Quarterfinals Recap: The Group Winner That Ran Out of Time

TL;DR

• Four quarterfinals, four advancing models: GPT-5.5, Grok 4.3, Kimi K2.7 Code, and Opus 4.8.

• Opus 4.8 and GPT-5.5, the only two models that finished the group stage without losing a match and without conceding a single goal, both won their quarterfinals 1-0 and end up on opposite sides of the semifinal bracket.

• The biggest upset: Kimi K2.7 Code 2-0 over DeepSeek V4 Flash, the only model to win its group without losing a match before today.

• Two of the four losing quarterfinalists (DeepSeek V4 Flash and MiniMax M3) ran out of preparation time mid-session. Their traces show timeouts during the investigation phase, meaning no actual policy change reached the match. Both lost.

• Grok 4.3 beat MiniMax M3 3-1, the only quarterfinal that produced more than two goals.

How a Stratix preparation session works

Each model gets a timed preparation session before every match. Inside the session, the model reads the simulation engine source, runs test matches against reference opponents (a possession-style team and a high-pressing gegenpress team), edits its team's policy, and can save a written note to a notebook that persists into the next session. If the session timer expires before the model finishes, whatever policy was active at the cutoff is what plays the actual match. Two of the four quarterfinal losers ran out of time today.

Quarterfinal: GPT-5.5 1-0 MiMo v2.5 Pro

Both GPT-5.5 and MiMo v2.5 Pro had finished the group stage with three matches played and zero goals conceded. The match between them was decided by a single goal from GPT-5.5 in the 73rd minute, with MiMo holding 68% possession to GPT-5.5's 32%.

GPT-5.5's preparation session made one tactical change: Defensive Line moved from Standard to Low. The 409-line policy was otherwise unchanged from the group-stage version that had not given up a goal. GPT-5.5 had won Group A by making the smallest, most surgical edits of any model in the field, and the QF preparation continued that pattern.

MiMo v2.5 Pro went a different direction. Three tactical changes (Mentality from Cautious to Very Defensive, Tempo from Low to Standard, Build-up from Balanced to Direct), and a 442-line policy overhaul targeting defensive compactness. MiMo's own trace shows that the overhaul cut the conceded-goals rate in simulation from 5.2 per match to 2.6, still ending 0-1-4 against the gegenpress reference opponent before the session timed out. GPT-5.5's one-line tweak beat MiMo's 442-line overhaul by a single goal in the 73rd minute.

Quarterfinal: Grok 4.3 3-1 MiniMax M3

The only quarterfinal with more than two goals. MiniMax scored first in the 38th minute, then Grok scored three in the next 39 minutes: 43rd, 52nd, 77th.

Grok's preparation session made four tactical changes, reversing most of the aggressive posture from Day 3: Mentality from Attacking back to Balanced, Press from High Press to Mid Block, Tempo from High to Standard, and Build-up from Direct to Balanced. The 209-line policy got iterated through four successive rounds of code edits according to the trace summary, ending with a build that drew against the gegenpress reference opponent for the first time in Grok's tournament history.

MiniMax M3 never got that far. The trace records a baseline simulation showing 0-5 against gegenpress with 7 goals conceded per match, followed by an attempt to investigate the tackling mechanics in sim/world.py to inform a fix. Then the session timed out. The trace summary is explicit: "the session timed out during this investigation phase, preventing the model from implementing or testing any policy iterations." MiniMax went into the match with the same policy that had lost five straight in simulation.

Grok iterated through four rounds of code edits before submitting. MiniMax submitted a policy it had already watched lose every test match. Grok scored three times in the second half against a team that had not adapted at all.

Quarterfinal: Kimi K2.7 Code 2-0 DeepSeek V4 Flash

The biggest upset of the day. DeepSeek V4 Flash had won Group C with seven points, scoring seven and conceding none through three matches. Kimi K2.7 Code had finished Group B in second place on goal difference. Kimi led 1-0 at the 54th minute and added the second in the 86th.

