New Relic AI Monitoring

Editor’s note: This New Relic AI Monitoring review is based on my hands-on evaluation of New Relic AI Monitoring on New Relic’s free tier in October 2026, using the Python APM agent and the OpenTelemetry path. I weighed pricing, setup, LLM tracing, agent and MCP coverage, privacy controls, and rivals.

Quick verdict: I recommend New Relic AI Monitoring for teams already on New Relic APM who want LLM latency, errors and token cost in the same trace as the rest of the stack. The trade-off: AI Evaluation only enters public preview in November 2026, so teams that need LLM-as-judge scoring today still need Langfuse, Arize or Braintrust beside it.

LLM calls become spans inside existing distributed traces across 6 APM agents, with OpenTelemetry ingest GA since August 2026. There is no separate AI charge: AI data bills as ordinary ingest plus full-platform seats.

In this review, I’ll take a closer look at New Relic AI Monitoring’s pricing, features and language coverage, so you can see exactly whether it belongs in your LLM stack.

Key Takeaways

• 100 GB/month and 1 full-platform user are free, then Original ingest costs $0.40/GB.
• LLM spans share the distributed trace with APM data, so a slow answer traces back to the database or service behind it.
• Model inventory shows errors, latency and token usage per model.
• Library coverage is uneven: Ruby instruments only the ruby-openai gem, and Go only OpenAI and AWS SDK for Go v2.
• Evaluation is not GA: AI Evaluation is preview-only from November 2026, and full-platform seats run $99 to $349 a month each.

Pros and Cons

How Much Does New Relic AI Monitoring Cost?

New Relic AI Monitoring has no separate price: AI observability is listed as a platform capability, so LLM data bills through data ingest and user seats. The 100 GB free allowance is shared with every other signal the account sends, so it is not an extra 100 GB reserved for LLM telemetry.

• Free ($0): For a single full-platform user, with 100 GB/month of ingest, unlimited Basic users and 500 synthetic checks.
• Standard ($10 first full-platform user, then $99 each): For teams needing up to 5 full-platform users. Core users cost $49 each.
• Pro ($349/user/month annually): For unlimited full-platform users, SAML SSO and extended retention. Monthly pay-as-you-go costs $418.80 per full-platform user.
• Enterprise (custom quote): For 24/7 support and regulated workloads, including FedRAMP Moderate and HIPAA eligibility with Data Plus.

I budget the data option separately from the seat edition:

• Original: $0.40/GB beyond the free 100 GB/month.
• Data Plus: $0.60/GB beyond the allowance, with longer retention.
• EU data center storage: An additional $0.05/GB per month.

Plan	Full-platform user price	Max full users	Ingest	Notes
Free	$0	1	100 GB/month included	Unlimited Basic users; no Core users
Standard	$10 first; $99 each additional/month	5	100 GB free; Original $0.40/GB beyond	Core users $49/month
Pro	$349/month annually; $418.80 monthly	Unlimited	Original $0.40/GB; Data Plus $0.60/GB beyond allowance	SAML SSO, extended retention
Enterprise	Custom quote	Contract terms	Usage-based	24/7 support; FedRAMP Moderate and HIPAA eligibility with Data Plus

A Standard account with 2 full-platform users and 300 GB of monthly Original ingest costs $189/month: $10 + $99 + (200 x $0.40 = $80). Nothing in that total is an AI fee.

New Relic markets the platform as “One platform. One price.” The engineer assistant, New Relic AI, is a different product billed as Advanced Compute in CCUs, with no published per-CCU rate. Datadog LLM Observability meters LLM spans instead, with Pro at $160/month annually for 100,000 LLM spans.

Is New Relic AI Monitoring Good Value for Money?

• Existing APM customers get the strongest value: With seats already paid, turning on AI Monitoring adds only the extra GB of LLM events, not a second observability subscription.
• Content capture drives the ingest bill: Prompts and completions are text-heavy, so record_content.enabled decides much of what you pay for.
• Seats outweigh ingest: 10 Pro engineers cost $3,490/month on an annual commitment before data charges. Arize AX Pro costs $50/month with unlimited users, though its 50,000-span allowance fits a smaller workload.
• Retention differs by data type: Original keeps Llm* custom events for 30 days but spans for only 8. Data Plus extends those defaults to 120 and 98 days, and 30 extra days cost $0.05 per GB ingested per month.

