New Relic AI Monitoring Review (2026)

The meter is the problem

New Relic AI Monitoring is a competent product attached to the wrong billing model for the workload it monitors.

New Relic bills on data ingested. Ingest runs at $0.40 per GB, or $0.60 with Data Plus. That is a perfectly sensible meter for APM, where a span records a method call, a status code and some timings.

LLM traces are not like that. A single trace carries the prompt, the completion, tool-call payloads and retrieved context, and routinely runs 5 to 50 KB. RAG applications logging large retrieved-context blocks sit at the top of that range.

Charging per gigabyte for the most verbose telemetry in software is a structural mismatch, and the reported consequences are severe. One agent stack ingesting 4 TB per month reportedly went from a $1,400 monthly APM bill to between $9,500 and $12,000 once AI Monitoring volumes were counted. Vendr’s 2026 data reportedly shows even 5 to 20 user deployments landing at $30,000 to $80,000 per year.

We report these as third-party figures rather than measurements we took. But they are consistent with how the meter works, and we would treat them as a warning rather than an outlier.

The third meter nobody accounts for

This one is genuinely easy to miss, and we have not seen it flagged in any comparison.

New Relic charges on three separate meters: data ingested, user seats, and compute units (CCUs) for AI features. Enabling New Relic AI or Transaction 360 starts the CCU meter.

The subtlety is what CCUs track. They measure usage by people, not by systems. Consumption scales with how much your team interacts with the AI features, not with your application traffic. Every cost model we have seen built for New Relic accounts for ingest and seats and stops there.

If you are budgeting this product, ask your account team explicitly how CCUs will be consumed under your intended usage pattern, and get the answer before you enable the feature rather than after.

Seats, and the one genuinely good part of the pricing

Meter	Rate
Ingest	$0.40/GB, or $0.60/GB with Data Plus
Basic users	Free, unlimited
Core users	$49/user/mo
Full platform users	$99 each on Standard (capped at 5), or $349/user/mo on Pro
CCUs	Metered separately once AI features are on

The free tier is genuinely generous on one axis: 100 GB per month of ingest, one full platform user and unlimited basic users. Compared with Datadog’s 40,000 LLM spans, 100 GB is a lot of room to evaluate.

Unlimited free basic users is the other real strength. In a large organisation where many people need read access to dashboards and few need to configure anything, this keeps the seat cost far more contained than per-seat competitors like LangSmith.

It monitors. It does not evaluate.

This is the second thing that should decide it for many teams.

There are no output-quality evaluation metrics. No faithfulness scoring, no relevance scoring, no hallucination detection with actual machinery behind it. No annotation workflow. No dataset curation.

New Relic’s marketing references identifying quality issues including hallucination, bias and toxicity, and critics have noted the gap between that framing and what is actually implemented. What you get is monitoring - token usage, cost, latency, prompt and response capture, model comparison. Those are useful. They tell you what your system did and what it cost. They do not tell you whether the answer was correct.

If evaluation matters, you are buying a second platform regardless, which changes the economics of buying this one at all.

The APM data model shows through

AI Monitoring is layered on APM 360, so LLM workloads inherit a data model built for HTTP requests and database calls.

That model assumes a unit of work that starts, does something, and finishes. Multi-turn conversations, tool calls and agent loops do not have that shape - their natural unit is a session, not a span. You can make it work, but you are fitting a new problem into an old container, and the ergonomics suffer. Tools designed for agents from the start, like HoneyHive, model sessions natively.

Reducing the bill costs you the data

Worth understanding before you assume sampling will save you.

Because the meter is volume, you reduce cost by sending less - truncating prompts and completions, dropping retrieved context, sampling traces. The bill falls proportionally.

The problem is that the retrieved context you dropped is usually exactly what you need to debug a RAG failure. Datadog’s tail-based sampling is smarter here: it keeps every error and slow request while dropping successful fast traffic, so you lose the boring traces and keep the interesting ones. Trimming payload size degrades every trace uniformly, including the ones you will need.

If you must run this, decide early and deliberately what you genuinely need to retain.

Should you use it?

Use New Relic AI Monitoring if your organisation is already standardised on New Relic for APM, logs, Kubernetes and SRE workflows, your LLM volume is modest, and correlation with infrastructure matters more than evaluating output quality.

Don’t use it if you run high-volume or RAG-heavy workloads where trace payloads are large, you need output-quality evaluation, or you are choosing an observability platform fresh.

Bottom line: the right answer for an existing New Relic shop with light LLM usage, and an expensive mistake for almost everyone else. The ingest meter is structurally wrong for LLM telemetry, the CCU meter is easy to miss entirely, and there is no evaluation capability at all. If you are not already committed to New Relic, Langfuse, Logfire or a dedicated eval platform will serve you better for less.

---

Pricing structure and the three-meter model verified against published documentation and third-party analyses on 31 July 2026. The bill-increase figures and Vendr deployment costs are third-party reports, not our measurements, and are flagged as such. This is a researched directory entry - we have not yet instrumented this platform with our reference application.