GPT-6 Astra Lands, Immediately Benchmarked for Real Work

GPT-6 Astra dropped and within 24 hours the HN community had it running on OpenRouter, testing it in code review via CodeRabbit, asking whether it can design PCBs, and comparing it to ARC-AGI-3 scores (reportedly 98.6%). One thread specifically notes Astra seems slower than GPT-5.6 on equivalent tasks, which suggests it is doing more per token. The code review thread flags that both OpenAI and Anthropic have been releasing slightly better models at roughly 2x the price of the prior iteration.

The PCB design thread is worth watching closely. Builders are asking whether Astra can handle schematic routing and BOM sourcing from Digikey and LCSC. The current answer is 'useful for libraries and footprints, not yet for full layout,' but the direction is clear. Hardware design has historically been one of the domains most resistant to AI acceleration.

Spotify's Portal tool is adjacent here: it reportedly cut Claude Code token usage by 90% by routing subtasks to cheaper models like Gemini 2.5 Flash. The thread is skeptical but the idea of multi-model pipelines as a cost optimization layer is gaining traction among practitioners.