Reviewer’s Notes: I recommend Free for a pilot that fits in 100 GB, then Standard while 5 or fewer engineers need full access. I would move to Pro only when you need more full-platform users or SSO, not simply because an application now calls an LLM.

My Experience With New Relic AI Monitoring

I enabled AI Monitoring through the Python APM agent, and content capture was a separate decision from switching monitoring on. Helicone takes a different route, sending requests through its proxy instead of instrumenting the application with an agent.

Enabling AI Monitoring on the Python Agent

The Python agent supports AI Monitoring from version 9.8.0 (Node.js from 11.13.0, Java from 8.12.0, Go from 3.31.0, Ruby from 9.8.0). I kept distributed tracing on and high security mode off, because either setting blocks AI capture entirely.

I set the AI switches in newrelic.ini:

ai_monitoring.enabled = true
ai_monitoring.record_content.enabled = false

The license key came from my secret store through the environment, never from the config file. NEW_RELIC_AI_MONITORING_ENABLED is the environment-variable equivalent of the monitoring switch.

Python agent 13.3.0 (July 2026) added server-side settings for ai_monitoring.enabled, record_content.enabled and streaming.enabled, and server-side values override local ones. Check the UI setting before you assume a newrelic.ini change controls capture.

Deciding What Content to Capture

Setting ai_monitoring.record_content.enabled to false keeps prompts and responses inside the app. It also stops token counts from being forwarded unless the application registers set_llm_token_count_callback to count locally.

Streaming has its own streaming.enabled switch, but only in Python, Node.js and Go. Feedback and token callbacks also differ by language, so the 6 agents do not expose identical controls.

I opened All capabilities > AI monitoring > Drop filters for the privacy layer:

1. Pick an event type: Each filter targets one event type.
2. Pick an attribute: One regex applies to one attribute of that event.
3. Repeat: Stripping one kind of data from every AI event type takes six filters.

Dropped data cannot be recovered, and New Relic does not guarantee filters fully resolve disclosure risk. New customers drop data with NRQL through Pipeline Control, since legacy NRQL drop rules reached end of life on June 30, 2026.

The OpenTelemetry Route

I set OTEL_EXPORTER_OTLP_ENDPOINT to the regional OTLP endpoint on port 443 and passed the license key from my secret store as the api-key header in OTEL_EXPORTER_OTLP_HEADERS. I then set OTEL_SERVICE_NAME, installed a GenAI instrumentation package and queried Span where gen_ai.system is set to confirm data arrived.

Content capture is off by default, and the opt-in flag depends on the package: OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT or TRACELOOP_TRACE_CONTENT. Token counts depend on the instrumentation version, and long prompts are truncated unless you raise OTEL_ATTRIBUTE_VALUE_LENGTH_LIMIT.

Reviewer’s Notes: Pin agent versions and test upgrades in staging. Users reported that Python 13.0.1 crashed every LangGraph streaming turn until they downgraded, that the Python 13.x agent broke SigV4 signing on Bedrock AgentCore calls, and that Node 14.3.3 broke OpenAI responses.parse() even with AI Monitoring disabled. Turning capture off is not a substitute for regression testing the agent itself.

AI Responses, Traces and the Model Inventory

An instrumented LLM call becomes a span in the existing distributed trace, so the waterfall shows model latency beside database and downstream service calls. That shared waterfall is the strongest reason to keep LLM tracing inside an APM platform.

AI Responses Page

I opened APM & Services, selected the entity and went to AI monitoring > AI responses. The page combines aggregate activity with response-level detail:

• Summary tiles: Total responses, average response time and average tokens per response.
• Time series: Response activity across the selected period.
• Response table: Prompt/response pairs, token counts and the model, subject to capture settings.

Each response row opens the trace waterfall, with errors highlighted in red next to the database and service spans of the same request.

Feedback takes application code: record_llm_feedback_event ties a positive or negative rating to a trace ID, and the AI responses page then filters by it. .NET lists no feedback method.

