I’m thinking in all directions now.
Start with the question: how many tokens per second could you get from the fast model with commercial inference? For an agent doing many small jobs, speed matters alongside token cost.
For a sense of the ceiling, the dedicated inference shops have been posting startling numbers for a while. Cerebras said in late 2024 that its service ran Llama 3.1 70B at 2,100 tokens per second, which it called 16x faster than the fastest GPU solution, with the figures checked by Artificial Analysis. That was an older, open model on purpose-built hardware. If a current fast model gets anywhere near that on a commercial endpoint, the arithmetic of agent workflows changes.
And if it’s fast enough — well, maybe it replaces Claude, with the proper workflows. Not for the heavy lifting, but for the long tail of small, repetitive jobs that make up most of an agent’s day, where the marginal intelligence of a frontier model is wasted on the task.
That starts to make OpenClaw more practical. An always-on assistant that’s quick and cheap per token is a different proposition from one that drains a frontier-model allowance while it waits for something to happen.
It might also mean I don’t need to rotate through three Max subscriptions anymore. Anthropic’s Max plan comes in two sizes — $100 a month for 5x Claude Pro’s rate limits, or $200 for 20x — so three of them is real money every month. Maybe just one, for the heavy lifting. The rest goes to whatever is fast and cheap.
Then the wild one. Launch 60 websites and have them all autoresearch-optimizing 24/7, every day. Karpathy’s autoresearch is a loop in which an agent modifies code, trains for five minutes, checks whether the result improved, keeps or discards, and repeats. It was built for LLM training, but the loop doesn’t much care what it’s optimizing. Point it at a site’s conversion rate instead of a validation loss and the shape holds. Sixty of those, running all night on a model fast enough that each iteration costs cents, is a business that improves itself while you sleep.
