EBR-bench updates include card ban and multi-agent testing

Astra: GPT-6 Astra is an advanced AI model developed for high-performance reasoning and task completion. It demonstrated exceptional results on EBR-bench by identifying and exploiting powerful card combinations in the Earthborne Rangers game. The news centers on Astra's perfect initial scores, its continued strong performance after a key card ban, and its role in prompting benchmark adjustments for better measurement of tactical learning. EBR-bench: EBR-bench is a benchmark designed to evaluate AI models' ability to learn from experience through repeated play of the complex board game Earthborne Rangers, which features long-horizon tasks requiring strategic deck-building and tactical adaptation. The benchmark measures how well models improve scores over multiple playthroughs by exploring card combinations and minimizing fatigue mechanics. In this news, EBR-bench is being updated with a card ban and multi-agent testing options to maintain its challenge as top models approach saturation. GPT-5.6 Sol: GPT-5.6 Sol is an AI model in the GPT lineage optimized for strong performance in analytical and exploratory tasks. It will have its EBR-bench results updated under the revised benchmark rules after the powerful card was banned from default testing. The news positions it as one of the recent models tracked to monitor when saturation effects emerge in long-horizon evaluations. Claude Opus 5: Claude Opus 5 is a high-capability AI model in the Claude series focused on complex problem-solving and multi-step reasoning. Its EBR-bench performance is being re-evaluated with the updated experimental settings, including the card ban. The development notes its inclusion in future reporting to reflect changes in how models are assessed on strategic gameplay. Claude Fable 5.1: Claude Fable 5.1 is an AI model from the Claude family designed for advanced language and reasoning tasks. It is included among the models whose EBR-bench scores will be updated and reported under the new default settings following the card ban. The news highlights its evaluation alongside other recent models to track progress in long-horizon benchmarks. AI Benchmarks: Benchmarks like EBR-bench are evolving to emphasize genuine strategic learning over single exploitable strategies in complex, multi-step environments. Evaluation Methods: Multi-agent setups are being tested as a way to encourage broader strategy exploration in AI systems without consistently raising overall performance metrics. Model Capabilities: Leading AI models continue to show rapid progress in long-horizon tasks involving experience-based adaptation and deck exploration in intricate game settings.