Cost Tracking

CostTrackingService records token usage and computes estimated cost for every LLM call. Per-model pricing tables for 30+ models. Now cache-aware (as of 2026-07-10) — Anthropic prompt-cache reads are priced separately at their lower rate.

What it tracks

Per call:

• Input tokens — prompt + context sent to the model
• Output tokens — model response
• Cache read — tokens served from the provider’s prompt cache (Anthropic)
• Cache creation — first-time cache writes (Anthropic)
• Estimated cost — 6-decimal precision

Cost is a weighted sum:

cost = uncached_input × input_rate
     + cache_read     × cache_read_rate      (Anthropic)
     + cache_creation × cache_creation_rate  (Anthropic)
     + output         × output_rate

All rates per million tokens.

Viewing costs

• Per message — token count under each response
• Per conversation — cumulative cost for the current thread
• Aggregate summary — breakdown by provider + model

Spend limits

{
  "codebuddy.agent.costThreshold": 1000,
  "codebuddy.agent.costThresholdWarningPercent": 80,
  "codebuddy.agent.dailySpendCap": null
}

• codebuddy.agent.costThreshold — maximum spend (USD) per conversation before CodeBuddy pauses and asks whether to continue. Default 1000.
• codebuddy.agent.costThresholdWarningPercent — percentage of the threshold at which a warning is shown (1–99). Default 80.
• codebuddy.agent.dailySpendCap — maximum total daily spend (USD) across all conversations. null disables it.

Pricing (approximate, mid-2025)

Per million tokens.

Provider	Model	Input	Output
Anthropic	Claude Sonnet 4	$3.00	$15.00
Anthropic	Claude Opus 4	$15.00	$75.00
Anthropic	Claude Haiku	$0.25	$1.25
OpenAI	GPT-4o	$2.50	$10.00
OpenAI	GPT-4o-mini	$0.15	$0.60
OpenAI	o3-mini	$1.10	$4.40
Google	Gemini 2.5 Pro	$1.25	$10.00
Google	Gemini 2.5 Flash	$0.15	$0.60
Groq	Llama 3.3 70B	$0.59	$0.79
DeepSeek	DeepSeek Chat	$0.27	$1.10
DeepSeek	DeepSeek Reasoner	$0.55	$2.19
Qwen	Qwen Plus	$0.80	$2.00
xAI	Grok	$5.00	$15.00
Ollama	Local models	Free	Free

When a model isn’t in the pricing table, conservative fallback: $3.00 / $15.00 per million.

Anthropic cache rates live in the same table under separate columns (cacheReadPerMillion, cacheCreationPerMillion) — see cost-tracking.service.ts for exact numbers.

Cost summary shape

interface ICostSummary {
  totals: {
    inputTokens: number;
    outputTokens: number;
    cacheReadTokens?: number;
    cacheCreationTokens?: number;
    estimatedCost: number;
    requestCount: number;
  };
  providers: Array<{
    provider: string;
    model: string;
    inputTokens: number;
    outputTokens: number;
    cacheReadTokens?: number;
    estimatedCost: number;
    requestCount: number;
  }>;
  conversations: Array<{
    threadId: string;
    provider: string;
    model: string;
    inputTokens: number;
    outputTokens: number;
    estimatedCost: number;
    requestCount: number;
  }>;
}

Reducing spend

• Smaller models for simple tasks — GPT-4o-mini or Gemini Flash for renames, formatting, trivial edits.
• Local models — Ollama for tasks that don’t need frontier intelligence.
• Project rules — clear rules reduce iteration.
• codebuddy.agent.costThreshold / dailySpendCap — per-conversation and daily USD safety nets.
• Anthropic prompt caching — already active; hit rate ~99% on turn 2+. Long conversations become cheap.
• Start fresh conversations for unrelated tasks — old context doesn’t help and costs tokens.

Related

• Providers — provider selection + failover chain
• Caching §1 — Anthropic prompt caching detail
• Telemetry — token counts persisted as OTel span attributes