AI Observability metrics reference

Latest Dynatrace
Reference
8-min read
Published Jul 02, 2026

This page lists every standard metric and span attribute that Dynatrace can ingest for GenAI workloads, grouped by where the data comes from.

OpenTelemetry-native metrics: First-class OTel metric instruments (gen_ai.client.*, gen_ai.server.*, gen_ai.workflow.*, gen_ai.invoke_agent.*, gen_ai.execute_tool.*) emitted directly by your instrumentation (Traceloop SDK, OneAgent, or other OTel/OpenLLMetry-based SDKs). Dynatrace ingests these as-is, and doesn't calculate them.
Dynatrace span-derived visualizations: Built by Dynatrace directly from gen_ai.* span attributes. No separate metric instrument is emitted for these; AI Observability computes tiles, filters, and breakdowns straight from the span data.
Vendor-native metrics: These are sent directly by the AI platform or gateway (Amazon Bedrock, Azure OpenAI, NVIDIA NIM, Kong AI Gateway), either as provider-specific span attributes or as their own metrics.
Other OTel or third-party sources: These are metrics that you derive yourself (for example, by parsing inference-server or application logs) and then forward to Dynatrace.

OpenTelemetry-native metrics

All signals are ingestible and queryable via DQL. Some signals are visualized in tiles, as noted in the tables.

OpenTelemetry's GenAI semantic conventions call every property (as documented in this section) an "attribute," including the ones attached to metric data points. Once ingested into Dynatrace, properties on metrics become dimensions; only span, log, and resource properties keep the "attribute" name. The tables on this page use a single Attribute column for both, since most properties here start life as span attributes. In cases where a property is also usable as a metric dimension, this is called out directly in its row (for example, in the Type or Description cell).

GenAI client metrics (OpenTelemetry)

These are the standard OpenTelemetry semantic-convention metrics for GenAI clients. They are metric instruments, not span attributes; your instrumentation must be configured to emit them. Possible configuration options include, for example, Traceloop's should_enrich_metrics=True or a custom OTel MeterProvider. If a metric that already has a tile is missing, that tile shows empty or zeroed values without an error.

The Requirement level column in the per-metric attribute tables below is stricter than the upstream OTel semantic conventions. gen_ai.request.model and gen_ai.response.model are marked Required here because AI Observability AI Observability needs both to populate the model filter, prompts table, and overview charts. OTel semconv classifies them as "conditionally required" and "recommended" respectively.

Metric	Instrument type	Unit	Description	Visualization
`gen_ai.client.token.usage`	Histogram	`{token}`	Number of input and output tokens used.	Shows cost tiles, with token usage split by type
`gen_ai.client.operation.duration`	Histogram	`s`	GenAI operation duration.	Shows latency charts (p99, mean)
`gen_ai.client.operation.time_to_first_chunk`	Histogram	`s`	Time to receive the first chunk, measured from when the client issues the generation request to when the first chunk is received in the response stream. Only emitted for streaming calls.	Visualization not available
`gen_ai.client.operation.time_per_output_chunk`	Histogram	`s`	Time per output chunk, recorded for each chunk received after the first one, measured from the end of the previous chunk to the end of the current chunk. Only emitted for streaming calls.	Visualization not available

Metric: `gen_ai.client.token.usage`

Attribute	Requirement level	Value type	Description	Example values
`gen_ai.operation.name`	Required	string	The name of the operation being performed.	`chat`, `generate_content`, `text_completion`
`gen_ai.provider.name`	Required	string	The GenAI provider identified by the instrumentation.	`openai`, `aws.bedrock`, `azure.ai.openai`
`gen_ai.token.type`	Required	string	The type of token being counted. This splits the cost tile lanes.	`input`, `output`
`gen_ai.request.model`	Required	string	The model a request is being made to.	`gpt-5`
`gen_ai.response.model`	Required	string	The model that generated the response.	`gpt-5-0616`

If this metric is missing, all cost tiles show $0 with no visible error.

If the gen_ai.token.type attribute is missing, the tiles show an undifferentiated cost total instead of separate input/output lanes.

Metric: `gen_ai.client.operation.duration`

Attribute	Requirement level	Value type	Description	Example values
`gen_ai.operation.name`	Required	string	The name of the operation being performed.	`chat`, `generate_content`, `text_completion`
`error.type`	Conditionally required	string	Error class the operation ended with, if it errored.	`timeout`, `500`
`gen_ai.provider.name`	Conditionally required	string	The GenAI provider identified by the instrumentation.	`openai`, `aws.bedrock`
`gen_ai.request.model`	Required	string	The model a request is being made to.	`gpt-5`
`gen_ai.response.model`	Required	string	The model that generated the response.	`gpt-5-0616`

If this metric is missing, all latency tiles show as empty, with no error surfaced.

