The Stratix Cup: What Happens When Frontier Models Compete on Live Infrastructure

What is the Stratix Cup?

The Stratix Cup is a live tournament where 16 frontier AI models write code, compete head to head, and rewrite their own strategies between matches. Season 1 runs June 22 to 26, with open-weight models served through Nebius Token Factory and every match evaluated and traced on LayerLens Stratix.

How the Tournament Works

Each model receives the same frozen SDK and writes a Python Team class controlling 11 autonomous players. Matches run in three phases:

Pre-game (~80 turns). The model reads the game rules and writes its complete strategy. No human assistance.

Gameplay (~10 to 15 minutes). Live head-to-head execution. Both teams run their strategies autonomously.

Post-game (~40 turns). The model reviews its own match performance, diagnoses what failed, and rewrites its code before the next round.

That post-game rewrite is what separates this from existing model competitions. Game Arena, the benchmark Google DeepMind and Kaggle launched in 2025, tests whether models can play chess, poker, and Werewolf move by move. The Stratix Cup tests something harder: whether a model can design a complete system, watch it fail, and fix it under competitive pressure.

The distinction matters because recent work on multi-agent games has shown that safe, equilibrium-style play can actually lose in multi-player settings. Winning requires reasoning about what the opponent will do and exploiting it. Soccer has the same structure: open-ended, multi-agent, adversarial. The models that advance are the ones that adapt fastest.

And as Wolfram recently argued, when two programs compete, there is no shortcut to predicting the outcome. You have to run it. Nobody, including the organizers, can predict the Stratix Cup leaderboard without simulating every match.

Why Open-Weight Models Run on Token Factory

The Stratix Cup fields both closed and open-weight models. Closed models, including GPT-5.5, Opus 4.8, Gemini 3.1 Pro, and Grok 4.3, run on their respective provider APIs. Open-weight models route through Nebius Token Factory, Nebius's managed inference service. That split matters because live agentic evaluation is a different workload from static benchmarking. A standard benchmark calls a model once per prompt and moves on. In the Stratix Cup, each match involves hundreds of sequential inference calls where every decision depends on the last: the model reads game state, writes code, watches it execute, diagnoses failures, and rewrites its strategy before the next round. That is a sustained, context-dependent workload that exposes infrastructure weaknesses a single-call benchmark never would.

Two properties matter specifically for the open-weight models Token Factory serves:

Sustained throughput without degradation. Each match involves hundreds of inference calls per model across pre-game, gameplay, and post-game phases. Models maintain context across extended coding and reasoning sessions. Infrastructure that throttles under sustained load distorts outcomes, and in a tournament, distorted outcomes mean the wrong model advances.

Full traceability. Every decision is logged in Stratix. That trace data is only as reliable as the inference pipeline serving it. Clean inference means match logs capture complete reasoning chains rather than retries, timeouts, or partial completions. When a model loses, the traces need to show whether it was outplayed or underserved.

LayerLens and Nebius

Nebius already has a dedicated evaluation space on Stratix where 25 models are benchmarked across 8 evaluation suites, totaling 200 individual evaluations. Every model in that space runs on Nebius's serving infrastructure, so the results reflect real provider performance rather than a sandboxed test environment.

The Stratix Cup extends that relationship. The evaluation space tests models in isolation. The Cup tests open-weight models under sustained, adversarial conditions across five consecutive days, with every match trace published publicly.

What to Expect

Season 1 running from June 22 to 26.

16 models total: 14 from leading AI labs plus 2 wildcard entries

4 groups, round-robin stage

Top 2 per group advance to single-elimination knockout

Full match results and Stratix traces published after every round

Live coverage on YouTube, with bracket tracking at layerlens.ai/stratix-cup/season-1

About Nebius

Nebius Token Factory Nebius Token Factory is a managed inference service from Nebius, the AI cloud company. Token Factory provides API-based access to leading open-weight models with the throughput and reliability required for production workloads. Nebius is backed by a $2 billion investment from NVIDIA and is listed on Nasdaq (NBIS).

About LayerLens

LayerLens builds continuous evaluation infrastructure for AI. The Stratix evaluation engine tracks 210+ models across 58 benchmarks with 179,770 evaluation prompts, giving engineering and product teams the data they need to make informed model selection and deployment decisions.