Insider's opinion on Astra capabilities

@Lentils80 post on X "Over the past few days, two GPT Astra checkpoints, "ultima-alpha" and "vega-alpha", were undergoing testing "ultima-alpha" appears to be the release candidate intended for the public, while "vega-alpha" is the cybersecurity-focused variant meant for security work in select enterprises Based on extensive testing on my part, when OpenAI said Astra is built for long-running tasks and orchestration they really meant it. It can run for an incredibly long time even without setting "/goal", fully autonomous, and it's very capable at orchestration and guiding the subagents it spawns For the research community, it's very good at applying existing academic literature. Tried it at some hard graphics optimization stuff, so a LOT of complex math involved, and it did great It also writes code with great quality and maintainability (for an LLM ofc), ranking the best out of all models in that I'd say, but most normal people will probably just run it as the main agent and cheaper models as subagents Additionally, creative writing appears to be way better than 5.6 Sol imo, still not the best but noticeably less slop" - Better than Fable on Code, but worst on Frontend and 3D (Not sure if he was talking about 5 or 5.1)