AI Entities

AI entities lists every service reporting AI data with response time, throughput and error rate. Each APM entity also gets its own AI responses page, so I moved from the affected service to its individual responses without leaving New Relic.

I use the entity view to find the failing service, then the response waterfall to separate model latency from dependency latency.

Model Inventory and Model Comparison

The model inventory splits model questions across four tabs:

• Overview: Model-level activity.
• Errors: Failures grouped by model and by error type.
• Performance: Slowest models and outlier detection.
• Cost: Tokens against limits, tokens by model, and prompt versus completion tokens.

Model comparison takes 2 models and filters by service and time range, which covers before-and-after deployment checks. It compares performance and tokens per completion, but neither measures answer quality.

Best for: full-stack incident investigation across model, database and service calls. Skip if: chain-level prompt debugging matters more than APM correlation; LangSmith fits that better.

Agents, MCP Servers and OpenTelemetry Coverage

LangGraph, Strands and AutoGen agent monitoring needs no configuration beyond enabling AI Monitoring on a supported APM agent. That framework support is separate from the limited-preview agentic suite, which matters for production planning.

Agent views report operational measurements, not answer quality:

• Invocation latency: Time spent in each agent invocation.
• Agent and tool errors: Error rates for agents and tools.
• Token usage: Consumption attributed to each agent.
• Entity map: Agents and tools as nodes, alongside a trace waterfall.

Python instruments FastMCP 2.3.3 and later; Node.js and Java instrument their MCP SDKs. New Relic claims MCP request-lifecycle capture links tool performance to databases, services and queues.

Python 13.3.0 fixed duplicate tool entities for LangChain tools backed by a separately instrumented remote MCP server, so run 13.3.0 or later if your entity map has to show tool ownership accurately.

Agentic AI Monitoring and the New Relic AI MCP Server were announced on November 4, 2025 as limited previews. They cover an AI Inventory of agents and tools plus an Agents Service Map, with no GA date or separate price.

OpenTelemetry support became GA on August 26, 2026, giving an agent-free route for stacks the six agents miss. New Relic’s examples cover:

• OpenAI
• Anthropic
• LangChain
• AWS Bedrock

If you do not need APM context at all, open-source Arize Phoenix or self-hosted Langfuse handle LLM tracing without New Relic seats. I recommend New Relic’s OpenTelemetry route when the LLM spans must sit next to service traces.

Reviewer’s Notes: Confirm preview access before making agent maps a deployment requirement. OpenTelemetry ingest is GA, but that does not make Agentic AI Monitoring or the New Relic AI MCP Server generally available.

Language and Framework Coverage: Where the Gaps Are

New Relic supports 6 APM languages, but library depth differs sharply: Python and Node.js instrument far more AI libraries than Ruby or Go. The exact SDK and API operation is the compatibility requirement, not the language name.

Language	Minimum agent	Instrumented AI libraries
Python	9.8.0	OpenAI 0.28.0+; Boto3/Bedrock 1.28.57+; LangChain 0.1.0+; LangGraph 1.0.7+ with LangChain 1.2.4+; Google Gen AI SDK 1.3.0+; FastMCP 2.3.3+
Node.js	11.13.0	OpenAI; Anthropic; Bedrock, including text-only Converse; Google ADK; Google Gen AI; LangChain/LangGraph; MCP SDK
Java	8.12.0	Bedrock; Bedrock Converse; MCP Java SDK; Spring AI
Go	3.31.0	Go OpenAI 1.19.4+; AWS SDK for Go v2 1.6.0+
Ruby	9.8.0	ruby-openai 3.4.0+ only
.NET	Bedrock: 10.23.0; OpenAI integrations: 10.37.0	AWSSDK.Bedrock 3.7.200+; OpenAI 2.0.0+; Azure.AI.OpenAI 2.0.0+

.NET sets no agent-wide minimum: Bedrock needs agent 10.23.0, while OpenAI and Azure.AI.OpenAI need 10.37.0.

Node.js coverage also stops at specific operations:

• OpenAI: Chat, completions and embeddings, not audio, files or images.
• Anthropic: Chat and completions, not files.
• Google Gen AI: Chat, embeddings and text, not audio, cache, image, PDF or video.
• Bedrock InvokeModel: Titan, Claude, Cohere and Llama 3, text only.

