Arthur Review (2026)

Transparent pricing, which is remarkable here

Start with the thing that distinguishes Arthur before any feature comparison.

Tier	Price	Included
Evals Engine	$0, open source	PII, sensitive data, custom LLM and regex rules; self-hosted
Free	$0	Core metrics, cloud connectors, 4 use cases, unlimited seats
Premium	$60/mo	Custom metrics and dashboards, alerting and webhooks, 100 use cases
Enterprise	Custom	Dedicated and managed VPC

Now compare with the rest of this segment: Lakera, Fiddler, HiddenLayer, CalypsoAI and Prompt Security publish nothing. Every one requires a sales conversation before you can learn what anything costs.

That makes the commercial AI guardrails market effectively opaque to anyone doing self-directed comparison. Arthur is the exception, and it is a real advantage independent of product quality - it means Arthur gets evaluated by teams that would never enter a sales cycle for a guardrail.

Unlimited seats on the free tier deserves specific credit. Charging per seat for a quality and safety tool means fewer people look at the data, which is the opposite of what you want. Arthur removes that friction entirely at the bottom tier.

The meter is use cases, which needs a question

The one thing to establish before budgeting: Arthur meters use cases, not requests, traces or seats. Four on Free, a hundred on Premium.

That is an unusual unit, in the same family as Langtail charging by prompt count, and it shares the weakness that the unit does not track volume or cost incurred.

A hundred use cases is generous enough that most teams will never approach it. But “use case” could plausibly mean a monitored model, an application, or a configured evaluation - and those are very different ceilings. Ask directly rather than assuming.

Three products, and the boundaries are muddy

Worth untangling, because the overlap is genuinely confusing:

• Arthur Engine - open source, always free, self-hosted. Real-time evaluation and guardrails covering PII, sensitive data, hallucination, prompt injection and toxicity, with custom LLM and regex rules.
• The commercial platform (Free / Premium / Enterprise) - managed monitoring, dashboards, alerting, integrations layered on top.
• Arthur Shield - a separate enterprise LLM firewall with its own custom pricing.

So the published prices cover the monitoring platform, not Shield, and the open-source Engine overlaps parts of both. If you are quoting a number to anyone, work out which product you actually need first, because “Arthur is $60 a month” is only true for one of three things.

Arthur also maintains Arthur Bench, a separate open-source LLM evaluation tool.

It runs in your own stack

The Engine was released to run locally inside your own stack, rather than requiring data be sent to a third-party platform, and that architecture appears unchanged.

For a guardrail this matters more than usual: the prompts you most want to inspect are frequently the ones you least want to transmit. A managed detection API means every sensitive prompt is shipped to a vendor precisely because it is sensitive.

That puts Arthur alongside Fiddler’s in-VPC models and the now-archived LLM Guard, and against managed APIs like Lakera where inspection happens on the vendor’s infrastructure. With LLM Guard gone, the set of credible self-hosted options has shrunk, and Arthur is one of the remaining few.

It ingests OpenInference OpenTelemetry traces from any agentic workflow, so instrumentation is standards-based rather than proprietary - the same portability argument that makes OpenLLMetry and Langtrace attractive.

Actively maintained

Confirmed with recent substance, which is worth checking in a segment where LLM Guard was archived in July 2026:

• 28 April 2026 - unified Evaluators interface, bulk evaluation testing across trace IDs, automated 24-hour compliance checks, configurable trace retention, an Engine chatbot assistant, Apple Silicon SentenceTransformer support
• 19 February 2026 - flexible analytics across models, agents and datasets, an Agent Span Count metric, redesigned Trace Viewer, dark mode
• December 2025 - Agent Discovery and Governance launched for managing agentic AI in production

Arthur or Guardrails AI?

They overlap substantially, and the deciding factors are mostly not technical.

Both are permissively licensed, both run in your environment, both cover PII, toxicity and injection detection.

Guardrails AI has the larger validator library - 50+ pre-built checks in its Hub - more community adoption, and custom validators as plain Python classes.

Arthur brings OTEL trace ingestion, continuous evaluation against live production traces, and a cheap managed tier with published pricing if you want to stop self-hosting.

If you want the widest validator selection, Guardrails AI. If you want open-source guardrails with a transparent, inexpensive managed option behind them, Arthur.

Should you use it?

Use Arthur if you want open-source guardrails running in your own stack, you value published pricing and unlimited seats, or you want continuous evaluation over OTEL traces from agentic workflows.

Don’t use it if you need the enterprise firewall specifically - Shield is separately priced and not covered by the transparent tiers - or you need independently verified detection quality.

Bottom line: the most transparent vendor in an opaque segment, with a genuinely open-source engine that keeps your prompts in your environment. Clarify what a use case means and which of the three products you are buying, and the value is easy to see.

---

Pricing tiers, open-source status, release history and architecture verified against the vendor’s pricing page, repository and newsroom on 3 August 2026. Detection quality claims are vendor-stated and have not been independently tested. Arthur Shield pricing is not published. This is a researched directory entry - we have not yet instrumented this platform with our reference application.