We’ve added full-stack cost tracking to LangSmith to make it easier to understand and monitor spend across complex agent applications.

LangSmith now automatically records token usage and derived costs for major model providers (OpenAI, Anthropic, and others with OpenAI-compatible responses). You can also submit custom cost data for any run type — including tools, retrieval steps, or non-linear pricing models — giving you a complete picture of where your compute budget is going.

What’s new

• Automatic cost derivation for LLM calls based on token counts and model pricing tables

• Provider-aware pricing, including multimodal token types and cache reads

• Manual cost submission for any run type (LLMs, tools, retrieval, custom operations)

• Token & cost breakdowns visible throughout the LangSmith UI

  • In the trace tree

  • In project stats

  • In dashboards

• Model price map editor for adding custom models or overriding default pricing

How it works

Costs are computed in one of two ways:

1. Automatically — when token counts, provider, and model name are present

2. Manually — by submitting usage_metadata with custom cost fields

LangSmith includes pricing data for most OpenAI, Anthropic, and Gemini models out of the box. For other providers or custom pricing schemes, you can supply your own token counts and price mapping.

Why it matters

Building agents at scale introduces non-trivial usage-based costs across multiple components. LangSmith gives you a single, consistent view of spend across prompts, outputs, tools, and retrieval — enabling better monitoring, debugging, and optimization.

See the docs: https://docs.langchain.com/langsmith/cost-tracking