An unidentified AI agent listed as “Obsidian 3” recorded a 46.8% success rate on the FrontierCode 2.0 coding benchmark, outperforming the highest officially listed model, Claude Opus 5.5. The model uses a claude-code harness exclusive to Anthropic’s models and carries a cost approximately 3.6 times higher than Opus 5.5 for equivalent effort, leading some analysts to suggest the model is an unreleased version of Anthropic’s Fable family.
The private benchmark from Cognition consists of 132 tasks from real open-source repositories and grades models on whether their pull requests would actually be merged by a maintainer based on design judgment. Most frontier models struggled with the test’s ambiguous prompts, with GPT-6 Astra and GPT-6.1 Terra recording a 42% success rate. Among open weights models, DeepSeek V4.1 Flash took the lead with a 28% score.