This weekend concluded the second Midnight Code Cup: 24 hour final round took place in Belgrade on July 4-5.
29 teams competed for the champions title, team members coming to Serbia from 20 different countries.

Credits

The champions: team taxi confidently took the lead and performed strongly throughout the whole round. A close game was played in the rest of the top places up until the end, teams superfight and itmo_1.2017 ended up taking second and third place respectively.

Final results: https://2026.midnightcodecup.org/standings/
Problems: https://2026.midnightcodecup.org/problems/
First Photos (more will be added to the same folder later): Folder
Instagram: https://www.instagram.com/midnightcodecup
X: https://x.com/midnightcodecup

Amazing Jury: Aksenov239, qwerty787788, tourist, nsychev, cdkrot, naagi, Gassa, Arpa, pashka, yulik.daniel, xiaowuc1, isaf27, mnaeraxr, aropan, Solenoid555, NALP, SergeyMelnikov.

Wonderful Sponsors: Recraft, Nebius, OpenAI, Alber Blanc, Revolut, Logical Intelligence, Joom, Pinely, GPT4Telegram, KINO.Impact.

Contest director: Sofia Tekhazheva
Org support, the coolest: lperovskaya, Karina Tinicheva and Recraft Team.

Huge thanks to everyone who helped pre-solve the problems and promote the competition!
Congratulations to the finalists and the winners!

How It Went

As well as a 24-hour contest can go! Finalists reported having lots of fun, which was our goal. Now, to the problems.

Problems

This year the problem set was gradually revealed throughout the contest. It started with problems B, E, F, G, H, K and L. After the first hour, problem A was opened. Then, problems D, E and I were revealed at each following hour. Finally, problem C was added in the evening to relax teams a little bit.

Here’s a short description of each problem:

A: Arcimboldo’s Style

Inspired by Giuseppe Arcimboldo's portraits assembled from fruits, vegetables, and other everyday objects, this problem asked contestants to recreate a target image as a collage of AI-generated cliparts. Each test provided a target image, a shared collection of small PNG cliparts (generated with Recraft.ai), and a limit N on the number of clipart placements. Contestants had to position cliparts — without rotation or scaling, with overlaps allowed and repetition permitted — on a transparent canvas of the target's size, drawing them in a specified order so the result matched the target as closely as possible pixel-wise.

Look at their beautiful creations: https://2026.midnightcodecup.org/problems/A/analysis/

And this cool solution visualization: https://youtu.be/408AX8hyTBE

Read Editorial: https://2026.midnightcodecup.org/problems/A/analysis/analysis

B: Beat the Feat

In this problem, a fictional malware corrupted WAV recordings by cutting each one into equal-sized chunks and shuffling them randomly, and contestants had to reconstruct the originals from the corrupted audio alone. The contest ran as 48 half-hour stages, one per original 40.96-second recording, each offering 10 records of progressively easier difficulty — from 12,288 tiny chunks (147 samples each) down to just 24 large ones — with every record open for submission only during its own three-minute window. The answer for each record was the permutation of chunk indices that restores the original audio exactly, so the challenge combined signal-processing heuristics for stitching audio fragments with heavy automation, since solutions had to be computed and submitted via the API under tight, rolling deadlines.

C: Capture the Flag

The contestants were asked to stretch their legs, go outside, and take a picture of their whole team in front of a flag.

D: Datacenter Imprisonment

In this problem, the magician Merlin was trapped in a datacenter doubling as a maze: an N×N grid whose empty cells form a tree, with a unique path from the top-left corner to the bottom-right.

Contestants submitted a single C++ agent, which was spawned as several isolated copies running in lockstep turns. On each turn, an agent could probe one cell, send a message to another agent, or read a message. The agents had to cooperatively explore the maze and find the full path in the fewest turns.

Watch selected round replays and compare how every team’s agents explored each maze.

E: Expedition into Space

In this problem, teams piloted a virtual spaceship through 2D space hunting for 15 hidden keys worth 160 points each. The ship was controlled through a REST API (or a web client) by setting an acceleration vector capped at magnitude 10, with stop and reset commands, and the only sensory feedback was a rate-limited endpoint returning a 1000×1000 PNG image centered on the ship. Exploration was open-ended: teams had to fly around, visually inspect what they found, and could discover "tools" along the way — strings passed to the renderer that changed how the view was drawn, unlocking otherwise invisible content. The problem combined physics-style navigation, automated image analysis, and puzzle-solving, with found keys submitted to the checksystem for scoring.

Try to find all the keys in space at https://expedition.midnightcodecup.org/ !

