
ALT One-page summary of Claude Fable 5.2: $8 / $40 per million tokens; Artificial Analysis Intelligence Index 65 vs Opus 5.5's 59 at $26.54 vs $10.04 per task; biggest leaps RoboDojo 64.9 vs 39.4, Remote Labor Index 41.3 vs 21.3, SWE-bench Pro 93.0 vs 89.9; MNIST weight test 81.4% vs 71.0%; four prompting tips

ALT A one-page overview of the Context Language Models paper. Top left, a cartoon robot rubs out lines in its own context.txt notebook. Title: Context Language Models, language models that manage their own context. Top row: The idea, where a standard LM appends its output to the context but a context language model writes its whole next context. Context As A File, where the context is mirrored to a file the model edits between turns. Suffix Cache Reuse, matching SGLang's accuracy with 65% of the FLOPs. Bottom row: Its own strategies, a subagent scoreboard, a new notes chat role and a compact_turns() helper reused 37 times. ContextBench, four synthetic tasks where CLM stays near 100%. Learn it, where one sentence steers it or RL with Stepwise GRPO takes Qwen3.5-9B from 28.8% to 42.5% on BrowseComp-Plus. Results (zero-shot): 59.4% on BrowseComp-Plus, 70% of the FLOPs on TerminalBench, and a 65% speedup in a six-agent swarm.