Metric: `gen_ai.client.operation.time_to_first_chunk`

Attribute	Requirement level	Value type	Description	Example values
`gen_ai.operation.name`	Required	string	The name of the operation being performed.	`chat`, `generate_content`, `text_completion`
`gen_ai.provider.name`	Required	string	The GenAI provider identified by the instrumentation.	`openai`, `gcp.gen_ai`, `gcp.vertex_ai`
`gen_ai.request.model`	Required	string	The model a request is being made to, if available.	`gpt-5`
`server.port`	Conditionally required	int	GenAI server port, if `server.address` is set.	`80`, `8080`, `443`
`gen_ai.response.model`	Required	string	The model that generated the response.	`gpt-5-0616`
`server.address`	Recommended	string	GenAI server address.	`example.com`

Metric: `gen_ai.client.operation.time_per_output_chunk`

Attribute	Requirement level	Value type	Description	Example values
`gen_ai.operation.name`	Required	string	The name of the operation being performed.	`chat`, `generate_content`, `text_completion`
`gen_ai.provider.name`	Required	string	The GenAI provider identified by the instrumentation.	`openai`, `gcp.gen_ai`, `gcp.vertex_ai`
`gen_ai.request.model`	Required	string	The model a request is being made to, if available.	`gpt-5`
`server.port`	Conditionally required	int	GenAI server port, if `server.address` is set.	`80`, `8080`, `443`
`gen_ai.response.model`	Required	string	The model that generated the response.	`gpt-5-0616`
`server.address`	Recommended	string	GenAI server address.	`example.com`

GenAI model server metrics (OpenTelemetry)

These metrics describe the operational behavior of GenAI model servers (functional and performance metrics), rather than the client's view of an operation. They're most relevant for self-hosted or gateway-fronted model servers.

Metric	Instrument type	Unit	Description	Visualization
`gen_ai.server.request.duration`	Histogram	`s`	Generative AI server request duration, such as time-to-last-byte or last output token.	Visualization not available
`gen_ai.server.time_to_first_token`	Histogram	`s`	Time to generate the first token of the response, for streaming requests.	Visualization not available
`gen_ai.server.time_per_output_token`	Histogram	`s`	Time per output token generated after the first token, for successful responses.	Visualization not available

Common attributes

All three Gen AI model server metrics share this attribute set:

Attribute	Requirement level	Value type	Description	Example values
`gen_ai.operation.name`	Required	string	The name of the operation being performed.	`chat`, `generate_content`, `text_completion`
`gen_ai.provider.name`	Required	string	The GenAI provider identified by the client or server instrumentation.	`openai`, `gcp.gen_ai`, `gcp.vertex_ai`
`error.type`	Conditionally required	string	Error class the operation ended with, if it errored. Applies to `gen_ai.server.request.duration` only.	`timeout`, `500`
`gen_ai.request.model`	Required	string	The model a request is being made to, if available.	`gpt-5`
`server.port`	Conditionally required	int	GenAI server port, if `server.address` is set.	`80`, `8080`, `443`
`gen_ai.response.model`	Required	string	The model that generated the response.	`gpt-5-0616`
`server.address`	Recommended	string	GenAI server address.	`example.com`

GenAI workflow, agent, and tool metrics (OpenTelemetry)

These metrics measure orchestration boundaries above a single model call: a full workflow, a single agent invocation, or a single tool execution. When instrumentation can only observe a single provider-facing client call, use gen_ai.client.operation.duration instead.

Metric	Instrument type	Unit	Description	Status
`gen_ai.workflow.duration`	Histogram	`s`	End-to-end duration of a workflow execution, from application code initiating the workflow to the workflow completing.	Visualization not available
`gen_ai.invoke_agent.duration`	Histogram	`s`	End-to-end duration of a single in-process agent invocation, from invocation start until the agent emits its last response chunk or errors.	Visualization not available
`gen_ai.execute_tool.duration`	Histogram	`s`	Duration of a single tool execution.	Visualization not available

Metric: `gen_ai.workflow.duration`

Attribute	Requirement level	Value type	Description	Example values
`error.type`	Conditionally required	string	Error class the operation ended with, if it errored.	`timeout`, `500`
`gen_ai.workflow.name`	Conditionally required	string	Human-readable workflow name provided by the application, if available.	`multi_agent_rag`, `customer_support_pipeline`

Metric: `gen_ai.invoke_agent.duration`

Attribute	Requirement level	Value type	Description	Example values
`error.type`	Conditionally required	string	Error class the operation ended with, if it errored.	`timeout`, `500`
`gen_ai.agent.name`	Conditionally required	string	Human-readable name of the invoked GenAI agent, when available.	`Math Tutor`, `Fiction Writer`
`gen_ai.request.model`	Recommended	string	The model configured for the agent, if applicable.	`gpt-5`

Metric: `gen_ai.execute_tool.duration`

Attribute	Requirement level	Value type	Description	Example values
`gen_ai.tool.name`	Required	string	Name of the tool used by the agent.	`Flights`
`error.type`	Conditionally required	string	Error class the operation ended with, if it errored.	`timeout`, `500`
`gen_ai.agent.name`	Conditionally required	string	Human-readable name of the agent executing the tool, when applicable.	`Math Tutor`, `Fiction Writer`
`gen_ai.tool.type`	Recommended	string	Type of the tool used by the agent, if available.	`function`, `extension`, `datastore`

Dynatrace span-derived visualizations

These attributes are set directly on spans by your instrumentation. Dynatrace computes the following tiles, filters, and breakdowns straight from them; there's no separate metric instrument behind these visualizations.

