Good bicycle, good pelican: https://tools.simonwillison.net/markdown-svg-renderer#url=ht...
With GLM 5.1 and 5.2, the big problem was tool calling and long-horizon coherency. 5.3 was more trustworthy at the cost of longer thinking traces, and now Flash seems to improve on it once again with a more concise, smaller model. As long as there aren't any noticeable regressions, I could see myself defaulting to this for >90% of my day-to-day coding work.