Skip to content
AgentsIndex

Patronus AI

Research lab selling RL environments and post-training data, scored by training open models on its own tasks.

Best for
Frontier AI labs that need custom training data, tasks, and simulated/RL environments to train and stress-test agents -- not for teams wanting a self-serve LLM-evaluation tool, since that product line is no longer the company's focus.
Look elsewhere if
  • you want to try it without paying
  • you need a public API

What we checked

Pages last read
Facts checked
4 facts

From 4 sources

AI answers
Too few read

58 buyer-question answers read, November 2026

At a glance

Categories
Agent tools and integrationsAI data services

For developers

Works with
Hugging Face, NVIDIA, MongoDB

Patronus AI in our research

All research

Study

AI Tool Pricing Index, November 2026

Pricing

Contact sales

82 of the 361 tools we checked ask you to contact sales for a price.

Edition price, checked 12 November 2026.

Pricing pages of the AI tools, November 2026 editionas of n = 361 tools checked

What is Patronus AI?

Patronus AI is a research lab that builds training data, tasks, and simulated/reinforcement-learning environments -- including a "digital world model" -- for frontier AI labs training agentic models. Before a 2026 pivot it sold AI evaluation products (Lynx for hallucination detection, Glider as a small evaluator LLM, Percival for agent-trace analysis); those are no longer presented as its current offering. Pricing is not published -- its pricing page says only to contact Patronus to discuss training needs, project scope, and pricing.

Patronus now sells RL environments and post-training data

The homepage calls Patronus AI "A behavioral research lab for AI" that studies "where models fail and how to improve them." What it sells is reinforcement learning environments in five named domains: alignment, coding, knowledge work, science and computer use. On the domains page, an environment is "a sandbox with the tools an agent needs, a task and a reward that scores the result," assembled from parts such as "a codebase or simulated apps, mock websites and the data behind them." Buyers can take parts on their own: "you can buy parts of an RL environment on their own."

The coding domain page says its environments "start from real repositories, with their file tree, commit and pull request history, tests, docs and run commands," and that "Experts review every task before it is used." The post-training data page adds a second service: "We can also run expert pairwise reviews of agent transcripts."

Every purchase is scoped by hand, and no price is published

No rate card exists, and /pricing now redirects to the contact page. That page opens with "We take on a limited number of new partners each quarter." The buying FAQ describes the process: "Start with a note through our contact page with the domain and what you need, and our team will follow up to scope the work."

The same FAQ says Patronus "has a set of ready-made RL environments, but most of our work is custom." Delivery time "varies with scope. We usually send samples first and go from there." Two terms that matter to a lab are also left to the contract. Data ownership "is set in each agreement," and running on a buyer's own harness is "possible," confirmed only during scoping.

The quality evidence comes from Patronus's own training runs

The post-training data page leads with "7 to 12 points" of "gain in success on held-out tasks at RL step 40." Patronus trained "an open 27B model on two of our task families, computer use in GitHub and game engine development." GitHub tasks went "from 35.6% to 47.8% with SFT then RL" and game engine tasks "from 64.3% to 73.1% with RL alone."

The page states its own limits plainly: "The sets were small, 46 held-out GitHub tasks and 54 game engine development tasks," and RL alone came within "about 2 points" of SFT then RL, "which is within the noise." It lists five controls. Near-duplicate tasks are kept on one side of the split, every held-out task runs eight times, and 50 passing runs were reviewed by hand, of which "48 were genuine solves." No outside party has reproduced these numbers on the pages read.

The evaluation platform and its API are retired

Patronus was known for an LLM evaluation platform. Its documentation site now carries a retirement notice: "The LLM evaluation platform documented on this site is retired and no longer offered. Its evaluators and evaluation API are retired."

The evaluation work survives as research. The research page still lists Lynx ("An Open Source Hallucination Evaluation Model"), GLIDER and FinanceBench beside newer agent work: TRAIL, MEMTRACK, DETOUR, reward hack detection in code environments, HalluWorld, SpeedrunBench, GlobeBench and a masked diffusion world model paper for agentic RL. Several carry arXiv numbers or NeurIPS, ICML and ACL venues. Some releases are public.

