Yesterday, a client wanted a draft, and I opened the same chat window again

At 9 p.m. on Tuesday, Shanghai-based freelance designer Lin Ran received a client message: “Can you send me a draft first?” She stared at the screen and opened her usual tools again. The problem wasn’t lack of skill; with so many models, choosing one became harder. I’ve been there too. I used to rush to replace every tool whenever something new launched, only to find that the old workflow felt more natural the next day. The novelty lasted two days.

This launch matters less for the names than for the way work gets done

OpenAI launched GPT‑6.2 Astra and Sol. Astra is built for work that needs multiple steps and tool coordination, while Sol leans more toward execution and programming. For us, the useful question isn’t the model label. It’s whether the model can break client materials into tasks, call the right tools, and deliver the result. In plain terms, Astra feels more like a colleague who researches, plans, and checks progress. Sol feels more like an assistant handling clearly defined technical instructions.

Who’s using it? Zhou Hang, a one-person company founder in Hangzhou, has the model turn client interview notes into a requirements list, generate a first draft, and then checks the brand voice himself. He doesn’t send the whole email untouched; he delegates the most time-consuming整理 work while keeping final review. That fits side-project operators: the goal isn’t “full automation,” but freeing people from moving information around by hand.

To replicate this today, I wouldn’t move the entire workflow

Money: If an existing subscription already includes a usable version, the added cost may be 0 yuan. If a separate plan is required, I’d check the official pricing before upgrading for the novelty alone. Time: Thirty minutes is enough for one comparison; sixty minutes can cover one real task end to end. Technical barrier: No model-parameter knowledge is needed. I only need to copy a piece of client material, state the goal clearly, and judge whether the answer is reliable. First step: Open the model selector in the OpenAI product I use and check whether Astra or Sol is available on my current account. If not, I’ll note it and try when access opens.

How do I test it without fooling myself? I use a real task, not vague questions. I’ll copy a redacted client brief, have both models attempt it, then compare three things: whether they missed any requirements, whether the output is directly actionable, and how long my own review takes. For contracts, quotes, or private client data, I remove names, contact details, and sensitive information before pasting anything in.

I’d decide whether to switch based on my current workload

If I’m just getting started, I wouldn’t switch yet. Not testing now is fine; mastering the tools I already have is often more cost-effective than chasing every release. I’d wait until I’m doing the same repetitive整理 work at least three times a week, then run a small test.

If I already have one or two stable clients, I’d test one workflow: interview-note organization, a requirements list, or an email draft. I’d keep the old process as a fallback and stay responsible for the final client delivery.

If I’m scaling, I’d first track which task type consumes the most delivery time, then decide whether to upgrade the plan or connect an automated workflow. I wouldn’t change three places at once. I’d let one part run for a week, record the hours saved and number of rework sessions, and only then consider expanding.

This tool isn’t for everyone, and none of us needs to chase every launch. What matters isn’t the model name. It’s whether the tool can save clients one step of waiting and leave us with a little more time from the revenue we earn.