Metrics

Token usage, timing, and per-model breakdown for runs, messages, sessions, and tool calls.

Class	Used On	Purpose
RunMetrics	RunOutput.metrics, TeamRunOutput.metrics	Run-level totals with per-model breakdown
MessageMetrics	Message.metrics	Per-API-call tokens and timing
SessionMetrics	agent.get_session_metrics()	Aggregated across all runs in a session
ModelMetrics	Inside RunMetrics.details	Per-model aggregate by (provider, id)
ToolCallMetrics	ToolExecution.metrics	Tool execution timing

All metric classes except ToolCallMetrics inherit from BaseMetrics, which provides the shared token and cost fields.

BaseMetrics

Shared token fields inherited by RunMetrics, MessageMetrics, SessionMetrics, and ModelMetrics.

Parameter	Type	Default	Description
input_tokens	int	0	Tokens in the prompt/input.
output_tokens	int	0	Tokens generated by the model.
total_tokens	int	0	Total tokens used (input + output).
audio_input_tokens	int	0	Audio tokens in the input.
audio_output_tokens	int	0	Audio tokens in the output.
audio_total_tokens	int	0	Total audio tokens.
cache_read_tokens	int	0	Tokens served from cache.
cache_write_tokens	int	0	Tokens written to cache.
reasoning_tokens	int	0	Tokens used for reasoning steps.
cost	Optional[float]	None	Cost of the API call(s).

RunMetrics

Used on RunOutput.metrics and TeamRunOutput.metrics. Includes all BaseMetrics fields plus:

Parameter	Type	Default	Description
timer	Optional[Timer]	None	Internal timer that computes duration and time_to_first_token. Excluded from to_dict().
time_to_first_token	Optional[float]	None	Time from run start to first token (seconds). Set once per run.
duration	Optional[float]	None	Total run time (seconds).
details	Optional[Dict[str, List[ModelMetrics]]]	None	Per-model breakdown keyed by model type (e.g., "model", "output_model", "memory_model").
additional_metrics	Optional[Dict[str, Any]]	None	Extra metrics (e.g., eval_duration).

details keys

The details dictionary uses model type strings as keys. Each key maps to a list of ModelMetrics objects (one per unique model (provider, id) pair).

Key	Description
"model"	Primary agent model
"output_model"	Output model for structured output
"parser_model"	Parser model
"memory_model"	Memory model
"reasoning_model"	Reasoning model
"session_summary_model"	Session summary model
"culture_model"	Culture model
"learning_model"	Learning model
"compression_model"	Compression model
"followup_model"	Followup model

Eval agent metrics are prefixed with eval_. If an eval agent uses model types "model" and "output_model", the details keys become "eval_model" and "eval_output_model".

Metrics is preserved as a backward-compatible alias for RunMetrics. Both from agno.metrics import Metrics and from agno.models.metrics import Metrics still work.

MessageMetrics

Used on Message.metrics. Per-API-call metrics. Includes all BaseMetrics fields plus:

Parameter	Type	Default	Description
timer	Optional[Timer]	None	Internal timer that computes duration and time_to_first_token. Excluded from to_dict().
duration	Optional[float]	None	Duration of this API call (seconds).
time_to_first_token	Optional[float]	None	Time to first token for this API call (seconds).
provider_metrics	Optional[Dict[str, Any]]	None	Provider-specific metrics (e.g., Ollama timing, Cerebras timing).

SessionMetrics

Returned by agent.get_session_metrics(). Aggregated across all runs in a session. Includes all BaseMetrics fields plus:

Parameter	Type	Default	Description
details	Optional[Dict[str, List[ModelMetrics]]]	None	Per-model breakdown, same structure as RunMetrics.details. Tokens summed across runs.
additional_metrics	Optional[Dict[str, Any]]	None	Aggregated additional metrics from all runs.

ModelMetrics

Per-model aggregate stored inside RunMetrics.details[model_type]. Includes all BaseMetrics fields plus:

Parameter	Type	Default	Description
id	str	""	Model ID (e.g., "gpt-4o").
provider	str	""	Provider name (e.g., "OpenAI Chat", "OpenAI Responses", "Anthropic").
provider_metrics	Optional[Dict[str, Any]]	None	Provider-specific data for this model.

ToolCallMetrics

Used on ToolExecution.metrics. Standalone class (does not inherit BaseMetrics).

Parameter	Type	Default	Description
timer	Optional[Timer]	None	Internal timer that computes duration. Excluded from to_dict().
start_time	Optional[float]	None	Unix timestamp when the tool call started.
end_time	Optional[float]	None	Unix timestamp when the tool call ended.
duration	Optional[float]	None	Tool execution time (seconds).

Provider field availability

Not all providers populate all fields. This table shows which BaseMetrics fields each provider sets on MessageMetrics.

Provider	Standard tokens	cache_read	cache_write	reasoning	Audio tokens	cost	provider_metrics
OpenAI Chat	✅	✅		✅	✅	✅
OpenAI Responses	✅	✅		✅
Anthropic Claude	✅	✅	✅				server_tool_use, service_tier
Google Gemini	✅	✅		✅			traffic_type
Groq	✅						completion_time, prompt_time, queue_time, total_time
Mistral	✅	✅		✅
AWS Bedrock	✅	✅	✅
LiteLLM	✅	✅		✅	✅
Perplexity	✅	✅		✅	✅
Azure AI Foundry	✅	✅		✅
Cerebras	✅						time_system, time_prompt
Ollama	✅						total_duration, load_duration, prompt_eval_duration, eval_duration
Cohere	✅
Meta Llama	✅
Meta Llama (OpenAI)	✅	✅		✅	✅	✅
HuggingFace	✅
IBM WatsonX	✅

“Standard tokens” = input_tokens, output_tokens, total_tokens. “Audio tokens” = audio_input_tokens, audio_output_tokens.