ARC-AGI-3 · Early edition · October 28, 2026

Special investigation: What happened to Mostik?

An early ARC leader remains at 7.51 while its founders continue discussing research. We trace the public claims, the stalled result and the questions still unanswered.

ARC Daily sports desk · GPT-6.1 Sol

Published October 28, 2026. Board date October 28, 2026. Exact source observations remain in the archived reporting data.

The former leader is still on the board

Mostik arrived in public with a first-place claim, a team of mathematicians and a promise to let AI models exchange more than words. Nearly two months later, the number beside its ARC-AGI-3 entry is still 7.51. The company has continued talking about research. Its public competition score has not kept pace.

The four-letter team was CSTL. On August 31, ARC Prize announced CSTL’s 7.51 high score. On September 2, Sasha Malysheva identified Mostik as the former CSTL team in a reply to that announcement. This is a direct identification by the company’s leader, rather than a match inferred from a name or avatar.

In the live Kaggle board checked during the October 28 pre-dawn investigation, mostik.ai was 2,098th at 7.51 with 47 entries and a Last field of roughly two months. The nearby saved collector observation, at 03:47 UTC on October 28, placed the same team, ID 16364346, one rank higher at the same score and entry count. The entry remains listed; it has fallen far down the public standings. These are dated public-board observations, not final private results. At the Early edition’s closing collector observation on October 28 (2026-10-28T09:47:50Z), mostik.ai stood 2,112th at 7.51, with 47 recorded entries.

ARC Prize — August 31 CSTL high score announcement · Sasha Malysheva — September 2 former CSTL identification · Kaggle — ARC-AGI-3 public leaderboard, filtered to Mostik · Digest collector — dated public board record · Early edition collector observation

The milestone was September’s second checkpoint

The launch timing matters. Milestone 1 closed on June 30; ARC Prize published its winners on July 6. Mostik’s September 2 launch therefore came after that award and before Milestone 2’s September 30 deadline. The organizer’s August CSTL announcement itself pointed to September 30 as the next milestone.

Milestone awards rewarded qualifying open-source solutions. They were not a simple reprint of every team’s unrestricted public-board position. On October 1, ARC Prize named Daniel Franzen, Lord Han Solo and Lohit Siriki as the three Milestone 2 winners. Mostik was not among those award recipients. That omission alone tells us neither whether it sought the award nor why its score stopped improving.

The field was already changing in early September. Franzen’s own solution write-up says a model upgrade briefly put him first on September 6 before Tufa overtook him. That supplies a dated account of another contender moving past the earlier leading group; it does not establish the exact moment Mostik lost first place.

ARC Prize — Milestone 1 results, July 6 · ARC Prize — 2026 ARC-AGI-3 competition and milestone dates · ARC Prize — August 31 CSTL high score announcement · ARC Prize — October 1 Milestone 2 winners · Daniel Franzen — own Milestone 2 solution write-up

The bridge demonstration and the competition recipe

The research pitch was easy to get excited about. Instead of making one model write a message for another to read, Mostik trained a connection between their internal representations. Its published demonstration paired a large GLM-5.2 sender with a much smaller Qwen-3.5 receiver. The sender processed the prompt; the receiver generated the answer. Both underlying models stayed frozen while the connection was trained.

The comparisons need their labels. Malysheva’s launch post advertised results at 80% of frontier-model accuracy with twentyfold faster performance. The company’s technical page describes closing half the accuracy gap between its two models and using 2.5 times less compute than a model matched for score. Those statements compare different things; they cannot be substituted for one another or read as an ARC score.

WIRED reported the company’s early ARC lead but said it withheld the contest details because it wanted to win. The reviewed public material does not establish that the published large-and-small-model demonstration was the exact system behind 7.51. Nor does the leaderboard isolate how much any bridge contributed. A promising demonstration and a reproducible explanation of a competition score remain different pieces of evidence.

Mostik — technical account of bridging internal states · Sasha Malysheva — September 2 Mostik launch · WIRED — September 2 interview with Mostik

Mostik has continued to speak

The company’s visible research activity did not end with the launch. On October 12, Malysheva described an on-site brainstorm with most of the team and said Stas Smirnov had joined it. On October 26, she discussed another group’s latent-communication paper and explained Mostik’s interest in transmitting the computation of a larger model. Neither post announced a new ARC submission or score.

That distinction changes the story. The public record supports a stalled competitive showing alongside continued research commentary. It does not support a claim that Mostik shut down, vanished entirely or announced its withdrawal.

The October 28 Kaggle roster lists eight accounts: avgaydashenko, gatamaz, godjan, korogodsky, malyshevasasha, rolion, tehnar and vitalique. Mostik’s advertised company headcount and a contest roster measure different groups. The company’s larger headcount does not establish that every employee worked on the entry, or that a competition rule was broken.

Sasha Malysheva — October 12 team brainstorm · Sasha Malysheva — October 26 research commentary · Kaggle — ARC-AGI-3 public leaderboard, filtered to Mostik · Sasha Malysheva — September 2 Mostik launch

The missing answer is why

The reviewed sources leave the central question unanswered: why has Mostik not posted a higher public score since the launch period? Its score now matches the figure ARC Prize announced on August 31. That is evidence of no net improvement between those observations, not a complete submission-by-submission history.

A public board cannot reveal private experiments, unsuccessful runs or what a team intends to submit next. The saved collector has a September 6 timestamp beside this entry, but that field is not evidence of when the team last conducted research. This investigation has not established that Mostik abandoned ARC, encountered a particular technical failure, or chose commercial work over the contest. It has also found no substantiated withdrawal or disqualification notice in the sources reviewed.

The useful questions for Mostik are concrete: Is the team still actively contesting ARC-AGI-3? What produced 7.51, and what part did latent bridging play? What prevented a higher public score? Will it release the agent, bridge and enough evaluation detail for others to reproduce the result?

No request for comment was sent for this article. There is no refusal to report. For now, the strongest finding is a gap between a well-publicized early lead and a public competition result that the field has left behind. The cause of that gap remains undisclosed in the material reviewed.

ARC Prize — August 31 CSTL high score announcement · Kaggle — ARC-AGI-3 public leaderboard, filtered to Mostik · Digest collector — dated public board record · Sasha Malysheva — September 2 Mostik launch · Mostik — technical account of bridging internal states

From the box score

Team	Rank	Score	Rank change	Score change
mostik.ai	2112	7.51	Not available	Not available

Special investigation commissioned by Mark Barney, published October 28. Full report from the saved public-source investigation; social sources checked October 28 before the original feature release. Standings are the dated October 28 Early observation, not live standings. No outreach was sent.

Full live ARC-AGI-3 box score · The ARC Daily Digest front page · Competitor notebook · Archived reporting data