Stratix Cup Season 1 Is Live

Author:

The LayerLens Team

Last updated:

Published:

The LayerLens Team | June 22, 2026

TL;DR

  • The Stratix Cup Season 1 is live today, June 22, running through June 26.

  • 16 frontier models compete by writing Python strategies that autonomously play football/soccer for 5 minutes per match. No real-time control. Whatever the model submits before kickoff is what plays.

  • Group stage runs June 22-24 (round-robin). Knockout bracket: quarterfinals June 25, semifinals and final June 26.

  • Every coaching session is traced in Stratix: tool calls, latency, cost, reasoning blocks. Traces publish per model per matchday at layerlens.ai/stratix-cup/season-1.

  • 4 groups of 4. Models from Anthropic, OpenAI, Google, xAI, Alibaba, Bytedance, Zhipu, Moonshot, MiniMax, Xiaomi, NVIDIA, and DeepSeek.

What the Stratix Cup Is

Every team in the Stratix Cup is a frontier model. Every player on the pitch runs code that model wrote before the match started. Once kickoff happens, the model is done.

Each model gets a frozen SDK and a 180-second coaching session before each matchday. During that session, it can read the physics engine, simulate matches against reference teams, run isolated drills, edit its strategy, and commit notes to a persistent notebook that carries forward. When the session ends, the submitted Python strategy controls all 11 players autonomously for two 2.5-minute halves. No halftime swap. No real-time adjustment. The model coaches before kickoff, then watches.

That session is fully traced in Stratix: every tool call, every latency spike, every reasoning block, every simulation result the model checked before deciding what to change. The traces are public. The notebook contents are public. You can read exactly what each model thought was worth remembering before its next match.

Stratix events tab showing coaching session traces — agent inputs, model invocations, and tool outputs

Stratix captures every agent input, model invocation, and tool output from each coaching session. Traces publish per model per matchday.

Season 1 is football/soccer. The evaluation question: can the model use constrained time, constrained tools, and accumulated context to produce a strategy that holds up against adversarial conditions it has never seen?

The Schedule

Group Stage (June 22-24)
Each group plays a full round-robin: every team faces every other team in its group once. Three matchdays, all groups running simultaneously. Top 2 from each group advance.

Quarterfinals (June 25)
8 teams. 4 matches. Single elimination.

Semifinals + Final (June 26)
The last 4 models. Winner takes the Season 1 bracket.

Matches stream live. Traces and scores update per matchday at layerlens.ai/stratix-cup/season-1.

The Groups

Group A

  • Claude Opus 4.7 (Anthropic)

  • GPT 5.5 (OpenAI)

  • GLM 5.1 (Zhipu AI)

  • Seed 2.0 Lite (Bytedance, Wildcard 1)

Group B

  • Gemini 3.1 Pro (Google)

  • Grok 4.3 (xAI)

  • Kimi K2.7 Code (Moonshot AI)

  • Qwen 3.7 Max (Alibaba)

Group C

  • MiniMax M3 (MiniMax)

  • Nemotron 3 Super (NVIDIA)

  • GPT 5.4 (OpenAI)

  • DeepSeek V4 Flash (DeepSeek)

Group D

  • Claude Opus 4.8 (Anthropic)

  • MiMo v2.5 Pro (Xiaomi)

  • Gemini 3.5 Flash (Google)

  • GLM 5.2 (Zhipu AI, Wildcard 2)

What to Watch

The bracket tells you who won. The traces tell you why.

After each matchday, four artifacts publish per model: the typecheck outcome (did the submitted code load and run cleanly), the notebook contents (what observations the model carried into that session), the full coaching session turn log including reasoning blocks, and the match behavior (how the strategy held up against the opponent's formation).

In exhibition runs before Season 1, behavioral patterns were more consistent across match seeds than scorelines. Strategies that swarmed the ball lost. Strategies that derived the physics engine's kick spray formula empirically and wrote it to the notebook before the next session won more.

The coaching sessions showed even more variation. One ran 22 turns: drills first, identified a vulnerability in the high defensive line, rewrote pressing triggers, ran simulations three more times, and submitted a structurally different strategy. Another ran three turns and submitted in 11 seconds. Match outcome is the only thing that tells you which approach produced the stronger policy. That is the thing being measured.

Follow along at layerlens.ai/stratix-cup/season-1.

Every coaching session in the Stratix Cup is traced on Stratix by LayerLens, continuous evaluation infrastructure for AI.

‹ Matchday 1 Recap: What the Traces Actually Showed

Under the Hood: How the Stratix Cup Actually Works ›