Pymetrics is a set of about a dozen short online games that some employers ask applicants to play early in the hiring process, usually before any interview. Since August 2022 it has belonged to Harver, a hiring assessment company, and Harver now sells it as its game-based platform. The games do not test what you know. They record how you play (how much risk you take on a balloon, how fast you press a key, how you split money with a stranger) and compare the pattern with a model built from people already doing the job well. Harver says the whole thing takes approximately 25 minutes on average. This guide walks through the games one by one, what Harver and an independent auditor say they measure, how the result is actually scored, which employers have published that they use it, how to ask for an accommodation, and how to prepare without paying for a prep pack.
The short version
| Question | Answer | Source |
|---|---|---|
| Who owns it | Harver, since the acquisition completed on August 11, 2022 | Harver press release |
| How many games | A core set of twelve; Harver's page says "12+" | Wilson et al., FAccT 2021; Harver |
| How long | About 25 minutes on average | Harver |
| What it measures | Effort, risk tolerance, decision making, attention, focus, learning, fairness, generosity, emotion | Harver |
| What the employer receives | A tier: Do Not Recommend, Recommend or Highly Recommend | BABL AI audit, July 2025 |
| Right answers | None. "There are no 'right' or 'wrong' ways of performing the exercises." | Paramount notice |
| Accommodations | Colorblindness, ADD or ADHD, dyslexia versions; other needs through the employer | Paramount notice |
What Pymetrics is, and the Harver change
Pymetrics started as an independent company. On August 11, 2022, Harver announced it had completed its acquisition of "pymetrics, the leader in unbiased soft skills assessments." The release said pymetrics had been used by millions of candidates and was compliant in "100+ countries, 30 languages." It did not say the pymetrics name would stay, but it has: Harver's software page in 2026 calls it "Harver's game-based platform," and the old login page at pymetrics.com now says only that "Pymetrics has been acquired by Harver" and points to a login.
Two practical consequences for candidates. First, the pymetrics.com candidate page that older guides link to is now a stub, so the most reliable information is now in Harver's own pages, in employer notices, and in audit documents. Second, the auditor's 2025 report calls the product "Harver's Soft Skills Platform," so if your invitation uses that name or Harver's branding instead of pymetrics, it is the same family of games.
How the system is built matters more than the brand. A 2021 peer-reviewed paper written by Northeastern University researchers together with pymetrics staff, Building and Auditing Fair Algorithms, describes the process step by step:
- The employer's high performers in the target role play the games. That "in group" typically holds 50 to 100 people.
- A model learns which gameplay features separate them from a comparison group drawn from the wider pymetrics player database.
- The model is tested for adverse impact on a separate group of players who gave demographic data, typically over 10,000 people.
- Applicants play the same games, and the model predicts which of them look like the high performers.
So there is no universal good score. The same play could rank well for one role at one company and poorly for another.
The games, one by one
Harver does not publish a list of game names, and its own page pictures only the balloon game. The list below comes from a 2026 candidate guide by My Consulting Coach, which describes twelve games, and it matches the shorter descriptions in a June 2026 guide from Truffle. Treat it as a careful outside description, not an official one. Employers can configure the suite, so you may not see all twelve.
| Game | What you do | Trait it is said to target |
|---|---|---|
| Balloon | Inflate balloons for money without popping them, and learn their color patterns | Risk taking |
| Tower | Move disks between pegs, in the style of the Tower of Hanoi puzzle | Planning |
| Money Exchange 1 | Split money with a partner character | Fairness |
| Money Exchange 2 | Earn money, split it, and judge whether splits feel fair | Fairness |
| Keypress | Press a key as many times as you can in a time window | Effort, speed |
| Hard or Easy | Choose low-pay easy tasks or high-pay hard ones with uncertain success | Effort, risk |
| Digits | Recall a growing string of numbers | Working memory |
| Stop | Press only when a target shape appears, and hold back otherwise | Impulse control |
| Arrows | Respond to arrow direction under rules that change with color | Attention |
| Lengths | Pick the longer of two nearly identical shapes, fast | Fine discrimination |
| Cards | Pick cards to earn as much as you can by spotting patterns | Pattern learning |
| Faces | Match a facial expression to a described situation | Reading emotion |
The balloon game is the one with a public first-hand description. A February 2021 MIT Technology Review piece by Hilke Schellmann describes each click on "Pump" earning five virtual cents, with everything lost if the balloon pops before "Collect." After 39 balloons her feedback read: "You stick to a consistent approach in high-risk situations. Trait measured: Risk." That mechanic is the same as the Balloon Analogue Risk Task that psychologists have used since Lejuez and colleagues published it in 2002, which fits the FAccT paper's statement that the games are "derived from peer-reviewed psychological studies."
The same paper says pymetrics had "recently started offering an additional suite of numerical and logical reasoning games." If your invitation mentions reasoning or numbers, expect some items that do have correct answers, unlike the twelve core games.
What the games measure, in the vendor's words
Harver's game-based assessment page lists nine attributes: effort, risk tolerance, decision making, attention, focus, learning, fairness, generosity and emotion. Its FAQ says the games "evaluate candidates' soft skills such as attention, decision-making, risk tolerance, and emotional intelligence," and the page claims they are "validated for fairness across gender, ethnicity, and socioeconomic status." Its software page adds that the system collects "1000s of data points" while you play.
An employer's notice describes the same thing in plainer terms. Paramount's automated employment decision tools page, last updated June 28, 2023, says the assessment is "designed to measure a series of cognitive, social, and emotional soft skills," is played on a computer or mobile device, takes about 25 minutes, and has "no 'right' or 'wrong' ways of performing the exercises."