Node.js also needs openai 4.0.0 or later and @anthropic-ai/sdk 0.33.0 or later.

NVIDIA NIM models need no extra monitoring steps. Python’s supported-library list omits Anthropic, although the agent instruments the Anthropic SDK.

Users have also reported narrower gaps:

• OpenAI Assistants API: Python monitors chat completions but not the Assistants API; the request for support dates to January 2025.
• Anthropic through LangChain: Python 13.5.0 recorded zero tokens and no content for those calls, with a fix targeted for 13.6.0.
• Bedrock inference profiles: Node.js lacked inference-profile model ID support, with implementation work opened in August 2026.

Check New Relic’s compatibility list against your exact library versions and test the API operations your app actually calls. A supported provider is not blanket coverage for every endpoint, streaming path or framework wrapper.

AI Evaluation: Quality Scoring Arrives in Preview

New Relic announced AI Evaluation on October 6, 2026, with public preview scheduled for November 2026. It is an asynchronous LLM-as-a-judge service that scores live telemetry and attaches the scores to traces as attributes; I did not use it in my hands-on testing.

The announced scope goes beyond latency and token accounting:

• Quality scoring: Hallucination, prompt-injection and data-leak detection.
• Guardrail checks: Injection, jailbreak, PII, toxicity and bias.
• RAG metrics: Faithfulness and answer relevancy.
• Pre-production workflows: Prompt Playground, versioned datasets and experiments.

New Relic claims it evaluates the whole application transaction rather than isolated LLM calls. Scores on the same traces used for incident work are appealing, but a preview is not a production commitment.

No GA date, price or evaluator list is published, so I would test the preview separately from today’s monitoring purchase.

As of October 2026, positive and negative feedback events are the only GA quality signal in AI Monitoring. Braintrust (50,000 scores on Pro) and Arize AX (unlimited evals on every tier) already sell evaluation as a GA product, so pair one with New Relic if you need automated scoring now.

How Does New Relic AI Monitoring Compare to Competitors?

New Relic AI Monitoring is the APM-native option for existing New Relic teams, not a default replacement for an LLM evaluation platform.

Tool	Pricing headline	Best for
New Relic AI Monitoring	100 GB free; Original $0.40/GB beyond; Standard $10 first full user, $99 additional; Pro $349/user/month annually	Existing New Relic APM teams
Datadog LLM Observability	40,000 LLM spans free; Pro $160/month annually for 100,000 spans	Datadog customers
Langfuse	Free self-hosting; Core $29/month; Pro $199/month	Open source and self-hosting
LangSmith	Plus $39/seat/month; 10,000 base traces	LangChain and LangGraph teams
Arize AX / Phoenix	AX Pro $50/month, 50,000 spans, unlimited users; Phoenix open source	Evaluation with unlimited users
Braintrust	Pro $249/month	Evaluation and prompt iteration
Helicone	Pro $79/month	Proxy-based setup

• Datadog LLM Observability: The closest like-for-like rival for teams already on Datadog APM. Its free tier includes 40,000 LLM spans with 15-day retention, tool, workflow and retrieval spans do not count toward billing, and Sensitive Data Scanner is included.
• Langfuse: I recommend Langfuse when self-hosting and open-source control matter more than infrastructure correlation. Core includes 100,000 units for $29/month, and self-hosting is free, but it is not a full-stack APM.
• LangSmith: I prefer LangSmith for teams built on LangChain and LangGraph. Plus costs $39 per seat with 10,000 base traces, kept 14 days, or 180 days for extended traces, so the bill grows with every collaborator.
• Arize AX and Phoenix: I recommend Arize for GA evaluation without per-seat costs. AX Pro includes unlimited users and evals for $50/month, but caps you at 50,000 spans and 10 GB. Phoenix is the open-source, local-first route.
• Braintrust: I recommend Braintrust when scoring and prompt iteration drive the purchase. Pro includes 5 GB of processed data and 50,000 scores for $249/month, with 30-day retention, and it sits beside an APM rather than replacing one.
• Helicone: I recommend Helicone for a proxy-first setup instead of an agent rollout. Pro costs $79/month, and the free tier includes 10,000 requests, 1 GB and 1 seat, at the cost of routing traffic through its proxy or SDK with less APM context.

