Is Galileo Luna-2 Enterprise-only?

The Evidence

Galileo's Luna-2 overview says it plainly: "Luna-2 is only available in the Enterprise tier of Galileo." The release note that introduced the models, dated 18 June 2025, read "Luna-2 is now available for Enterprise Customers". After the product became Splunk Agent Observability on 7 August 2026, Splunk's copy of the docs kept the line: "Luna-2 is only available in the Enterprise tier."

The pricing page is consistent with that. It does not mention Luna-2 on any plan, and it lists low-latency dedicated inference servers and real-time guardrails under Enterprise only. Galileo's Protect docs explain why the gating matters beyond cost: runtime protection "either requires Luna-2 on the enterprise tier of Galileo, or custom code-based metrics", because a check on the hot path needs a fast model.

What Luna-2 Is

Luna-2 is Galileo's family of small language models built for one job: running evaluation metrics fast and cheaply enough to score every trace rather than a sample. The docs describe 3B and 8B variants fine-tuned from Llama, a listed rate of $0.02 per million tokens, and latency that depends on hardware and request size - about 15 milliseconds for a small request on H100-class GPUs, and over a second for a 100,000-token input. Luna-2 "works with most of the out of the box metrics" and supports custom LLM-as-a-judge metrics.

Two related pieces are Enterprise as well. Luna Studio, "the self-service fine-tuning web app for Luna-2 custom evaluators", is part of the enterprise tier and is deployed into your own cluster or cloud. And although open-source Agent Control lists Galileo's Luna among its pluggable evaluators, the controls are free and the Luna-2 models behind that evaluator are not.

Luna-2 gives Galileo customers a vendor-maintained judge model tuned for its evaluation metrics at a low published token rate. It is most relevant to Enterprise teams scoring large volumes of production traces, where general-purpose judge models become a meaningful cost and latency burden.

Other Ways to Judge Cheaply

Without Luna-2 - on Galileo Free or Pro, or on any other platform - the cost of LLM judging comes down through four levers, roughly in this order.

1. Code checks first. Many "judgments" are not judgments: did a tool call error, did the agent exceed a step budget, touch a file it should not, return valid JSON. Deterministic checks cost nothing per call and never disagree with themselves. Keep the model for semantic questions such as groundedness or tone.
2. Sample. Judging 10% of sessions cuts judge calls tenfold, and a recurring failure still shows up in the sample. Oversample what the code checks flag, and where your platform lets you, key the sample on a session identifier so a rerun lands in the same bucket.
3. A smaller judge you host. An open-weight model on your own GPUs trades per-token fees for infrastructure. It pays off at high, steady volume, and it needs calibrating against human labels before you trust its scores.
4. Shorter prompts. Give the judge the part of the transcript its criterion needs - the final answer and retrieved context for groundedness, the tool events for tool use - rather than the whole session.

Reduce LLM judge cost works through the arithmetic, and LLM judge vs code-based evaluator covers where the line between the first two levers sits.

How Failproof AI Handles Judge Cost

Failproof AI does not provide a proprietary judge model like Luna-2. It supports code-based evaluations, LLM judges, prebuilt policy packs and the evaluation suite your team already uses. Deterministic checks handle failures such as tool errors, step limits and malformed output without consuming judge tokens, leaving model-based evaluation for behavior that needs semantic judgment.

The evaluations feed a broader failure-intelligence workflow. Failproof analyzes results across production sessions, groups related evidence into findings, recommends fixes and alerts owners. Teams can reduce model cost through sampling and shorter judge inputs, then use the resulting findings to decide which deterministic checks or behavioral policies should run continuously.

When None of This Matters

At a few thousand short sessions a month, a general-purpose judge may cost less than the engineering effort required to host and calibrate a smaller model. Start by measuring the token bill and judge accuracy. Sampling or self-hosted models become useful when evaluation spend is material or latency limits how much production traffic you can score.