Cost tracking

Opik has been designed to track and monitor costs for your LLM applications by measuring token usage across all traces. Using the Opik dashboard, you can analyze spending patterns and quickly identify cost anomalies. All costs across Opik are estimated and displayed in USD.

You can use the Opik dashboard to review costs at three levels: spans, traces, and projects. Each level provides different insights into your application's cost structure.

Individual spans show the computed costs (in USD) for each LLM spans of your traces:

If you are using one of Opik's integrations, we automatically aggregates costs from all spans within a trace to compute total trace costs:

Track your overall project costs in:

1. The main project view, through the Estimated Cost column:

2. The project Metrics tab, which shows cost trends over time:

You can retrieve the estimated cost programmatically for both spans and traces. Note that the cost will be None if the span or trace used an unsupported model.

import opik

client = opik.Opik()

span = client.get_span_content("<SPAN_ID>")
# Returns estimated cost in USD, or None for unsupported models
print(span.total_estimated_cost)

import opik

client = opik.Opik()

trace = client.get_trace_content("<TRACE_ID>")
# Returns estimated cost in USD, or None for unsupported models
print(trace.total_estimated_cost)

If you are not using one of Opik's integration, Opik can still compute the cost. For you will need to ensure the span type is llm and you will need to pass:

1. provider: The name of the provider, typically openai, anthropic or google_ai for example (the most recent providers list can be found in opik.LLMProvider enum object)
2. model: The name of the model
3. usage: The input, output and total tokens for this LLM call.

You can then update your code to log traces and spans:

from opik import track, opik_context

@track(type="llm") # Note - Specifying the type is this is important
def llm_call(input):
  opik_context.update_current_span(
    provider="openai",
    model="gpt-3.5-turbo",
    usage={
      "prompt_tokens": 4,
      "completion_tokens": 6,
      "total_tokens": 10
    }
  )
  return "Hello, world!"

llm_call("Hello world!")

import opik

client = opik.Opik()

trace = client.trace(
  name="custom_trace",
  input={"text": "Hello world!"},
)

# Logging the LLM call
span = trace.span(
  name="llm_call",
  type="llm",
  input={"text": "Hello world!"},
  output={"response": "Hello world!"},
  provider="openai",
  model="gpt-3.5-turbo",
  usage={
    "prompt_tokens": 4,
    "completion_tokens": 6,
    "total_tokens": 10
  }
)

When you need to set a custom cost or use an unsupported model, you can manually set the cost of a span. There are two approaches depending on your use case:

If you're manually creating spans, you can set the cost directly when creating the span:

from opik import track, opik_context

@track
def llm_call(input):
  opik_context.update_current_span(
    total_cost=0.05,
  )
  return "Hello, world!"

llm_call("Hello world!")

With Opik integrations, spans are automatically created and closed, preventing updates while they're open. However, you can update the cost afterward using the update_span method. This works well for implementing periodic cost estimation jobs:

from opik import Opik
from opik.rest_api.types.span_public import SpanPublic

# Define your own cost mapping for different models
TOKEN_COST = {
    ("openai.chat", "gpt-4o-2024-08-06"): {
        "input_tokens": 2.5e-06,
        "output_tokens": 1e-05,
    }
}

# This part would be custom for your use-case and is only here for example
def compute_cost_for_span(span: SpanPublic):
    provider = span.provider or span.input.get("ai.model.provider")
    model = span.model or span.output.get("gen_ai.response.model")
    usage = span.usage

    if (provider, model) in TOKEN_COST:
        model_cost = TOKEN_COST[(provider, model)]
        cost = (
            usage["input_tokens"] * model_cost["input_tokens"]
            + usage["output_tokens"] * model_cost["output_tokens"]
        )
        return cost
    return None

def update_span_costs(project_name, trace_id=None):
    opik_client = Opik()

    # Find LLM spans that don't have estimated costs
    spans = opik_client.search_spans(
        project_name=project_name,
        trace_id=trace_id,
        filter_string='type="llm" and total_estimated_cost=0',
    )

    for span in spans:
        cost = compute_cost_for_span(span)

        if cost:
            print(f"Updating span {span.id} of trace {span.trace_id} with cost: {cost}")
            opik_client.update_span(
                trace_id=span.trace_id,
                parent_span_id=span.parent_span_id,
                project_name=project_name,
                id=span.id,
                total_cost=cost,
            )

# Example usage in a CRON job
if __name__ == "__main__":
    update_span_costs("your-project-name")

This approach is particularly useful when:

• Using models or providers not yet supported by automatic cost tracking
• You have a custom pricing agreement with your provider
• You want to track additional costs beyond model usage
• You need to implement cost estimation as a background process
• Working with integrations where spans are automatically managed

Opik currently calculates costs automatically for all LLM calls in the following Python SDK integrations:

• Google ADK Integration
• AWS Bedrock Integration
• LangChain Integration
• OpenAI Integration
• LiteLLM Integration
• Anthropic Integration
• CrewAI Integration
• Google AI Integration
• Haystack Integration
• LlamaIndex Integration

Cost tracking is supported for the following LLM providers (pass the name in parentheses as the span provider; the most common ones are also available in the opik.LLMProvider enum):

• OpenAI (openai) - Models hosted by OpenAI (https://platform.openai.com)
• Anthropic (anthropic) - Models hosted by Anthropic (https://www.anthropic.com)
• Anthropic on Vertex AI (anthropic_vertexai) - Anthropic models hosted by Google Vertex AI
• Google AI (google_ai) - Gemini models hosted in Google AI Studio (https://ai.google.dev/aistudio)
• Google Vertex AI (google_vertexai) - Gemini models hosted in Google Vertex AI (https://cloud.google.com/vertex-ai)
• AWS Bedrock (bedrock) - Models hosted by AWS Bedrock (https://aws.amazon.com/bedrock)
• Groq (groq) - Models hosted by Groq (https://groq.com)
• TypeSafe AI (typesafe) - Jev System One models hosted by TypeSafe AI (https://typesafe.ai)

You can find a complete list of supported models for these providers in the model_prices_and_context_window.json file.