Core span attributes

AI Observability visualizes the following attributes, which are set directly on spans and exist alongside the general OTel/Dynatrace fields service.name, span.status_code, and duration.

Requirement levels reflect what AI Observability needs to render correctly, which is stricter than the upstream OpenTelemetry semantic conventions for GenAI spans. Attributes marked Required here are mandatory for the app to show data; missing any of them leaves filters, tiles, or table columns empty.

Attribute	Requirement level	Visualization
`gen_ai.provider.name` (or deprecated `gen_ai.system`)	Required	Shows provider filter, service explorer table
`service.name`	General OTel resource attribute, not GenAI-specific.	Shows service filter, service explorer table
`gen_ai.operation.name`	Required	Shows operation breakdown, trace analysis
`gen_ai.request.model`	Required	Shows prompts table, distributed tracing
`gen_ai.response.model`	Required	Shows model filter, overview charts
`gen_ai.usage.input_tokens` / `gen_ai.usage.output_tokens`	Required	Shows prompts table, cost tile fallback when the `gen_ai.client.token.usage` metric isn't emitted
`span.status_code`	General OTel span field, not GenAI-specific.	Shows service health tile (success/failure rate)
`duration`	General OTel span field, not GenAI-specific.	Shows average and p50/p90 latency in the Explorer per-service and per-model tables
`gen_ai.agent.name`	Conditionally required	Shows agent vs. LLM distinction, agent filter
`gen_ai.conversation.id`	Conditionally required	Shows groups spans into conversation threads in the prompts view
`gen_ai.input.messages` / `gen_ai.output.messages`	Required	Shows prompts table content

For agent-specific and provider-specific span attributes, see Additional span attributes and Vendor-native metrics and attributes.

For the full attribute list, required vs. optional levels per view, and provider-specific attributes, see Semantic Conventions for GenAI agent and framework spans.

Additional span attributes

The OTel GenAI semantic conventions define more gen_ai.* span attributes that apply across providers than the ones AI Observability currently visualizes, see Core span attributes. None of these additional span attributes have a dedicated tile yet, but they're ingestible and queryable via DQL if your instrumentation sets them. For provider-specific attributes (OpenAI, AWS Bedrock, Azure AI Foundry), see the matching vendor under Vendor-native metrics and attributes.

More common span attributes

Requirement levels follow Semantic Conventions for GenAI spans.

Attribute	Requirement level	Description
`gen_ai.request.max_tokens`	Recommended	Maximum number of tokens requested.
`gen_ai.request.temperature`	Recommended	Sampling temperature for the request. Visualized in the Model versioning and A/B testing dashboard, but not in the custom app pages.
`gen_ai.request.top_p`	Recommended	The `top_p` sampling setting for the request.
`gen_ai.request.stop_sequences`	Recommended	Sequences that stop generation, if requested.
`gen_ai.request.seed`	Conditionally required	Seed used for deterministic sampling, if included in the request.
`gen_ai.request.choice.count`	Conditionally required	Number of response choices requested, if available and not `1`.
`gen_ai.response.id`	Recommended	Unique identifier for the response.
`gen_ai.response.finish_reasons`	Recommended	Reasons the model stopped generating, per choice.
`gen_ai.usage.cache_read.input_tokens`	Recommended	Input tokens served from a prompt cache. Distinct from the Bedrock-specific `gen_ai.prompt.caching` metric. For more information, see Amazon Bedrock.
`gen_ai.usage.cache_creation.input_tokens`	Recommended	Input tokens used to write to a prompt cache.
`error.type`	Conditionally required	Error class the operation ended with, if it errored. Same field also used as an attribute on the GenAI client metrics. For more information, see GenAI client metrics.

Agent span attributes

Set on spans for agent invocations. See Semantic Conventions for GenAI agent and framework spans for the full specification.

This is done in addition to the already-visualized gen_ai.agent.name and gen_ai.input.messages/gen_ai.output.messages, see Core span attributes

Attribute	Requirement level	Description
`gen_ai.agent.id`	Conditionally required	Provider-assigned stable identifier of the agent (not a transient in-memory instance ID), if applicable.
`gen_ai.agent.description`	Conditionally required	Free-form description of the agent, if provided by the application.
`gen_ai.agent.version`	Conditionally required	Version of the agent, if provided by the application.
`gen_ai.data_source.id`	Conditionally required	Identifier of the data source used for RAG, if applicable.
`gen_ai.output.type`	Conditionally required	Content type requested from the model (for example `text`, `json`, `image`), if the request includes an output format.
`gen_ai.system_instructions`	Opt-in	System message or instructions given to the model, separate from the chat history.
`gen_ai.tool.definitions`	Opt-in	Tool definitions available to the agent.

