Gemini 4 Argon
Third among 13 tested configurations: 42.46 V2WScore, 19.6% task success and 91.0% build rate. Inspect the component metrics and actual successful and failed executions.
Published result snapshot · 27 October 2026
Testing.
The result is not yet verified. This page will become the model's report once its evaluation is complete.
We will report the overall ranking alongside build rate, task success, stage progress, geometry and dynamics. Strengths and limitations will be tied to those metrics and actual model outputs.
Feedback and recovery claims require the agent's iteration logs. No such capability is inferred from a final score or rollout alone.
Follow the four-model evaluation ↗The measured result.
The overall score combines functionality, geometry and dynamics, with executability as a gate. Gemini's V8 result was published on 27 October 2026. Comparison rows use the same 222-instance, 39-family score scope and scoring contract, while retaining their original publication dates. The paper's results are a separate historical set.
| Model | V2WScore ↑ | Success ↑ | Progress ↑ |
|---|---|---|---|
| #1 · Claude Opus 5.5 | 52.41 | 33.6% | 0.61 |
| #2 · Claude Fable 5.1 | 44.13 | 22.2% | 0.45 |
| #3 · Gemini 4 Argon | 42.46 | 19.6% | 0.41 |
| #4 · GPT-6.1 Sol | 41.19 | 10.9% | 0.30 |
Beyond the overall score.
| Metric | Value | Coverage |
|---|---|---|
| V2WScore | 42.46 | n=222 / 222 |
| Task Success | 19.6% | n=222 / 222 |
| Stage Progress | 0.41 | n=204 / 222 |
| Scene CD | 5.50 cm | n=202 / 222 |
| Object Shape CD | 1.04 cm | n=201 / 222 |
| Object Size Error | 3.09 cm | n=201 / 222 |
| Terminal Relative Translation Error | 32.38 cm | n=152 / 222 |
| Terminal Relative Rotation Error | 107.02 deg | n=73 / 222 |
| Fixed-ref Task | 11.4% | n=172 / 222 |
| Fixed-ref Progress | 0.24 | n=172 / 222 |
| Fixed-ref Translation APE | 18.83 cm | n=172 / 222 |
| Build Rate | 91.0% | n=213 / 222 |
Actual model executions.
Four verified Gemini rollouts show a completed pick-and-place and failed assembly and placement. The matching two-handed-lift record has no delivered submission and no rollout video. These inputs were inherited from a directed Opus-success/Fable-failure selection; they do not estimate Gemini's overall success frequency.
Gemini 4 Argon V8 · published 27 October 2026. Formal V2WScore uses all 222 instances across 39 task families. Component metrics are family-macro averages with their own finite coverage; task success is not a passed-instance count. Earlier model records retain their own publication dates. This report reads published records; it does not represent a fresh runtime reevaluation. Result snapshot and provenance ↗
