LLM-as-a-judge alternative

An LLM judge checks a sample. Latent checks every answer.

A frontier judge model takes seconds and about a cent to read each answer, so most teams judge a sample. Latent reads the model's activations in 33 ms for $0.02 per 1,000 answers, so it checks every one and sends the judge only the risky ones.

Start free Book a call

Time to check an answer33 ms vs 3.1 s

Cost per 1,000 answers$0.02 vs $10.61

Judge bill when Latent picks what to judge10× lower

Flags that are right, Latent picking vs judge alone93% vs 82%

A judge on every answer costs seconds and dollars

We timed Claude Opus as a judge against Latent on the same 1,498 RAGTruth answers, every checker given the source document.

Checker	Median time per answer	Cost per 1,000 answers
Claude Opus as judge	3.1 to 3.4 seconds	$10.61
Latent's reader, hosted by Latent	150 to 170 ms	See pricing
Latent's reader, on the same machine	33 ms	$0.02 on a GPU you run
Latent inside your model on vLLM	No measurable time	No extra cost

A million answers a month is about $10,600 of judge calls, and each one adds seconds. So most teams judge a sample, and the answers outside it go unchecked. How we measured

Latent reads the model that wrote the answer

A judge reads the source and the answer and then writes out a verdict. Latent reads the activations behind the answer once and generates no text, which is why it takes milliseconds. Inside your own model on vLLM, it reads activations your model already computes, so there is no second pass at all.

A judge is also a second model of the same kind as the one it checks, so it can share the mistakes it is looking for, and it can decline to answer. In our rebuild of the OpenAI agent incidents, the judge declined 4.3% of its decisions.

Keep your judge for the riskiest 10%

Teams that like a judge's written reasons can keep it and send it only Latent's riskiest answers. On the same human-labeled answers, that cut the judge bill tenfold and made more of the flags right.

93%

of flags are real failures when Latent picks the riskiest 10% and Claude Opus judges only those, for $1.06 per 1,000 answers. Opus judging every answer reached 82% for $10.61.

1,498 RAGTruth answers with human span labels. Claude Opus at its September 2026 list price. September 2026.

Set Send to a judge on the console's Policy page, and each flagged answer is queued for a judge's written analysis after it is delivered, so your reviewers get the explanation where it matters.

Latent's flags hold up against human labels

Every number below was read on answers held out of the fit. The benchmarks page lists each one with its data, sample and labels.

96%

of the answers Latent flags are real failures, when it flags 10% of other models' RAG answers.

RAGTruth held-out set: 3,498 answers to 583 prompts from GPT-4, GPT-3.5, Llama 2 7B, 13B and 70B, and Mistral 7B. Human span labels. Read by Latent's reader, Llama 3.1 8B Instruct. September 2026.

4.2×

the failure rate of random review in Latent's top 10% on finance filings, after one calibration run on 1,000 of the model's own answers. Detection rose from AUROC 0.586 to 0.851, in 453 seconds for $6.88 of judge calls.

FinQA 10-K excerpts: 187 held-out answers about 24 companies the fit never saw. Labels from a strict Claude Opus judge. Llama 3.1 8B Instruct on vLLM. September 2026.

Score every answer with one line

Install the Python SDK and turn it on at startup. Latent scores every answer from your OpenAI, Anthropic, Gemini and Ollama clients, and your calls go to your provider exactly as before.

pip install runlatent
export LATENT_API_URL=https://app.runlatent.ai
export LATENT_API_KEY=lk_...

import runlatent

runlatent.auto_instrument()  # once at startup: every answer is scored off-thread
r = runlatent.check(prompt, answer, context=sources)  # or score one answer before you return it

The quickstart walks through setup and the first scored answer.

Questions

What is LLM-as-a-judge?: Using a language model to grade another model's output, here whether an answer is supported by its sources. A judge is flexible and writes its reasons, and it costs a model call and seconds for every answer it reads.
Is Latent as accurate as an LLM judge?: On the same 1,498 human-labeled RAGTruth answers, Claude Opus judging every answer was right on 82% of its flags. With Latent picking the riskiest 10% for Opus to judge, 93% of flags were right, for a tenth of the judge bill.
Do I still need a judge?: Keep one if you want written explanations. Latent sends it only the answers worth explaining.
Does it work with GPT, Claude and Gemini?: Yes. For models behind an API, Latent's reader model re-reads each finished answer. For open-weight models on vLLM, Latent reads your own model as it writes.

Research

Latent against an LLM judge: the time and cost of checking every answer, measured.
LLM hallucination detection: methods compared: judges, NLI models, sampling, token probabilities and internal-state probes.
LLM guardrails: what Latent does with a flagged answer.