Vendor-native metrics and attributes

This page lists only what AI Observability currently visualizes. Vendors and gateways expose far more native metrics and attributes; for each vendor's complete signal set, reference the vendor documentation as linked in the relevant section.

For the full list of AI platforms and gateways that Dynatrace integrates with, see AI Observability integrations.

For a high-level overview, see Which metrics are available?.

Amazon Bedrock

AI Observability visualizes the following metrics and attributes.

Metric or attribute	Type	Used for
`gen_ai.bedrock.guardrail.activation`, `.content`, `.sensitive_info`, `.contextual`, `.topics`, `.words`	Span attribute	Guardrail overview cards, contextual grounding score
`gen_ai.guardrail.grounding_type`	Span attribute (metric dimension)	Splits grounding vs. relevance guardrail tiles
`gen_ai.prompt.caching`	Metric (counter)	Cache-read/write token savings, distinct from the span attribute `gen_ai.prompt_caching`

See Amazon Bedrock Guardrails for the full guardrail feature set.

Dynatrace also ingests the OTel semantic-convention attributes aws.bedrock.guardrail.id (required) and aws.bedrock.knowledge_base.id (recommended), set when gen_ai.provider.name is aws.bedrock. (For more information, see AWS Bedrock semantic conventions.) These are distinct from the gen_ai.bedrock.guardrail.* attributes and aren't used by a tile today.

Azure OpenAI

AI Observability visualizes the following metrics and attributes.

Metric or attribute	Type	Used for
`gen_ai.prompt.prompt_filter_results`	Span attribute	Guardrail overview cards
`gen_ai.completion.content_filter_results`	Span attribute	Guardrail overview cards

See Content filtering for Azure OpenAI for the full set of filter categories and severity levels.

Dynatrace also ingests the OTel semantic-convention attribute azure.resource_provider.namespace (recommended), set when gen_ai.provider.name is azure.ai.inference. For more information, see Azure AI Inference semantic conventions. It identifies the Azure Resource Provider Namespace (Microsoft.CognitiveServices for Azure AI Inference operations) and isn't used by a tile today.

The ready-made Azure AI Foundry dashboard covers cost, latency, and token totals only. It doesn't add content-filter tiles beyond the custom app tiles.

NVIDIA NIM

NVIDIA NIM inference servers expose their own Prometheus-compatible metrics covering request latency, throughput, token counts, and GPU utilization. There's no dedicated NIM tile in the custom app pages, but the ready-made NVIDIA dashboard visualizes a subset directly:

Metric	Used for
`request_prompt_tokens`, `request_generation_tokens`	Token usage and cost-estimate tiles
`e2e_request_latency_seconds`	Average and p99 request-latency tiles
`time_to_first_token_seconds`	Time-to-first-token tile
`generation_tokens_total`	Throughput tile
`gpu_cache_usage_perc`	GPU cache utilization tile
`num_requests_running`	In-flight request count
`request_success_total`	Request-volume tile

Standalone vLLM deployments expose this same metric set, since NIM proxies vLLM's own metrics without renaming or wrapping them.

See NVIDIA NIM observability reference for the full Prometheus metric set.

Google Gemini / Vertex AI

Gemini and Vertex AI return safety ratings on each response: the safetyRatings array has a harm category, probability, and blocked flag per category, plus a finishReason of SAFETY when a response is blocked.

Neither the custom app pages nor the ready-made Gemini/Vertex AI dashboard visualize safety ratings yet. To do this, you need to forward them to Dynatrace and query via DQL. The ready-made dashboard does cover the same generic cost, latency, and token tiles as the other vendor dashboards.

Kong AI Gateway

Kong AI Gateway exposes its own AI Gateway observability signals covering token usage, cost, and latency across LLM traffic. There's no dedicated Kong tile in the custom app pages, but the ready-made Kong dashboard visualizes:

Metric	Used for
`kong_ai_llm_requests_total`	Request-volume tiles by provider and model
`kong_ai_llm_tokens_total`	Token usage and cost tiles, split by `token_type`
`kong_ai_llm_provider_latency_ms`	Average and p99 provider-latency tiles
`kong_kong_latency_ms`	Gateway-side latency by service and route

See Kong AI Gateway observability for the full signal reference.

OpenAI

Dynatrace also ingests the following OTel semantic-convention attributes, set when gen_ai.provider.name is openai. For more information, see OpenAI semantic conventions. None of them are used by a tile today.

Attribute	Requirement level	Description
`openai.api.type`	Recommended	Which OpenAI API served the request, for example `chat_completions` or `responses`.
`openai.request.service_tier`	Conditionally required	The service tier requested, if included and not `auto`.
`openai.response.service_tier`	Conditionally required	The service tier used for the response, if included.
`openai.response.system_fingerprint`	Recommended	Fingerprint tracking any change in the GenAI environment that generated the response.

The ready-made OpenAI dashboard follows the same pattern as Bedrock and Azure OpenAI's cost/latency/token tiles, built entirely on standard gen_ai.client.token.usage, gen_ai.token.type, and gen_ai.prompt.caching.

