Cost Tracking

Every AI call costs tokens. ThinkLang tracks usage automatically so you can monitor and control costs.

Show Cost After Execution

Add the --show-cost flag to see a cost summary after running a program:

thinklang run app.tl --show-cost

Output:

--- Cost Summary ---
Total calls: 5
Total tokens: 3200 input, 1800 output
Total cost: $0.063000

By operation:
  think: 3 calls, 3500 tokens, $0.045000
  infer: 2 calls, 1500 tokens, $0.018000

By model:
  claude-opus-4-6: 5 calls, $0.063000

Cost Report Command

View the cost report for the current session:

thinklang cost-report

This shows the same summary as --show-cost but can be run independently.

What Gets Tracked

Every think, infer, reason, and assert.semantic call is recorded with:

Field	Description
operation	Which primitive was called (think, infer, reason, semantic_assert)
model	The model used (e.g., claude-opus-4-6)
inputTokens	Number of input tokens consumed
outputTokens	Number of output tokens generated
costUsd	Estimated cost in USD
durationMs	How long the call took
prompt	First 100 characters of the prompt (for identification)

Cost Summary

The cost summary aggregates records into:

• Total input tokens, output tokens, cost, and call count
• By operation -- breakdown by think, infer, reason, semantic_assert
• By model -- breakdown by which model handled each call

Model Pricing

Costs are calculated from per-million-token pricing:

Model	Input (per 1M tokens)	Output (per 1M tokens)
claude-opus-4-6	$15.00	$75.00
claude-sonnet-4-5	$3.00	$15.00
claude-haiku-4-5	$0.80	$4.00

Reducing Costs

• Enable caching. Set THINKLANG_CACHE=true (the default). Identical requests are served from cache at zero cost.
• Use infer for simple tasks. infer uses shorter prompts than think, reducing token usage.
• Keep context small. Only include relevant data with with context. Large context means more input tokens.
• Use without context for exclusions. Strip unnecessary data before it reaches the AI.
• Choose the right model. Set THINKLANG_MODEL to a smaller model for tasks that do not require the most capable one.
• Use replay in tests. Run thinklang test --replay to avoid live API calls during test iteration.