Found first: a primary source the press has not covered yet.
The Accio Team has released Occamy-1.0, a 35-billion-parameter open-weight model that scores 82.20 on the Claw-Eval co-work benchmark, ahead of GPT-5.6 Sol (81.80) and DeepSeek V4 Pro (81.70) on that measure. Model weights and a subset of training data are publicly available. The paper is at arXiv:2609.11977, submitted September 4, 2026.
What the source says
Occamy-1.0 is a 35B-A3B mixture-of-experts model built on the Qwen3.6-35B-A3B checkpoint through staged post-training over replayable long-horizon execution trajectories. On the Claw-Eval average, it scores 82.20; GPT-5.6 Sol scores 81.80, DeepSeek V4 Pro 81.70, GLM-5.2 81.60, and Qwen3.8-Max leads the table at 83.90. Scores are mixed outside that benchmark: on WildClawBench, Occamy-1.0 scores 49.16 against GPT-5.6 Sol's 67.20, and on Terminal-Bench 2.1 it scores 59.0 against GPT-5.6 Sol's 88.80. On BFCL v4 tool calling, Occamy-1.0 (65.40) edges GPT-5.6 Sol (64.33). The work is credited to the Accio Team across 43 co-authors, with no individual institutional affiliations listed.
Why it matters
On the Claw-Eval co-work benchmark, a 35B open-weight model now scores above GPT-5.6 Sol and DeepSeek V4 Pro. For teams running agentic workloads at volume, a locally deployable model that matches frontier performance on co-work tasks removes dependency on external APIs and the associated per-call cost. The release of a training data subset alongside weights gives practitioners and researchers something to work with beyond scores alone. The paper also reports efficiency figures comparing Occamy-1.0 to the Qwen3.6-35B-A3B base: roughly 150K tokens per trajectory against roughly 186K, with trace wall time around 40 seconds against around 75 seconds.