GPT-6 Astra Tops New AI Science Research Benchmark

Overview

The wait is over. The Terminal-Bench-Science 0.1 leaderboard has officially launched, offering the first rigorous agentic benchmark for scientific research workflows. This new evaluation framework is designed to test how well large language models can autonomously navigate the complexities of real-world laboratory and computational science.

Two commercial giants currently dominate the top of the charts. GPT-6 Astra (max) leads the pack with a score of 63%, closely followed by Claude Opus 5.5 (xhigh) at 62%. While these results signal significant progress in agentic AI, they also highlight that there is substantial room for improvement before AI can fully replicate expert scientific reasoning.

Benchmark Construction and Methodology

Terminal-Bench-Science 0.1 was introduced in August 2026 as a collaborative effort led by Steven Dillmann and researchers at Stanford. The project was developed in partnership with the @terminalbench team and a global community of open-source scientific contributors. The goal was to move beyond simple knowledge recall and evaluate models in dynamic, interactive environments.

The benchmark consists of 70 expert-curated tasks spanning five major scientific domains. These domains include life sciences with 19 tasks, physical sciences with 17 tasks, mathematical sciences with 17 tasks, engineering sciences with 9 tasks, and earth sciences with 8 tasks. Each task is grounded in real research scenarios, ensuring that the evaluation reflects practical challenges rather than theoretical puzzles.

The testing methodology mirrors the original Terminal-Bench approach but is specialized for scientific inquiry. Each task places an AI agent into a sandbox environment equipped with specific data, tools, and instructions. The agent must then run the entire workflow from start to finish without human intervention. Automated tests grade every task on a strict pass/fail basis, and the final scores report the average pass@1 metric over three separate attempts.

This benchmark is not about answering questions; it is about executing complex, multi-step research protocols with precision and autonomy.

Key Performance Takeaways

The initial results reveal a clear hierarchy in the current landscape of scientific AI. Only two models have managed to score above the 50% threshold, underscoring the difficulty of the tasks. GPT-6 Astra (max) holds the top spot at 63%, while Claude Opus 5.5 (xhigh) follows closely at 62%. The standard max configuration of Claude Opus 5.5 scored slightly lower at 59%.

Recent releases have demonstrated substantial advancements in agentic capabilities, yet even the leading model has significant headroom. A score of 63% indicates that while top-tier models can handle a majority of tasks, nearly four in ten scientific challenges remain unsolved or incorrectly executed by the best available systems.

Domain-Specific Performance

Performance varies significantly across different scientific fields. Life sciences emerged as the lowest-scoring domain for most leading models. This finding is notable given the intricate and often unstructured nature of biological data. However, domain scores should be interpreted with caution due to noise; with only 8 to 19 tasks per domain, small sample sizes can skew overall averages.

Claude Opus 5.5 (xhigh) exhibited a stark contrast between disciplines. It passed 71% of mathematical sciences tasks, demonstrating strong logical and computational reasoning. In contrast, its performance dropped to just 46% on life sciences tasks. This disparity suggests that current models may struggle more with the ambiguity and variability inherent in biological research compared to the structured logic of mathematics.

The Open Weights Gap

The gap between proprietary commercial models and the best open-weight alternatives remains vast. The leading open weights models, GLM-5.3 (max) and DeepSeek V4.1 Flash (max), scored 10% and 9% respectively. These figures sit more than 50 points below the leaders, highlighting a significant chasm in capability for autonomous scientific agents.

While open models have made strides in general coding and reasoning benchmarks, the agentic scientific domain appears to be a particularly tough nut to crack. The combination of tool use, domain-specific knowledge retrieval, and long-horizon planning seems to require scale and refinement that current open-weight releases have not yet achieved.

Implications for AI Research

The launch of Terminal-Bench-Science 0.1 provides a critical benchmark for the future of AI in research. It shifts the conversation from passive language understanding to active scientific execution. For developers and researchers, these results serve as a realistic baseline for what autonomous agents can currently achieve.

The data suggests that while we are making rapid progress, AI is not yet a replacement for expert human scientists in complex, multi-domain workflows. The low scores in life sciences and the wide gap with open models indicate that significant architectural and training innovations are still needed.

As the community continues to contribute new tasks and refine existing ones, this benchmark will likely become a standard for measuring agentic maturity in the sciences. The current leaders have set a high bar, but the journey toward fully autonomous scientific discovery is far from over.

Looking Ahead

With 70 tasks already in place, the benchmark offers a solid foundation for ongoing evaluation. The scientific community is encouraged to submit new tasks and propose modifications to existing ones as part of the open-source collaboration.

For now, the leaderboard stands as a testament to the capabilities of GPT-6 Astra and Claude Opus 5.5, but also as a reminder of the hurdles that remain. The road to reliable, general-purpose AI scientists is long, and Terminal-Bench-Science 0.1 is the first clear mile marker on that path.