#most-interesting-this-week 1 min read
The most interesting stuff this week
Small decision models replacing big ones, skills that make coding agents audit and review, and short contexts that cost no accuracy.
Decision models
- Amazon saves millions by replacing Sonnet 4.6 with Strands Decider. A 4B decision model cut ticket labeling costs 98%.
- A Jev that can ask Opus for help gives Opus accuracy at 4x lower cost. Escalation only works if the cheap judge is calibrated.
Coding agent skills
- Get your coding agent off ~2019 JavaScript, onto modern web primitives. Browser vendors wrote 200 guides on native features for agents.
- security-audit-skill. The agent that checks a finding never found it.
- Elephant-Goldfish. Subagents with no context review specs before coding starts.
- Stroq. It shows which of your keys agent sessions sent out.
Agent runs
- You don't need long context. Frequent compaction costs frontier models almost no accuracy.
- Batch, Then Aggregate: Better Agentic Coding with Parallel Attempts. Merging four attempts solved 61% of tasks, one solved 47%.
Models
- Claude Fable 5.2 and Mythos 5.2. At high effort it matches Opus 5.5 for half the cost.
- Introducing Gemini 4 Hydrogen. The cheapest Gemini 4 model scores above Gemini 4 Argon.
- Muse Ember: Our smallest, most efficient model. It scores near Muse Spark for $0.14 a task, not $2.39.
- Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values. One small fine-tune removed nearly all of Qwen's propaganda answers.