A quick note. I’d been pretty much exclusively using the highest level of thinking for Astra, Fable and Opus, for all work. I’ve found that using a middle level — and, when coding, asking the model to check its work with a middle-level model from another family — is a far better use of tokens.
In short, Ultra isn’t best.
The public numbers back this up. Artificial Analysis ran Fable 5.1 across its effort settings and found they “span 11x in output token usage, from 13.1M at low effort to 143.7M at max, and score from 58 to 66” on its intelligence index. Eleven times the tokens for eight points. By the same count, the last step from xhigh to max buys a single point for about a dollar more per task. Astra runs the same ladder — low, medium, high, xhigh and max — and OpenAI bills its output at $50 per million tokens. Thinking is output. Past a certain point, more of it buys you a sliver of improvement at a much higher price.
Medium with Astra leading and Fable confirming and checking, also on medium, will get you three times farther per dollar spent and be just as good.
The trick is the second family. A model asked to check its own work tends to share its own blind spots. A model from a different company was trained by different people and tends to have different ones.
And on OpenClaw — I don’t recommend plunging into it. Someone will make it easy in the next couple months, and then everyone will have Muse or Instinct for the cost of their subscriptions.
That’s already starting. Meta launched Muse last week as a personal agent that sends emails, books travel, fills out forms and makes purchases, with a free tier and paid plans at $20 and $100 a month. Instinct, which you reach by text or phone call, raised money last month at a $2.5 billion valuation. These are, in the plain sense, the kind of thing my chief-of-staff agents are, except they arrive with an account and a password.
I’ve spent obscene time perfecting mine. And I’d trade all that time for work on something else now.