Other OTel or third-party sources

Some instrumentation libraries don't emit gen_ai.* names natively. These need a normalization step before AI Observability can read them.

OpenInference (used by frameworks like LangChain and LlamaIndex) emits its own convention (llm.model_name, llm.token_count.*, and so on). Normalize it to gen_ai.* using an OTel Collector transform processor or Dynatrace OpenPipeline. OpenInference example provides the full attribute mapping and known gaps (for example, guardrail details beyond the validator name aren't yet translated).
Custom log-derived metrics. For sources with no OTel exporter (for example, parsing Ollama server logs), you can compute your own metrics and forward them to Dynatrace using the standard metrics ingest API. These won't match gen_ai.* naming automatically, so use consistent naming and dimensions if you want them to appear alongside standard views.

AI Observability metrics reference

Latest Dynatrace
Reference
8-min read
Published Jul 02, 2026

This page lists every standard metric and span attribute that Dynatrace can ingest for GenAI workloads, grouped by where the data comes from.

OpenTelemetry-native metrics: First-class OTel metric instruments (gen_ai.client.*, gen_ai.server.*, gen_ai.workflow.*, gen_ai.invoke_agent.*, gen_ai.execute_tool.*) emitted directly by your instrumentation (Traceloop SDK, OneAgent, or other OTel/OpenLLMetry-based SDKs). Dynatrace ingests these as-is, and doesn't calculate them.
Dynatrace span-derived visualizations: Built by Dynatrace directly from gen_ai.* span attributes. No separate metric instrument is emitted for these; AI Observability computes tiles, filters, and breakdowns straight from the span data.
Vendor-native metrics: These are sent directly by the AI platform or gateway (Amazon Bedrock, Azure OpenAI, NVIDIA NIM, Kong AI Gateway), either as provider-specific span attributes or as their own metrics.
Other OTel or third-party sources: These are metrics that you derive yourself (for example, by parsing inference-server or application logs) and then forward to Dynatrace.

OpenTelemetry-native metrics

All signals are ingestible and queryable via DQL. Some signals are visualized in tiles, as noted in the tables.

GenAI client metrics (OpenTelemetry)

Metric	Instrument type	Unit	Description	Visualization
`gen_ai.client.token.usage`	Histogram	`{token}`	Number of input and output tokens used.	Shows cost tiles, with token usage split by type
`gen_ai.client.operation.duration`	Histogram	`s`	GenAI operation duration.	Shows latency charts (p99, mean)
`gen_ai.client.operation.time_to_first_chunk`	Histogram	`s`	Time to receive the first chunk, measured from when the client issues the generation request to when the first chunk is received in the response stream. Only emitted for streaming calls.	Visualization not available
`gen_ai.client.operation.time_per_output_chunk`	Histogram	`s`	Time per output chunk, recorded for each chunk received after the first one, measured from the end of the previous chunk to the end of the current chunk. Only emitted for streaming calls.	Visualization not available

Metric: `gen_ai.client.token.usage`

Attribute	Requirement level	Value type	Description	Example values
`gen_ai.operation.name`	Required	string	The name of the operation being performed.	`chat`, `generate_content`, `text_completion`
`gen_ai.provider.name`	Required	string	The GenAI provider identified by the instrumentation.	`openai`, `aws.bedrock`, `azure.ai.openai`
`gen_ai.token.type`	Required	string	The type of token being counted. This splits the cost tile lanes.	`input`, `output`
`gen_ai.request.model`	Required	string	The model a request is being made to.	`gpt-5`
`gen_ai.response.model`	Required	string	The model that generated the response.	`gpt-5-0616`

If this metric is missing, all cost tiles show $0 with no visible error.

If the gen_ai.token.type attribute is missing, the tiles show an undifferentiated cost total instead of separate input/output lanes.

Metric: `gen_ai.client.operation.duration`

Attribute	Requirement level	Value type	Description	Example values
`gen_ai.operation.name`	Required	string	The name of the operation being performed.	`chat`, `generate_content`, `text_completion`
`error.type`	Conditionally required	string	Error class the operation ended with, if it errored.	`timeout`, `500`
`gen_ai.provider.name`	Conditionally required	string	The GenAI provider identified by the instrumentation.	`openai`, `aws.bedrock`
`gen_ai.request.model`	Required	string	The model a request is being made to.	`gpt-5`
`gen_ai.response.model`	Required	string	The model that generated the response.	`gpt-5-0616`

If this metric is missing, all latency tiles show as empty, with no error surfaced.