F: Fast & Ferrous

In this problem, teams wrote a real-time strategy bot commanding robots on a shared grid arena. Games ran continuously in ~5-minute rounds: each team started with a base and a few robots, gathered crystals to build more, and earned points by shooting enemy robots, bases, and destructible walls, with range and line-of-sight checks on fire. Strategies streamed the game state every tick over gRPC or HTTP and submitted move, fire, and build commands. Point values inflated 10% hourly, so late games mattered more, and the final score was relative to the best team's cumulative points — rewarding continuous improvement of the bot's economy, micro, and targeting throughout the contest.

G: Golf Club

It all started with participants collecting Midnight coins from sponsors. To earn a coin, you had to complete a sponsor challenge: - Play a memory game with Revolut - Vibecode a browser game with Nebius in under 30 minutes: one the player could control only through a webcam - Solve two puzzles from Recraft: a traditional image puzzle and one where you assembled backend components into an actual service architecture - Tackle a Lego challenge from Joom - Beat the Rubik's cube from Alber Blanc

Once a team had collected its coins, it was invited to play golf, as rich people do. Prompt golf, that is.

In the golf challenge, teams solved competitive programming problems. Sounds easy and boring, doesn't it? To make the problem interesting, they weren’t allowed to write a single line of code — instead of that, they could only ask a small LLM (gpt-5.4-mini by OpenAI was used) to solve this problem, and the system only accepted the code written by this model. And, as you can guess from the problem title, not only the prompt must generate the correct solution, it must also be as short as possible — just as in the code golf. 16 best teams competed in duel rounds until the final where USA2 team grabbed the win at the last second!

Can you beat their record? Try for yourself: https://golf-club.midnightcodecup.org

H: HODL

In this problem, teams played as quant traders buying and selling integer-priced virtual stocks over the 24-hour contest through a web client and API. There was no cash budget — the only constraints were a cap of 100 units per stock, no short selling, and a shared 10-requests-per-second rate limit. Profit per stock was the sum of realized money from trades and the mark-to-market value of current holdings, with each stock scored independently: the team with the highest positive profit earned the full 48 points, others proportionally. Success came down to modeling each stock's price dynamics from its observed history and automating trading strategies that knew when to buy, when to sell, and when to simply HODL.

You can observe teams’ strategies and performance at https://2026.midnightcodecup.org/problems/H/stats/

I: Interactive Mind Games

In this problem, teams competed against each other in three independent game-theory games played in synchronized one-minute rounds throughout the contest. In Unique Smallest, every team submitted a number and less popular (then smaller) numbers ranked higher; in Blotto, teams split 100 troops across 5 battlefields and dueled every other team pairwise; in Door Rush, teams picked one of six capacity-limited rooms, either locking in early for a rush bonus or staying flexible to see crowd counts before a final choice — overcrowded rooms scored zero. Later rounds weighed exponentially more (up to 10×), and each game was scored relative to the best team's accumulated total, so success required automating play and adapting strategies to opponents' observed behavior over hundreds of rounds.

See all rounds

J: Just Wait till the Midnight

A traditional midnight quiz by lperovskaya. Short and sweet, to celebrate the Midnight.

K: Krowten eht Esrever

In this problem teams faced two black-box "beasts" (hidden models), each guarding a secret key. A team could send a strictly black-and-white 256×256 PNG (one request per second per endpoint) and get back only a score from 0 to 100 indicating how much the beast "liked" the image; pleasing a beast enough made it reveal its key. What each beast preferred was unknown, so the task was pure black-box optimization: probing with crafted images, inferring the hidden scoring function from responses, and running search or gradient-free optimization to climb toward images the beast loved — with each recovered key worth 600 points.

You can also please the beast at https://krowten.midnightcodecup.org

L: Legendary Grandmaster

In this problem the participants needed to solve chess problems as fast as they can. Once the top-10 emerged they were given a unique chance to play in an online simultaneous exhibition against the legendary Ian Nepomniachtchi ("Nepo") — chess Grandmaster, two-time winner of the Candidates Tournament, and two-time World Championship challenger. This was made possible by Alber Blanc.

Watch the stream.

The checksystem was once again created by the amazing nsychev.

The biggest surprise: although agents were running wild in this contest, almost all teams stayed on the floor throughout the whole 24 hours. As we hoped, human decision making was crucial if a team wanted to beat all the other talented contestants.

We feel like we achieved our goal of creating the problems people could have fun with! Until the next one =)

See you at Midnight (next year)!
Midnight Code Cup Team