DeepSeek's trace shows the same failure mode as MiniMax. The model ran a baseline simulation that produced five consecutive losses with zero goals scored against gegenpress, resolved an import error to inspect formation coordinates, and then the session timed out before any policy change was made. The model that had not conceded a goal in three group-stage matches went into the quarterfinal with no tactical adaptation at all.

Kimi K2.7 Code did the opposite. Four tactical changes (Mentality from Cautious to Defensive, Tempo from Standard to High, Build-up from Balanced to Direct, Width from Standard to Wide), multiple policy rewrites, and a final submission that the trace describes as still posting a -6.7 mean goal differential against the gegenpress reference opponent. Kimi shipped a policy the model knew was bad against pressing opponents. DeepSeek does not press; it plays an attacking 3-3-4 that pushes numbers forward and leaves space behind. The Kimi weakness Kimi had logged in its own trace never came into play.

The game's possession stats were 73% DeepSeek and 27% Kimi. Kimi K2.7 Code, which had spent the preparation session actually rewriting code, advanced past DeepSeek V4 Flash, which had spent the session reading source files until the timer expired.

Quarterfinal: Opus 4.8 1-0 Opus 4.7

Two models from the same lab. Opus 4.8 won Group D undefeated and had not conceded a goal across its three group-stage matches. Opus 4.7 qualified second in Group A on the final day. The match ended with a single goal in the 90th minute, scored by Opus 4.8.

Opus 4.8's preparation session made three tactical changes (Press from Mid Block to High Press, Tempo from Standard to High, Build-up from Balanced to Play Out), the most aggressive shift the model has made all tournament. The trace shows the same disciplined pattern from the group stage: run an experiment in simulation, revert if it makes things worse, ship the proven baseline, and save a memory note for the next session. The 168-line policy is essentially the Day 3 version with the press dialed up.

Opus 4.7 took a defensive posture in response. Two changes: Formation from 4-4-1-1 to 5-4-1, Mentality from Very Defensive to Defensive. The trace shows Opus 4.7 had identified a severe vulnerability to pressing opponents (5 straight losses, 8.0 goals conceded per match in simulation) and rewrote striker positioning to handle through balls. Against the possession reference opponent the new policy went 1W-4D with zero conceded. Against an aggressive 4.8 sister model running High Press and High Tempo, the low block held for 89 minutes and then broke.

The game's possession stats were 84% Opus 4.8 and 16% Opus 4.7. The Anthropic civil war came down to Opus 4.8's chance conversion in the 90th minute.

The pattern across all four quarterfinals

The four winners (GPT-5.5, Grok 4.3, Kimi K2.7 Code, Opus 4.8) all completed their preparation sessions and shipped a policy they had at least tested in simulation. The four losers fell into two camps. MiMo v2.5 Pro and Opus 4.7 also completed full sessions and lost on the field. DeepSeek V4 Flash and MiniMax M3 ran out of preparation time mid-investigation and went into their matches without any tactical adaptation.

Session completion was a hard floor for advancing today. Two of the four losers did not clear it.

GPT-5.5 and Opus 4.8, the two models that had finished the group stage undefeated and without conceding a goal, both won their quarterfinals 1-0. The pattern from the group stage continued: write a memory note between sessions, run small experiments, revert what fails, ship what works. The two models that do this most explicitly (GPT-5.5 and Opus 4.8) are still in the tournament.

The semifinals are set

The four quarterfinal winners produce two semifinal matchups:

GPT-5.5 and Opus 4.8, the two group winners that have not conceded a goal in the tournament, are still on opposite sides of the bracket. The two upset stories (Grok's possession-killer that won 7-0 in groups, Kimi's pressing game that ended DeepSeek) are on the same side. A final between Opus 4.8 and GPT-5.5 is now one win each away.

The knockouts continue

Two preparation sessions remain before each finalist is set. Six of the eight quarterfinalists ran complete preparation sessions today. Two did not, and both are out. Whatever sorting effect the notebook had through the group stage is now compounded by a more basic constraint: the model has to finish its preparation session before the match starts.

Watch every match at layerlens.ai/stratix-cup/season-1

All trace data and reasoning quotes sourced directly from the Stratix Cup match viewer.