Pattern: The Frontier Bifurcates — Open-Weight Models Close the Gap While Gated Models Pull Away

Simon Willison·medium signal

In the same week: DeepSeek V4 ships Apache 2.0 open weights at 80.6% SWE-Bench (within 0.2 points of Opus 4.6), while Anthropic's Mythos Preview at 93.9% SWE-Bench is explicitly withheld from general availability due to security capabilities. The frontier is splitting into 'open and competitive enough' (DeepSeek, Ling) vs 'gated and significantly superior' (Mythos). Builders get near-frontier performance for free, but the true ceiling requires API access through approved channels.