Founded in 2023 and backed by $70M across three rounds

The company page names founders Anand Kannappan (CEO) and Rebecca Qian (CTO) and says Patronus "was founded in 2023." Staff come "from Meta Superintelligence Labs, Amazon AGI, Apple and Google." It lists three rounds: a $3M seed in September 2023 led by Lightspeed Venture Partners, a $17M Series A in May 2024 led by Notable Capital, and a $50M Series B in June 2026 led by Greenfield Partners, with Datadog and Samsung among the participants.

A separate site, data.patronus.ai, runs the supply side.

Where it fits, and where it does not

The buyer is a team training models: post-training and RL researchers who need tasks, graders and simulated apps in a specific domain and can spend weeks scoping a custom build. The contact form offers "Working with Patronus" and "Selling or licensing data" as its main routes, which matches that audience.

It does not fit a team that wants to score its own LLM app or agent traces. That product is retired, and older reviews of Patronus as an evaluation tool describe something it no longer sells. A team that needs a published price, a self-serve trial or an API will find none of them here.

Patronus AI price and terms

Terms

Free to start
PartlyLimited free useOnly the published research datasets and benchmarks (for example FigmaTrace) are free
The terms in full

the RL environments are sold through scoped engagements, and the site describes no free plan or trial for them.

For AI agents

Developer docs

Patronus AI's pricing page

Pricing page, captured

patronus.ai/domains, checked

What we checked on Patronus AI

YesPartlyNo

What it costs to startYes

What the price counts

Commercial access is sold as ready-made or custom RL environments, with requirements scoped after buyers make contact.

Start with a note through our contact page with the domain and what you need, and our team will follow up to scope the work.

patronus.ai/domainsChecked

What it costs to startNo

Free to use

Buying an environment starts by contacting Patronus to scope the work, with no self-serve trial published.

On the page: Questions about buying environments, 'How do I buy RL environments from Patronus AI?'

patronus.ai/domainsChecked

How you build with itNo

Public API

The former public evaluation API has been retired; the present site offers research and custom environments instead.

Its evaluators and evaluation API are retired.

docs.patronus.ai/docs/evaluators/concep…Checked

How you build with itPartly

Open source

Research artifacts are released publicly, named one by one: the FigmaTrace trajectories and a fine-tuned model on Hugging Face. The environments sold to buyers are not published.

We are open sourcing the dataset, with 2883 training and 586 evaluation trajectories, along with our best performing Qwen3.8-27B model on Hugging…

patronus.ai/blog/figmatrace-a-comprehen…Checked

Patronus AI alternatives

Questions about Patronus AI

How much does Patronus AI cost?
Patronus AI gives its price on request. From patronus.ai/domains, checked 12 November 2026.
Does Patronus AI have an API or an MCP server?
Patronus AI has no API. From docs.patronus.ai/docs/evaluators/concepts, checked 30 October 2026.
Is Patronus AI open source, and can you host it yourself?
Part of Patronus AI is open source. From patronus.ai/blog/figmatrace-a-comprehensive-training-dataset-for-figma-design-workflows, checked 28 October 2026.
What have we checked on Patronus AI?
4 facts, under these headings: What it costs to start, how you build with it.
How do we check Patronus AI?
We hold 4 facts on Patronus AI, quoted from 4 pages. Facts and price last checked 12 November 2026. Last change found 21 March 2026.

How we check every listing

Patronus AI price history

We have read the price once so far, on 12 November 2026: price on request.

PricePlanFirst seenLast seenReadingsSource
Price on request12 November 202612 November 20261Pricing page, read 12 November 2026

From our research on 12 November 2026. A date is when we read the page, not when the price changed.

Similar in Agent tools and integrations

See all in Agent tools and integrations

Is Patronus AI yours?

Free, and the page then carries the verified owner mark. You confirm it with an email at patronus.ai. Claim this listing

Report an error if a fact is wrong.