Weekend magazine · The Arena Report
When the arena flickers
Two contests, two different headaches, one deadline that isn’t moving. What’s going wrong, who’s saying what, and what it means for Monday night.
Two contests, two different headaches, one deadline that isn’t moving. What’s going wrong, who’s saying what, and what it means for Monday night.
Saturday, October 31Reported with sources linkedEvery cause belongs to whoever named it
ARC Prize 2026 runs two contests on Kaggle at once, and both have been straining in the final stretch. In ARC-3, players say submissions have crashed with system errors or come back with a flat 0.00 after half an hour. In ARC-2, players say scores are taking more than a day to arrive. Kaggle’s staff say they are looking into the wider problems, though for now they can’t check individual ARC-3 submissions. ARC Prize says the deadline stands: Monday, November 2, 11:59 pm UTC, which is 6:59 pm Eastern.
Nobody in this story is a villain. The Kaggleverse is a huge, crowded arena, two giant contests are heading for the same deadline, and the lights are flickering. The legends are still in it.
ARC-2 teams can run on Kaggle’s L4x4 machines: four NVIDIA L4 graphics chips in one machine, with up to twelve hours per run. Kaggle’s ARC-2 page describes that competition’s L4x4 access as limited and says it uses GPU time at twice the rate of the older machines. Gemma’s competition page also explicitly offers L4x4 access; the hardware is not exclusive to ARC-2.
When the line gets long, a twelve-hour run waits its turn. Chan Kha Vu of V | A | M says the wait to start has passed twelve hours and scores are taking more than a day to come back, so a last-day entry can’t be scored in time. The ask, Chan says, is not a longer contest and not more submissions, just more time to pick the final ones. Masspeaks, in the ARC-2 top forty, started the forum thread with a proposal for what to do if the delays continue, and CPMP backed a fixed period for choosing. Kaggle’s María Cruz replied that Kaggle doesn’t plan to change the deadline or the submission process in the final days.
CPMP flagged the L4 line a month ago, and Kaggle’s Dustin answered that demand had risen sharply and that capacity had been added. Gemma’s competition page explicitly offers Kaggle L4x4 machines. CPMP pointed this out in his November 1 reply to our article, and he was right: our earlier wording gave the misleading impression that the hardware was exclusive to ARC-2. Both competitions offer that hardware.
Gemma’s discussion board supplies further evidence. On September 28, Ryan Holbrook paused rescoring because of L4x4 queues. On October 8, Dustin attributed long waits and erroneous failures to heavy L4 submission volume; Holbrook said the next day that doubling the machines had not contained demand. Dustin estimated a 33-hour queue on October 10, then about 14 hours on October 13. On October 16 he said the earlier queueing-error problem had been fixed. Those figures do not establish the queue length on October 31.
Late-October complaints continued in the same Gemma thread. On October 26, Jason Krijgsman reported a submission pending since the previous afternoon. On October 28, Starright reported scoring taking over a day and suspected renewed queue trouble. On October 31, Sawan Ruparel reported repeated submission timeouts after more than 14 hours or a day, asking whether his configuration was responsible. These participant reports establish continuing submission problems; they do not separate GPU queue time from execution time or establish the cause of each failure.
ARC-2 staff also acknowledged more sessions than machines. Both contests offer L4 hardware, and both had late-October complaints. Pressure on shared L4 capacity is a plausible explanation, consistent with CPMP’s concern. The available evidence does not establish how Kaggle divided that capacity between contests or how much of ARC’s delay came from Gemma.
Correction · November 1, 2026: We have corrected the implication that L4x4 access was exclusive to ARC-2. Thanks to CPMP’s reply; see Gemma’s Upgraded Accelerators section. We have added Gemma’s October 26, 28 and 31 participant reports alongside the earlier staff updates, including the reported improvement and fix. We distinguish the historical queue estimates from late-October reports and from our interpretation of shared capacity pressure.
ARC-3 teams can run on RTX 6000 graphics cards, with nine hours per run. Since October 22, players have reported three kinds of trouble: saves and submissions that fail with a “Kaggle error” because a big AI model won’t load; runs that show “complete” after about half an hour with a flat 0.00; and long, unpredictable waits before a run even starts.
Kaggle’s staff say bad stored data in one of their data-center regions was cleared about a week ago, that engineers are looking into the wider problems, and that for now they can’t investigate individual submissions because the team is stretched thin. Earlier, Kaggle’s Dustin said some affected runs were being rescored. Kaggle’s Walter Reade says use of the whole site is up two to three times since the ARC contests launched, driven by AI agents making contests easier to enter and by contests getting more complex.
Tufa’s Jeroen Cottaar broke out the stopwatch. Cottaar says Tufa’s submissions lose about forty minutes before their code even starts, and that a blank test submission sat about fifty minutes before anything ran. Real runs still stop at nine hours, so Cottaar reads it as time taken straight out of the game. CPMP replied that a wait before code starts is not unusual and not a problem. Cottaar answered with the blank-submission test. Two Grandmasters, one stopwatch, one polite disagreement.
Players are adapting. xz, whose one-person team is third on the board, starts playing ten minutes in whether or not the AI model is ready, and hits reset every seven minutes. Van-Phuc Huynh gives up some solving time, waiting longer and resetting more, so a run doesn’t end in a zero.
“I never saw Kaggle platform being so bad since I started kaggling.”CPMP, NVIDIA, on X, Saturday · source
“I’m not stressed, I’m sad and frustrated.”CPMP, ARC-2 forum, this week · source
“We can’t work that way.”CPMP, to Kaggle staff, after six saves and a mount error · source
“We’re quite convinced things are not back to normal.”Jeroen Cottaar, Tufa Labs · source
“We just ask to be able to have more time for selecting the final submission.”Chan Kha Vu, V | A | M, on the ARC-2 queue · source
“We also got a 0.00.”Son Pham, Team Turducken, ARC-3 · source
“Our team is looking into the broader platform issues reported in this thread.”María Cruz, Kaggle staff, ARC-3 forum · source
“Why not freeze it, disable submission?”Ivan, whose Kaggle name links to Ivan Sorokin’s account, NVARC3 · source
An ARC Prize moderator told competitors on Discord, after talking with ARC Prize’s president, that the deadline stands: Monday, November 2, 11:59 pm UTC. The moderator said the schedule has been public for months, that Kaggle’s leadership strongly advised against a change, and that ARC Prize is working with Kaggle on the slowdowns. We reported it Friday night. Kaggle’s contest pages carry the standard line that organizers can update the timeline, with any change announced on the competition website.
Each team can pick up to two final submissions; if a team doesn’t pick, Kaggle picks for it. The private leaderboard decides the final standings. Kaggle staff first said a run sent before the bell would keep scoring afterwards, then corrected that guidance on October 29. The safe plan, and ARC Prize’s advice too, is to plan for delays and have final runs finished well before the bell. The Final Hours has the rest.