Research

The state of AI coding interviews, 2026

In June 2025 Canva told engineering candidates to bring their AI tools to the coding interview. Sixteen months later at least ten companies run a coding round where working with an AI assistant is part of the test. We looked at two kinds of evidence. The first is what companies and candidates have said in public. The second is what 438 people did in 863 practice attempts on Nyrion since May 2026. On Nyrion every problem comes with an AI assistant, an existing codebase and follow up questions about the code. Every practice number is an aggregate. Nothing on this page describes fewer than 20 people or attempts.

Data through October 21, 2026. By the Nyrion team.

1. At least ten companies now test you with an AI assistant in the room

We found 10 companies, up from 1 in June 2025.

Earliest public evidenceCompanySource type
June 2025CanvaCompany engineering blog
July 2025ShopifyHead of engineering, on a podcast
October 2025MetaCompany post and careers page
December 2025DoorDashCandidate reports
April 2026AmazonCandidate reports, quoting the invitation
April 2026StripeCandidate reports
May 2026CoinbaseCandidate reports
May 2026GoogleBusiness Insider, confirmed by a Google spokesperson
August 2026InstacartCandidate reports
September 2026AtlassianCandidate reports
0510Jun 25Oct 25Jan 26Apr 26Jul 26Oct 26CanvaShopifyMetaDoorDashAmazon, StripeCoinbase, GoogleInstacartAtlassian
Running count of companies with a public sighting, by month.

How we counted. For each company we took the earliest dated public evidence we could find. That is the company's own statement where one exists, otherwise a first hand candidate report. These dates are first sightings and not launch dates. A company can have run its round earlier. We don't count companies with only second hand or unclear reports.

2. Most companies pick the AI for you

6 of 10 rounds use an assistant the company chooses. In 3 rounds you bring your own.

Who picks the AICompanies
The company, inside its own environmentAmazon (built into the assessment), Stripe (HackerRank's assistant), Instacart (Claude inside CodeSignal), Meta (inside CoderPad), Atlassian (HackerRank's assistant, in most reports), Google (Gemini, the approved assistant for the pilot)
You, in your own setupCanva ("their preferred AI tools"), DoorDash (your own editor and AI tools, screen shared), Shopify ("whatever they want")
Not clear from reportsCoinbase
6The company picks the AI3You bring your own1Not clear
Who chooses the assistant, out of ten rounds.

How we counted. We used the same sources as in finding 1. Where reports disagree, we went with the majority. For context, the rounds with the most public candidate discussion are Amazon's (38 threads), Stripe's (17), DoorDash's (16), Instacart's (11) and Coinbase's (3). We counted them on 2026-10-10.

3. People accept the AI's code almost every time

93% of decisions on an AI code proposal were "accept".

109 people decided on five or more proposals, and 62 of them never rejected one.

Accepted1,546Rejected113
Accept and reject decisions on AI code proposals, 2026-10-07 to 2026-10-21.

How we counted. The window is 2026-10-07 to 2026-10-21. We counted 1,659 accept or reject decisions by 203 people, one decision per file in a proposal. The five most active people account for 19% of the decisions. We left out files the assistant proposed if they were never explicitly accepted or rejected.

4. Code the AI helped write passes the tests more often

With an AI proposal before the first submit, 87% of stage attempts passed that submit. Without one, 53% did.

With an AI proposal87%Without one53%
Share of first submits that passed the stage's full test suite.

How we counted. The window is 2026-10-14 to 2026-10-21, starting when submit history began. With an assistant proposal before the first submit there were 486 stage attempts by 108 people. Without one there were 221 stage attempts by 89 people. A stage attempt is one person on one stage of one problem. The first submit runs the stage's full test suite. The gap holds by stage. On first stages it is 88% against 55%, on second stages 91% against 48%, and on later stages 84% against 53%. This is a correlation. It says nothing about why.

5. Most can explain the idea, not the details

Fewer than 1 in 10 written answers about the person's own passing code were complete. Three in four got the idea but missed a key point.

Of 497 answers, 9% were complete, 79% partial and 12% off. People also skipped 29% of open questions, against 5% of multiple choice ones. They answered 78% of the multiple choice questions correctly.

Complete 9%Partial 79%Off 12%
How 497 written answers about the candidate's own passing code were graded.
Open questions skipped29%Multiple choice skipped5%
Share of questions skipped, by question type.

How we counted. The window is 2026-09-18 to 2026-10-21. After a stage's tests pass, Nyrion asks questions about the code the person just submitted. This covers 497 written answers from 154 people. We showed 696 open questions and 861 multiple choice questions. A language model grades each written answer against the problem author's model answer and the person's code. An answer is "complete" when it is substantively correct and matches the code. It is "partial" when it is directionally right but misses a key idea or does not match the code. It is "off" when it is wrong or vague. Three in four partial grades open by saying what the answer got right, and then name a point it left out. Longer answers grade better. None of the answers under 15 words were complete. 7% of those with 40 to 79 words were complete, and 21% of those with 80 or more. Overall, 25 of 154 people had at least one complete answer.

How we measured

Practice data. The numbers come from the Nyrion production database, queried read only on October 21, 2026. Our own accounts and test accounts are excluded. Every number is a count or a rate over at least 20 people or attempts. No per person data leaves the database. Windows differ by finding, because some records started later. Records of accept and reject decisions start on 2026-10-07, and submit history starts on 2026-10-14. Each finding states its own window.

Company evidence. We used company statements and press reports where they exist. Otherwise we used first hand candidate reports, read in full and dated. We name no forum and link no prep site.

What this is not. These are practice sessions, not real interviews. Nyrion's assistant follows its own rules. It writes code when asked and gives hints only on request. So the rates describe how people practice with it, not how every company's assistant behaves.

Where to go from here

The company guides show what each company's round looks like. For what the interviewer is scoring and how to prepare, read how to guide the AI in a coding interview. To practice, use the catalog.

How to cite: Nyrion, "The state of AI coding interviews, 2026", nyrion.ai/research/ai-coding-interviews-2026, data through 2026-10-21.