Dynatrace and Splunk Agent Observability are the other enterprise APM suites with LLM monitoring; I would only shortlist them if you already run them.

New Relic meters GB and seats, Datadog meters LLM spans, and Braintrust combines processed data with scores. Equal monthly prices do not buy equal coverage, so compare them against your own span volume, retention and evaluation needs.

How I Reviewed New Relic AI Monitoring

I set up AI Monitoring on the free tier through the Python agent configuration and the OpenTelemetry route, then worked through the AI responses page, model inventory and agent views.

My evaluation covered:

• Pricing: Ingest allowances, seat escalation and retention.
• Setup: Agent switches, server-side overrides, drop filters and OTLP configuration.
• Tracing: AI responses, model inventory and service correlation.
• Coverage: All 6 agents, framework support and preview boundaries.
• Privacy: Content suppression, token callbacks and drop-filter scope.
• Upgrade risk: User-reported regressions in Python 13.x and Node.js 14.3.3 agents.

I weighed users’ agent-upgrade failures against the benefit of reusing existing APM instrumentation, and compared New Relic with Datadog, Langfuse, LangSmith, Arize, Braintrust and Helicone on price and on monitoring versus evaluation.

Prices current as of October 2026.

Should You Use New Relic AI Monitoring for LLM Observability?

I recommend New Relic AI Monitoring for teams already using New Relic APM or OpenTelemetry who need LLM latency beside service and database latency. Model inventory adds cross-model errors and token usage, and agent views cover LangGraph, Strands and AutoGen workflows.

There is no separate AI SKU, but the deployment stops being free once ingest passes 100 GB or a second engineer needs full access. Budget full-platform seats before expanding beyond a pilot, especially at Pro’s $349 per seat on an annual plan.

I do not recommend New Relic as your only LLM platform if you need GA evaluation or prompt management. Choose Langfuse for self-hosting, Arize for evaluation with unlimited users, or Braintrust for scoring and prompt iteration, and keep New Relic where full-stack correlation earns its cost.

Start with New Relic AI Monitoring on one free-tier service, enable ai_monitoring.enabled, and measure how many GB the prompt and response events add before turning on content capture everywhere.

FAQs

Is New Relic AI Monitoring free?

Yes, up to a point: New Relic includes 100 GB/month of ingest and 1 full-platform user free, with no separate AI Monitoring charge. That allowance covers a pilot, but every other signal the account sends shares it. Beyond it, Original ingest costs $0.40/GB, Data Plus costs $0.60/GB, and extra full-platform seats cost $99 on Standard or $349 on annual Pro.

Is New Relic AI Monitoring the same as New Relic AI?

No. AI Monitoring observes your LLM applications, while New Relic AI is the assistant for engineers, billed as Advanced Compute in CCUs. AI Monitoring uses ingest and seat billing, and no per-CCU price is published for the assistant, so keep the two budgets separate.

Does New Relic AI Monitoring support Anthropic and Gemini?

Yes, but support depends on the agent and SDK. Node.js instruments the Anthropic SDK and Google Gen AI, while Python lists the Google Gen AI SDK and also instruments Anthropic. Bedrock-hosted Claude works in Python, Node.js and Java, and the OpenTelemetry examples include Anthropic. Check exact operations, since Python 13.5.0 lost tokens on Anthropic calls through LangChain and Node.js skips multimodal Gemini inputs.

Can I stop New Relic from storing prompts?

Yes. Set ai_monitoring.record_content.enabled to false and the agent stops sending prompts and responses; token counts then need a local token-count callback. Drop filters add a second layer, but each targets one attribute of one event type, and New Relic does not guarantee complete removal. OpenTelemetry content capture is off by default.

Does New Relic AI Monitoring work without a New Relic agent?

Yes. OpenTelemetry support became generally available in August 2026, so GenAI instrumentation can send spans to New Relic’s regional OTLP endpoint. I set the service name and the api-key header alongside the endpoint. AI Monitoring must also be enabled on the account, and token visibility depends on the instrumentation version.