Existing benchmarks for autonomous AI scientists only evaluate final outputs—code, hypotheses, or papers—while discarding the reasoning process that produced them. This prevents auditing of methodology, diagnosing failure modes, or distinguishing systematic reasoning from lucky guesses. To address this gap we introduce OpenDiscoveryTrace, a public dataset containing 558 complete AI scientific agent trajectories. Each step records nine structured fields: thoughts, tool calls, observations, errors, revision triggers, and self‑reported confidence. The dataset spans 124 scientific tasks across drug discovery, materials science, genomics, and literature analysis. It covers seven models: three frontier models (GPT‑5.4, Claude Opus 4.6, Gemini 3.1 Pro) with 124 trajectories each, balanced across domains and difficulty; four open‑weight models (Qwen2.5‑7B, Mistral‑7B‑v0.3, Phi‑3.5‑mini, Qwen2.5‑1.5B) with 30 each; plus 60 live‑retrieval variant trajectories. Pilot analysis of 363 LLM‑judged trajectories shows that process traces reveal behavioral differences invisible to output‑only evaluation: the three frontier models achieve comparable success rates (84‑89%), yet Claude Opus 4.6 generates 30× more errors than GPT‑5.4 (2.5 vs 0.08 errors per trajectory, $p < 0.01$).
Review: Tracing the reasoning process offers a finer‑grained assessment of AI scientists, enabling error pattern detection and improving model interpretability.