Skip to main content

​
How token-based costing works

For every LLM span, LangSight calculates cost as: 
cost = (input_tokens / 1,000,000 × input_price_per_million)
     + (output_tokens / 1,000,000 × output_price_per_million)
Token counts and model ID come from span attributes. If you use the LangSight SDK directly, set them on the span: 
await langsight.send_span(ToolCallSpan.record(
    server_name="openai-mcp",
    tool_name="chat_completion",
    started_at=started,
    status=ToolCallStatus.SUCCESS,
    input_tokens=512,
    output_tokens=128,
    model_id="gpt-4o",
))
If you send OTLP traces, LangSight extracts these automatically from standard semantic convention attributes:
OTLP attributeMaps to
gen_ai.usage.input_tokensinput_tokens
gen_ai.usage.output_tokensoutput_tokens
gen_ai.request.modelmodel_id
Most agent frameworks (LangChain, LlamaIndex, PydanticAI) already emit these — no extra instrumentation needed.

​
Pre-seeded models

LangSight ships with pricing for 21 models. Prices reflect rates as of early 2026 and can be updated at any time.
ProviderModels
Anthropicclaude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5
OpenAIgpt-4o, gpt-4o-mini, o3, o3-mini, o1
Googlegemini-1.5-pro, gemini-1.5-flash, gemini-2.0-flash, gemini-2.5-pro
AWS Bedrocknova-pro, nova-lite, nova-micro
Metallama-3.1-70b, llama-3.3-70b, llama-3.1-8b
Model prices change frequently. Check Settings → Model Pricing to confirm prices match your current contract before using cost data for chargebacks or budgeting.

​
Updating prices

Settings → Model Pricing → find the model → Edit → update input_price_per_million and/or output_price_per_million → Save. Via API (admin only): 
PATCH /api/costs/models/gpt-4o
Content-Type: application/json
X-API-Key: ls_admin_key

{
  "input_price_per_million": 2.50,
  "output_price_per_million": 10.00
}

​
Adding custom models

For models not in the pre-seeded list (fine-tuned models, private deployments): Settings → Model Pricing → Add Model → fill in model ID and prices. Via API: 
POST /api/costs/models
Content-Type: application/json
X-API-Key: ls_admin_key

{
  "model_id": "my-fine-tuned-gpt4o",
  "provider": "openai",
  "input_price_per_million": 5.00,
  "output_price_per_million": 15.00
}
Spans with an unrecognised model_id appear in cost reports with $0.00 cost and a warning flag until pricing is added.

​
All sources filter

The All sources filter (previously labelled “All servers”) on the Cost Analytics page shows costs grouped by server_name from tool call traces. This field includes:
  • LLM providers — e.g. gemini, claude, gpt-4o (costs calculated from token counts)
  • Sub-agents — e.g. catalog, analyst (agents acting as tool providers in multi-agent pipelines)
  • Any other service tracked by the LangSight SDK
It does not show MCP infrastructure server names like catalog-mcp — those are health-checked by langsight monitor but are not cost sources themselves. The cost source is the service that consumes tokens or incurs per-call charges, not the MCP proxy in front of it. Filter by a specific source to see cost per LLM provider or per sub-agent: 
GET /api/costs/breakdown?project_id=production&window=7d&server_name=gemini

​
Per-project cost filtering

All cost queries are scoped to a project. Use the project switcher in the sidebar to filter the Costs page, or pass project_id to the API: 
GET /api/costs/breakdown?project_id=production&window=7d

​
Non-LLM tool call costs

Tool calls to MCP servers that are not LLM providers use fixed cost_per_call pricing rules from .langsight.yaml: 
costs:
  rules:
    - server: jira-mcp
      tool: "*"
      cost_per_call: 0.001
    - server: "*"
      tool: "*"
      cost_per_call: 0.0
These show as the “Tool call cost” line in the Costs page, separate from LLM cost.