Metric: `gen_ai.client.operation.time_to_first_chunk`

Attribute	Requirement level	Value type	Description	Example values
`gen_ai.operation.name`	Required	string	The name of the operation being performed.	`chat`, `generate_content`, `text_completion`
`gen_ai.provider.name`	Required	string	The GenAI provider identified by the instrumentation.	`openai`, `gcp.gen_ai`, `gcp.vertex_ai`
`gen_ai.request.model`	Required	string	The model a request is being made to, if available.	`gpt-5`
`server.port`	Conditionally required	int	GenAI server port, if `server.address` is set.	`80`, `8080`, `443`
`gen_ai.response.model`	Required	string	The model that generated the response.	`gpt-5-0616`
`server.address`	Recommended	string	GenAI server address.	`example.com`

Metric: `gen_ai.client.operation.time_per_output_chunk`

Attribute	Requirement level	Value type	Description	Example values
`gen_ai.operation.name`	Required	string	The name of the operation being performed.	`chat`, `generate_content`, `text_completion`
`gen_ai.provider.name`	Required	string	The GenAI provider identified by the instrumentation.	`openai`, `gcp.gen_ai`, `gcp.vertex_ai`
`gen_ai.request.model`	Required	string	The model a request is being made to, if available.	`gpt-5`
`server.port`	Conditionally required	int	GenAI server port, if `server.address` is set.	`80`, `8080`, `443`
`gen_ai.response.model`	Required	string	The model that generated the response.	`gpt-5-0616`
`server.address`	Recommended	string	GenAI server address.	`example.com`

GenAI model server metrics (OpenTelemetry)

Metric	Instrument type	Unit	Description	Visualization
`gen_ai.server.request.duration`	Histogram	`s`	Generative AI server request duration, such as time-to-last-byte or last output token.	Visualization not available
`gen_ai.server.time_to_first_token`	Histogram	`s`	Time to generate the first token of the response, for streaming requests.	Visualization not available
`gen_ai.server.time_per_output_token`	Histogram	`s`	Time per output token generated after the first token, for successful responses.	Visualization not available

Common attributes

All three Gen AI model server metrics share this attribute set:

Attribute	Requirement level	Value type	Description	Example values
`gen_ai.operation.name`	Required	string	The name of the operation being performed.	`chat`, `generate_content`, `text_completion`
`gen_ai.provider.name`	Required	string	The GenAI provider identified by the client or server instrumentation.	`openai`, `gcp.gen_ai`, `gcp.vertex_ai`
`error.type`	Conditionally required	string	Error class the operation ended with, if it errored. Applies to `gen_ai.server.request.duration` only.	`timeout`, `500`
`gen_ai.request.model`	Required	string	The model a request is being made to, if available.	`gpt-5`
`server.port`	Conditionally required	int	GenAI server port, if `server.address` is set.	`80`, `8080`, `443`
`gen_ai.response.model`	Required	string	The model that generated the response.	`gpt-5-0616`
`server.address`	Recommended	string	GenAI server address.	`example.com`

GenAI workflow, agent, and tool metrics (OpenTelemetry)

Metric	Instrument type	Unit	Description	Status
`gen_ai.workflow.duration`	Histogram	`s`	End-to-end duration of a workflow execution, from application code initiating the workflow to the workflow completing.	Visualization not available
`gen_ai.invoke_agent.duration`	Histogram	`s`	End-to-end duration of a single in-process agent invocation, from invocation start until the agent emits its last response chunk or errors.	Visualization not available
`gen_ai.execute_tool.duration`	Histogram	`s`	Duration of a single tool execution.	Visualization not available

Metric: `gen_ai.workflow.duration`

Attribute	Requirement level	Value type	Description	Example values
`error.type`	Conditionally required	string	Error class the operation ended with, if it errored.	`timeout`, `500`
`gen_ai.workflow.name`	Conditionally required	string	Human-readable workflow name provided by the application, if available.	`multi_agent_rag`, `customer_support_pipeline`

Metric: `gen_ai.invoke_agent.duration`

Attribute	Requirement level	Value type	Description	Example values
`error.type`	Conditionally required	string	Error class the operation ended with, if it errored.	`timeout`, `500`
`gen_ai.agent.name`	Conditionally required	string	Human-readable name of the invoked GenAI agent, when available.	`Math Tutor`, `Fiction Writer`
`gen_ai.request.model`	Recommended	string	The model configured for the agent, if applicable.	`gpt-5`

Metric: `gen_ai.execute_tool.duration`

Attribute	Requirement level	Value type	Description	Example values
`gen_ai.tool.name`	Required	string	Name of the tool used by the agent.	`Flights`
`error.type`	Conditionally required	string	Error class the operation ended with, if it errored.	`timeout`, `500`
`gen_ai.agent.name`	Conditionally required	string	Human-readable name of the agent executing the tool, when applicable.	`Math Tutor`, `Fiction Writer`
`gen_ai.tool.type`	Recommended	string	Type of the tool used by the agent, if available.	`function`, `extension`, `datastore`

Dynatrace span-derived visualizations

Core span attributes

AI Observability visualizes the following attributes, which are set directly on spans and exist alongside the general OTel/Dynatrace fields service.name, span.status_code, and duration.