What none of this proves is that the traits predict job performance. The FAccT auditors said plainly that they "did not investigate the ability of pymetrics' games to measure human capabilities" or whether those map to job performance, and that drawing a line from lab experiments to real work "is challenging." They asked pymetrics to publish its validation studies, which they had seen only as a confidential presentation.
How the score really works
The most useful document for a candidate is one most guides skip: the bias audit that New York City's Local Law 144 requires. Since enforcement began on July 5, 2023, an employer using an automated hiring tool on New York City candidates needs a bias audit done within the past year and has to make a summary public. Harver's latest summary we could find, prepared by BABL AI and dated July 17, 2025, is hosted by Standard Chartered. It says:
- The output is a recommendation tier, "Do Not Recommend, Recommend, or Highly Recommend," based on "a candidate's percentile rank against a validated model."
- The default line for Recommend is the 50th percentile.
- In the January to December 2024 data, the selection rate was 0.540 for women (119,242 applicants) and 0.545 for men (164,014 applicants).
Put simply, about 54 in every 100 players in that data landed on the Recommend side of the default line. That is a long way from the gatekeeper reputation the game has on some forums, and it tells you how to think about it: the games sort a large pool roughly in half for that role's model, and the employer then applies its own resume screen and interviews. The FAccT paper describes the same flow, with the employer free to "apply additional filters (e.g., resume screening)" after receiving the high scorers.
Which employers use it
Only employers that say so in a source you can check are listed here. Prep sites often name dozens of firms without evidence, and employers change vendors.
| Employer | What the source shows | Source, date |
|---|---|---|
| Paramount | Publishes a pymetrics notice with accommodations, and links its latest bias audit for NYC candidates | Paramount AEDT page, updated June 28, 2023 |
| Standard Chartered | Hosts the July 2025 Harver audit summary; its current early careers page describes a "Skills Insight Assessment" without naming pymetrics | Audit PDF; early careers, checked October 18, 2026 |
| BCG | Used it "in some offices and older processes"; "not the assumed default" in 2026 | My Consulting Coach, July 2026 (third party) |
The last two rows show the problem. Standard Chartered hosts a Harver audit while its careers page calls its first test a "Skills Insight Assessment," and a BCG prep guide says pymetrics is no longer the default there. Read your own invitation rather than assuming from the company name. If the link goes to a pymetrics or Harver domain, it is this test. If it goes to HireVue, see our HireVue interview guide, and if it is a coding test, see the CodeSignal assessment guide.
A related point on disclosure: in Greenhouse's 2026 Candidate AI Interview Report, 70% of the 2,950 job seekers surveyed "were never clearly told upfront that AI would be evaluating them." That survey was about AI interviews, not games, but the fix is the same: if a hiring step is automated, the employer's careers site or its NYC notice is where to look for how.
Accommodations
The built-in accommodations cover three conditions. Paramount's notice says applicants "will be prompted to self-identify as requiring accommodations for colorblindness, ADD/ADHD, and/or dyslexia" before they start, and that "an accommodated version of the pymetrics assessment will be administered immediately." The FAccT paper also describes built-in accommodations for color blindness and dyslexia, and translation into over 20 languages. Harver's current software page lists "Accessibility options" without detail.
For anything else, the route is the employer, not Harver. Paramount gives a phone number and an email address for "other disabilities." Practical steps:
- Choose the accommodated version on the setup screen if it applies to you. The choice is made before the first game, not after.
- If your need is not one of the three, email the recruiter or the employer's accommodations contact before you open the games, and keep the reply.
- If a physical condition makes fast keyboard or touch input hard, say so specifically, because some games ask for quick responses.
- Ask before you play, not after a session you were unhappy with.
How to prepare without paid prep packs
You cannot study for these games the way you study for a numerical reasoning test, because there is no answer key and the model differs by role. What you can control is whether your play reflects you on a normal day.
Before you start
- Read the invitation for the deadline, the device options and any extra reasoning section.
- Block about 30 minutes with no interruptions. Prep guides report that you cannot pause inside a game but can rest between them.
- Use a device you are used to. The FAccT paper notes that web and mobile versions can "elicit different behaviors," and that data may be scaled for platform differences, so comfort matters more than which device.
- Close other tabs and silence your phone. The timed games reward steady attention.
While you play
- Finish every game. The FAccT paper says that, by default, gameplay from users with more than two missing games is treated as incomplete and removed from analyses.
- Read each game's instructions fully. Several have rules that change partway, such as arrow colors.
- Play the way you would actually behave. A strategy designed to look "ideal" can only be ideal against one employer's model, which you cannot see.
- On the effort games, give real effort. Keypress and Hard or Easy are described as measuring exactly that.
Free practice, if you want it
Several games resemble classic psychology tasks that are free to try in a browser at the PsyToolkit experiment library, including the Balloon Analogue Risk Task, digit span, go/no-go and stop signal tasks, an arrow flanker task and the Iowa Gambling Task. A few minutes on each removes the surprise of the format. That is all paid practice packs can honestly offer too, since none of them has the employer's model.
After you finish
Prep guides report that a result can be shared with other employers on the platform for about 330 days and that retakes are not normally available on demand. Neither Harver nor the audit confirms those numbers publicly, so ask the recruiter if it matters for a second application.
Where this fits in an application
The games usually arrive after you submit an application and before any interview, so they come after the resume screen has done its work. Our guide to what an ATS is explains that earlier step. ApplyEze works at that earlier stage: it writes a resume for each posting, plus a cover letter when the form asks, and completes the application form on 12 named platforms. Assessments like Pymetrics, HireVue and CodeSignal arrive afterwards by email, and they are yours to take.