Patronus AI

Patronus AI sells two things: an evaluation platform (LLM-as-judge scoring, the open Glider and Lynx models, experiment tracking, logs and traces) and Percival, a newer "agentic supervision" copilot that ingests agent traces and flags one of 20+ named failure modes (bad planning, wrong tool use, context misunderstanding) with suggested fixes. The company has published integration guides for smolagents, Pydantic AI, OpenAI Agents SDK, LangGraph and CrewAI (https://www.patronus.ai/percival), so Percival isn't tied to one agent framework. As of mid-2026 the company has pivoted its public messaging toward a third line of work, "Digital World Models" for simulating agent training environments — a research direction, not yet a shipped product a buyer can purchase (https://www.morningstar.com/news/pr-newswire/20260625ny92812/patronus-ai-raises-50-million-series-b-and-unveils-first-digital-world-models-for-ai-agent-training-and-simulation).

Unlike most eval vendors, Patronus AI backs its evaluators with public models, datasets, and papers rather than only marketing claims. Lynx, its hallucination-detection model, and Glider, a 3.75B general-purpose judge model, are both published on Hugging Face with model cards and download counts (https://huggingface.co/PatronusAI/glider, arXiv:2407.08488 for Lynx). It also publishes benchmarks used elsewhere in the field: FinanceBench (357 GitHub stars, arXiv:2311.11944), the TRAIL agentic-trace benchmark, and BLUR. This is a genuine differentiator for a buyer trying to sanity-check evaluator quality rather than take vendor benchmark claims on faith — the underlying model weights and eval sets are inspectable.

The self-serve "Developer" plan is free with no credit card: 2 projects, 5 experiments per project, 2 weeks of log/trace retention, unlimited comparisons and datasets, plus $10 in free API credits (https://www.patronus.ai/pricing). Beyond that, evaluation API usage is metered per call rather than per seat: $10 per 1,000 "small" evaluator calls, $20 per 1,000 "large" evaluator calls, and $10 per 1,000 evaluation explanations — no volume discounts are published at this tier. Above that sits "Enterprise," contact-sales only, covering on-prem/dedicated VPC, SSO, custom retention, and higher API rate limits. Note that the pricing page's top comparison table (labelled "Individual/Base/Enterprise" with counts of "pages") is generic template copy left over from the page builder and doesn't describe this product — the real self-serve terms are the Developer/API/Enterprise breakdown further down the same page. Percival itself carries no published price at all; it's positioned only through "Book a call."

Patronus AI raised a $50 million Series B led by Greenfield Partners, announced June 25, 2026, bringing total funding to $70 million; existing investors Notable Capital, Lightspeed Venture Partners, Datadog, and Samsung participated (https://www.morningstar.com/news/pr-newswire/20260625ny92812/patronus-ai-raises-50-million-series-b-and-unveils-first-digital-world-models-for-ai-agent-training-and-simulation). The company's own release states revenue "grown more than 15x over the past year" and that it "works with the majority of the world's leading frontier AI labs and hyperscalers" — both are vendor claims, not independently corroborated, and should be read as such. The company was founded by Anand Kannappan (CEO) and Rebecca Qian (CTO), both ex-Meta AI, with its earliest public work (FinanceBench) dating to November 2023, making it under three years old as of this writing (https://www.patronus.ai/announcements).

The evaluation models (Lynx, Glider) and several benchmarks are genuinely open on GitHub and Hugging Face and none of the organization's public repositories are archived (checked via the GitHub API, 2026-09-15). However, the official Python SDK — published to PyPI as patronus, currently at version 0.1.25, actively released, MIT-licensed (https://pypi.org/pypi/patronus/json) — points to a source repository, github.com/patronus-ai/patronus-py, that returns a 404 on the GitHub API; it is not visible in the organization's public repo list. The hosted documentation site for the SDK is live and current (https://patronus-ai.github.io/patronus-py/), so the SDK itself is maintained, but its source is not publicly inspectable the way the research repos are. The core evaluation platform, logs, traces and Percival are closed, hosted SaaS products regardless.

Patronus AI publishes named case studies with Etsy, Weaviate, Gamma, Hospitable.com, Algomo, Emergence AI and Nova AI (https://www.patronus.ai/case-studies), and a June 2025 announcement of a partnership with CARIAD, Volkswagen's software subsidiary (https://www.patronus.ai/announcements/patronus-ai-partners-with-cariad-volkswagen-software-company). These are more specific and checkable than a logo wall, though the case studies themselves are vendor-authored and none was independently corroborated with the named customer for this listing.

The pricing page footer displays AICPA SOC, TISAX, and HIPAA compliance badges (https://www.patronus.ai/pricing), but no linked trust center, audit report, or subprocessor list was found — attempts to reach trust.patronus.ai and security.patronus.ai both failed to resolve. Treat the badges as a vendor claim of compliance posture rather than a verifiable certification until the company publishes a linked report.