An LLM inference span is named "{gen_ai.operation.name} {gen_ai.request.model}", for example "chat gpt-4". The span carries gen_ai.provider.name (required, such as openai or aws.bedrock), gen_ai.request.model, and token counts in gen_ai.usage.input_tokens and gen_ai.usage.output_tokens. Agent spans use operations such as create_agent and invoke_agent, and tool calls use execute_tool. The conventions also define spans and metrics for MCP clients and servers.
The conventions are still marked Development, so attribute and metric names can change between releases. In 2026 they moved from the main semantic-conventions repository to a dedicated repository, open-telemetry/semantic-conventions-genai, which uses OpenTelemetry Weaver to generate the docs. Instrumentation should pin a semantic conventions version and expect migrations.
OpenLIT SDKs emit spans and metrics that follow the GenAI semantic conventions for LLM providers, vector databases, and agent frameworks, with a single openlit.init() call. Because the output is standard OTLP, the same traces work in OpenLIT, Grafana, Datadog, or any OpenTelemetry backend without re-instrumenting.