Attribute	Requirement level	Visualization
`gen_ai.provider.name` (or deprecated `gen_ai.system`)	Required	Shows provider filter, service explorer table
`service.name`	General OTel resource attribute, not GenAI-specific.	Shows service filter, service explorer table
`gen_ai.operation.name`	Required	Shows operation breakdown, trace analysis
`gen_ai.request.model`	Required	Shows prompts table, distributed tracing
`gen_ai.response.model`	Required	Shows model filter, overview charts
`gen_ai.usage.input_tokens` / `gen_ai.usage.output_tokens`	Required	Shows prompts table, cost tile fallback when the `gen_ai.client.token.usage` metric isn't emitted
`span.status_code`	General OTel span field, not GenAI-specific.	Shows service health tile (success/failure rate)
`duration`	General OTel span field, not GenAI-specific.	Shows average and p50/p90 latency in the Explorer per-service and per-model tables
`gen_ai.agent.name`	Conditionally required	Shows agent vs. LLM distinction, agent filter
`gen_ai.conversation.id`	Conditionally required	Shows groups spans into conversation threads in the prompts view
`gen_ai.input.messages` / `gen_ai.output.messages`	Required	Shows prompts table content

For agent-specific and provider-specific span attributes, see Additional span attributes and Vendor-native metrics and attributes.

For the full attribute list, required vs. optional levels per view, and provider-specific attributes, see Semantic Conventions for GenAI agent and framework spans.

Additional span attributes

More common span attributes

Requirement levels follow Semantic Conventions for GenAI spans.

Attribute	Requirement level	Description
`gen_ai.request.max_tokens`	Recommended	Maximum number of tokens requested.
`gen_ai.request.temperature`	Recommended	Sampling temperature for the request. Visualized in the Model versioning and A/B testing dashboard, but not in the custom app pages.
`gen_ai.request.top_p`	Recommended	The `top_p` sampling setting for the request.
`gen_ai.request.stop_sequences`	Recommended	Sequences that stop generation, if requested.
`gen_ai.request.seed`	Conditionally required	Seed used for deterministic sampling, if included in the request.
`gen_ai.request.choice.count`	Conditionally required	Number of response choices requested, if available and not `1`.
`gen_ai.response.id`	Recommended	Unique identifier for the response.
`gen_ai.response.finish_reasons`	Recommended	Reasons the model stopped generating, per choice.
`gen_ai.usage.cache_read.input_tokens`	Recommended	Input tokens served from a prompt cache. Distinct from the Bedrock-specific `gen_ai.prompt.caching` metric. For more information, see Amazon Bedrock.
`gen_ai.usage.cache_creation.input_tokens`	Recommended	Input tokens used to write to a prompt cache.
`error.type`	Conditionally required	Error class the operation ended with, if it errored. Same field also used as an attribute on the GenAI client metrics. For more information, see GenAI client metrics.

Agent span attributes

Set on spans for agent invocations. See Semantic Conventions for GenAI agent and framework spans for the full specification.

This is done in addition to the already-visualized gen_ai.agent.name and gen_ai.input.messages/gen_ai.output.messages, see Core span attributes

Attribute	Requirement level	Description
`gen_ai.agent.id`	Conditionally required	Provider-assigned stable identifier of the agent (not a transient in-memory instance ID), if applicable.
`gen_ai.agent.description`	Conditionally required	Free-form description of the agent, if provided by the application.
`gen_ai.agent.version`	Conditionally required	Version of the agent, if provided by the application.
`gen_ai.data_source.id`	Conditionally required	Identifier of the data source used for RAG, if applicable.
`gen_ai.output.type`	Conditionally required	Content type requested from the model (for example `text`, `json`, `image`), if the request includes an output format.
`gen_ai.system_instructions`	Opt-in	System message or instructions given to the model, separate from the chat history.
`gen_ai.tool.definitions`	Opt-in	Tool definitions available to the agent.

Vendor-native metrics and attributes

For the full list of AI platforms and gateways that Dynatrace integrates with, see AI Observability integrations.

For a high-level overview, see Which metrics are available?.

Amazon Bedrock

AI Observability visualizes the following metrics and attributes.

Metric or attribute	Type	Used for
`gen_ai.bedrock.guardrail.activation`, `.content`, `.sensitive_info`, `.contextual`, `.topics`, `.words`	Span attribute	Guardrail overview cards, contextual grounding score
`gen_ai.guardrail.grounding_type`	Span attribute (metric dimension)	Splits grounding vs. relevance guardrail tiles
`gen_ai.prompt.caching`	Metric (counter)	Cache-read/write token savings, distinct from the span attribute `gen_ai.prompt_caching`

See Amazon Bedrock Guardrails for the full guardrail feature set.

Azure OpenAI

AI Observability visualizes the following metrics and attributes.

Metric or attribute	Type	Used for
`gen_ai.prompt.prompt_filter_results`	Span attribute	Guardrail overview cards
`gen_ai.completion.content_filter_results`	Span attribute	Guardrail overview cards

See Content filtering for Azure OpenAI for the full set of filter categories and severity levels.

The ready-made Azure AI Foundry dashboard covers cost, latency, and token totals only. It doesn't add content-filter tiles beyond the custom app tiles.

