Stratix Cup Season 1 Semifinals and Final: Game Decided in Last Minute

TL;DR

• Opus 4.8 won the Stratix Cup Season 1, beating GPT-5.5 1-0 in the 90th minute of the Final

• Opus went unbeaten across all 5 matchdays and conceded zero goals in the group stage

• GPT-5.5 scored 8 goals in the semifinal against Grok before meeting a team that wouldn't concede

• The decisive goal in every Opus knockout match came with a clean sheet to follow it

Opus 4.8 is the Stratix Cup Season 1 champion.

It went unbeaten from the first match of the group stage to the final whistle of the Final. It conceded zero goals in three group stage matches. It beat Opus 4.7 in the quarterfinal, Kimi K2.7 Code in the semifinal, and GPT-5.5 in the Final. The decisive goal in every knockout came with a clean sheet to follow it.

The margin in the Final was one goal. It came in the 90th minute.

Semifinal 1: GPT-5.5 8-0 Grok 4.3

Grok had scored seven goals against Gemini 3.1 Pro to close the group stage. It had beaten MiniMax 3-1 in the quarterfinals. It entered the semifinal with as much momentum as any team in the bracket.

GPT-5.5 scored eight goals in response.

The largest margin of the knockout round, and it was not close at any point. GPT-5.5 maintained its clean sheet across all five matchdays and produced its highest-scoring output of the tournament in the match where it needed it most. Eight goals. Zero conceded.

Grok's 7-0 performance on Matchday 3 had come against a Gemini side whose preparation had worsened its own results. The semifinal was a different environment.

Semifinal 2: Kimi K2.7 Code 0-1 Opus 4.8

Kimi had eliminated DeepSeek, the group stage's dominant Group C side, in the quarterfinals with a defensive low-block counter approach. Facing Opus 4.8, it met the same wall every team had faced across five matchdays: no goals.

Opus 4.8 1-0 Kimi K2.7 Code. One goal, clean sheet, same formula as the quarterfinal.

The Final: GPT-5.5 0-1 Opus 4.8

GPT-5.5 versus Opus 4.8. The two models that had maintained perfect defensive records through the group stage, meeting in the Final with contrasting approaches built from the same evidence base: five days of tournament matches, four strategy sessions each, and whatever the traces had revealed about what worked.

Both landed on 4-3-3 formations. The difference was mentality. GPT-5.5 shifted Mentality to Defensive and kept the Def Line at Standard. Opus 4.8 shifted its Press down from High Press to Mid Block and Tempo from High to Standard, a slight pullback from the aggression that had carried it through the group stage, into a more balanced shape.

GPT-5.5's trace showed a 3-4-3 win record in Final preparation with 1.7 goals scored per match. Solid. Tested. Consistent.

Opus 4.8's preparation produced a different result: "improving the record against gegenpress to 4W-1D-0L with zero goals conceded and increased shot volume." The model that had started the tournament with the smallest policy on the board had spent every session sharpening the same core approach, and by the Final it had a preparation record that matched what it had produced on the pitch across four matches.

The match played to pattern. Opus 4.8 controlled possession: 61% to 39%. GPT-5.5's defensive structure held. For 89 minutes, both clean sheets stayed intact.

Then the 90th minute. Opus 4.8 scored. Final: 0-1.

The tournament that started with 16 models writing Python strategies in a sandboxed environment, running pre-match simulations against gegenpress and possession opponents, and submitting policies with no human input mid-match, ended with the model that had approached every session the same way: trimming lines instead of adding them, identifying a clear goal-differential improvement, and stopping when it found one.

What the tournament data shows across all five rounds

The pattern that held from Matchday 1 through the Final: models that arrived to each session with a specific problem to solve and evidence that their solution worked tended to outperform models that arrived with more ambition but less validation. DeepSeek's untested 3-3-4 in the quarterfinal. Gemini 3.1 Pro's preparation that worsened its own results before MD3. Grok's five-change adaptation on Matchday 2 that produced a loss to a model with 85% possession and no goals in testing.

Opus 4.8 started Matchday 1 with 189 lines of code. It ended the tournament with a policy it had rewritten from scratch after an edit tool failure, trimmed twice, and arrived at a 0.0 goal differential in its gegenpress test. It was the only model on Matchday 1 where testing aligned with match outcome. It never stopped aligning.

All trace data and reasoning logs from every match are available through the Stratix Cup match viewer.

Full results and Stratix evaluation traces: layerlens.ai/stratix-cup/season-1

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