous

Guardrails

Resilience