Monday morning, I nearly drove a client crazy

At 10 a.m. on Monday, I was sitting in a Chengdu café, rewriting the introduction page for a lawyer’s client. This was the fourth revision, and she still wanted a different tone. I rewrote the copy while worrying about model costs, ready to buy more credits—I’d made that mistake before. I kept thinking, “Spending a bit more will be safer,” only to discover at month-end that every task had quietly added to the bill.

Astra and Sol arrived, but stronger isn’t always the better choice

OpenAI has launched Astra and Sol. Astra leans toward maximum capability; Sol leans toward better value for everyday work. In Artificial Analysis tests, Sol cost about 15% per task as much as comparable Astra tasks, roughly 85% less; however, Astra scored higher on complex reasoning benchmarks. Public feedback wasn’t unanimous either. Some people noticed odd traces of training in the output. The official team said some criticism had been overstated, but acknowledged that problems existed. What I care about isn’t the top of a leaderboard, but whether I can deliver my work reliably.

I know Lin Yue, an independent consultant in Hangzhou who creates paid knowledge products. She gives Sol daily topic-rewriting work and uses Astra to organize data for quarterly reports, while keeping an older model as a control. Her method is simple: run the same real-world tasks through each model, then compare the results and bills. I’ve started doing the same on my own client projects, and found that “cheaper” doesn’t mean “use without limits.” When complex tasks need frequent rework, the API calls I saved can reappear as time spent fixing the work.

Try this today: calculate the cost, then rerun your own tasks

Replication cost: For API pricing, Astra costs $8 per million input tokens and $40 per million output tokens. Sol costs $2 per million input tokens and $10 per million output tokens. Cached inputs cost $0.40 for Astra and $0.10 for Sol. I don’t have to pay the full amount immediately: I can open the model picker from an existing ChatGPT Plus, Pro, Team, Enterprise, or Edu account and test first. For developer API access, I need to register for the platform, add a payment method, and select Sol or Astra.

The technical barrier is modest: I need to turn my commonly used prompts and evaluation criteria into fixed tasks; coding isn’t necessary yet. Time cost: an ordinary chat user can complete a same-prompt comparison in about 10 minutes; API setup takes about an hour. The first step isn’t to top up credits. It’s to open the model picker, take one piece of client copy I sent yesterday, run it through both Astra and Sol, and record the pass rate, number of revisions, and cost. This tool isn’t for everyone, and personal results may vary. There’s no urgency to try it today.

Choose for the stage I’m at, not for the leaderboard

If I’m just starting out: I’d use Sol for rewriting, summaries, and first drafts of emails, recording each result alongside the time it took to reach “approved.” Once I have stable clients or a team account, I’d consider testing them separately.

If I already have one or two clients: I’d run a small comparison with similar tasks: half through Sol, half through Astra, using the same acceptance criteria as before. When the feedback differs, I’d use whether the client needed one fewer revision—not how polished the demo looked—as the deciding factor.

If I’m scaling: I’d track three numbers: success rate, manual rework time, and API cost, then decide how to route tasks. I wouldn’t migrate the whole team just because a new model launched, and I wouldn’t ignore Astra’s advantage on complex work. I might invite the team to test both for a week, but I’d keep one person who can clearly explain why tasks are assigned the way they are.