NVIDIA NIM

Metric	Used for
`request_prompt_tokens`, `request_generation_tokens`	Token usage and cost-estimate tiles
`e2e_request_latency_seconds`	Average and p99 request-latency tiles
`time_to_first_token_seconds`	Time-to-first-token tile
`generation_tokens_total`	Throughput tile
`gpu_cache_usage_perc`	GPU cache utilization tile
`num_requests_running`	In-flight request count
`request_success_total`	Request-volume tile

Standalone vLLM deployments expose this same metric set, since NIM proxies vLLM's own metrics without renaming or wrapping them.

See NVIDIA NIM observability reference for the full Prometheus metric set.

Google Gemini / Vertex AI

Kong AI Gateway

Metric	Used for
`kong_ai_llm_requests_total`	Request-volume tiles by provider and model
`kong_ai_llm_tokens_total`	Token usage and cost tiles, split by `token_type`
`kong_ai_llm_provider_latency_ms`	Average and p99 provider-latency tiles
`kong_kong_latency_ms`	Gateway-side latency by service and route

See Kong AI Gateway observability for the full signal reference.

OpenAI

Attribute	Requirement level	Description
`openai.api.type`	Recommended	Which OpenAI API served the request, for example `chat_completions` or `responses`.
`openai.request.service_tier`	Conditionally required	The service tier requested, if included and not `auto`.
`openai.response.service_tier`	Conditionally required	The service tier used for the response, if included.
`openai.response.system_fingerprint`	Recommended	Fingerprint tracking any change in the GenAI environment that generated the response.

Other OTel or third-party sources

Some instrumentation libraries don't emit gen_ai.* names natively. These need a normalization step before AI Observability can read them.

OpenInference (used by frameworks like LangChain and LlamaIndex) emits its own convention (llm.model_name, llm.token_count.*, and so on). Normalize it to gen_ai.* using an OTel Collector transform processor or Dynatrace OpenPipeline. OpenInference example provides the full attribute mapping and known gaps (for example, guardrail details beyond the validator name aren't yet translated).
Custom log-derived metrics. For sources with no OTel exporter (for example, parsing Ollama server logs), you can compute your own metrics and forward them to Dynatrace using the standard metrics ingest API. These won't match gen_ai.* naming automatically, so use consistent naming and dimensions if you want them to appear alongside standard views.

AI Observability metrics reference

OpenTelemetry-native metrics

GenAI client metrics (OpenTelemetry)

Metric: gen_ai.client.token.usage

Metric: gen_ai.client.operation.duration

Metric: gen_ai.client.operation.time_to_first_chunk

Metric: gen_ai.client.operation.time_per_output_chunk

GenAI model server metrics (OpenTelemetry)

Common attributes

GenAI workflow, agent, and tool metrics (OpenTelemetry)

Metric: gen_ai.workflow.duration

Metric: gen_ai.invoke_agent.duration

Metric: gen_ai.execute_tool.duration

Dynatrace span-derived visualizations

Core span attributes

Additional span attributes

More common span attributes

Agent span attributes

Vendor-native metrics and attributes

Amazon Bedrock

Azure OpenAI

NVIDIA NIM

Google Gemini / Vertex AI

Kong AI Gateway

OpenAI

Other ready-made dashboards

Other OTel or third-party sources

Related topics

AI Observability metrics reference

OpenTelemetry-native metrics

GenAI client metrics (OpenTelemetry)

Metric: gen_ai.client.token.usage

Metric: gen_ai.client.operation.duration

Metric: gen_ai.client.operation.time_to_first_chunk

Metric: gen_ai.client.operation.time_per_output_chunk

GenAI model server metrics (OpenTelemetry)

Common attributes

GenAI workflow, agent, and tool metrics (OpenTelemetry)

Metric: gen_ai.workflow.duration

Metric: gen_ai.invoke_agent.duration

Metric: gen_ai.execute_tool.duration

Dynatrace span-derived visualizations

Core span attributes

Additional span attributes

More common span attributes

Agent span attributes

Vendor-native metrics and attributes

Amazon Bedrock

Azure OpenAI

NVIDIA NIM

Google Gemini / Vertex AI

Kong AI Gateway

OpenAI

Other ready-made dashboards

Other OTel or third-party sources

Related topics

Metric: `gen_ai.client.token.usage`

Metric: `gen_ai.client.operation.duration`

Metric: `gen_ai.client.operation.time_to_first_chunk`

Metric: `gen_ai.client.operation.time_per_output_chunk`

Metric: `gen_ai.workflow.duration`

Metric: `gen_ai.invoke_agent.duration`

Metric: `gen_ai.execute_tool.duration`

Metric: `gen_ai.client.token.usage`

Metric: `gen_ai.client.operation.duration`

Metric: `gen_ai.client.operation.time_to_first_chunk`

Metric: `gen_ai.client.operation.time_per_output_chunk`

Metric: `gen_ai.workflow.duration`

Metric: `gen_ai.invoke_agent.duration`

Metric: `gen_ai.execute_tool.duration`