Ornith-1.5: Open Model Rivals Claude Opus 4.8 via Self-Improvement Loop

Ornith.ai has released Ornith-1.5, a family of foundation models built around a training loop in which the model generates its own work. Extending the self-scaffolding approach of Ornith-1.0, the new version closes the loop: the model proposes progressively harder tasks, builds task-specific scaffolds (tools, decomposition, orchestration), and produces solution rollouts that feed reinforcement learning. All three stages — task generation, scaffold construction, and solutions — are optimized jointly with GRPO, so reward from a rollout propagates back to reward better curricula and harnesses, not just better answers. The effect is a curriculum that automatically tracks the model’s capability frontier, targeting roughly a 20% success rate so tasks stay challenging but still yield usable training signal, with additional reward terms enforcing task validity, verifiability, and novelty.

The lineup spans a 397B mixture-of-experts flagship, a 35B MoE, and a 9B dense model, with the vendor claiming state-of-the-art results among open models at each size. Ornith-1.5-397B posts 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE — effectively matching Claude Opus 4.8 (85.0 and 59.0) while beating open competitors like GLM-5.2 and DeepSeek-V4-Flash. The 35B activates only 3B parameters per token yet outruns similarly sized dense models on agentic coding, and the 9B ships a quantized mobile build that runs on iPhone and Android while reportedly outperforming larger models such as Gemma 4-31B and Qwen 3.6-35B.

The notable angle isn’t just benchmark parity with a leading closed model, but the mechanism: reducing reliance on human-curated tasks and hand-built evaluation harnesses in favor of a system that manufactures its own training data. The design also explicitly rewards harnesses that resist reward hacking and evaluator shortcuts — a tacit acknowledgment that self-generated curricula can degenerate if the grader can be gamed. All figures are vendor-reported (averaged over five runs), so independent verification of both the benchmarks and the robustness of the self-improvement loop remains the open question.