30 days: 67B tokens. Hundreds of thousands of lines shipped. Subagents covering hundreds of hours a week.
I don't ration the model. I treat agents as headcount with a job, keep judgment in the loop, and spend where the bottleneck is the model.
GPT-6 Astra made that cheaper.
@tetsuoai@lingxi Reversible is the right constraint. I’d keep the overnight agent on reversible cleanup and have the morning check compare the repo itself, not just the agent’s report. I’ve had an automation say a rollout was complete while the old process was still running.
@deanwperkins I check the actual path, not the feature list: can the right account sign in, can a real message leave, and does the system of record show the work completed? I’ve seen a polished dashboard hide an unfinished launch.
@brycent The “by default” part is the whole point. I’d rather narrow the agent’s tool surface before the first run than explain later why it had access it never needed. I do the same with client auth: one credential per client, never shared.
@garrytan@capydotai The speed-up is useful. Before I add another agent path, I map what is already running and check whether the gap is real. That keeps a faster workflow from becoming one more system to remember.
@brycent repost had me looking at this — very excited, and agree this is likely a unicorn.
Only tidbit is I guess I missed the preorder window, and the discount right now is immaterial, so instead of buying it immediately my reaction was to pause and wait… I pushed through it.
But realistically a sizable preorder locks customer in and removes friction. Something to think about @lumeriaskin
@elonmusk I hit 67b tokens on Claude and Codex so swapped a lot of my workflow to @grok SuperHeavy… and maxed out with 2 days to go in the weekly reset. This broke a bunch of long running routines I’ve since balanced out.
Apart from that LOVING 4.6.
Bro, please bring back daily resets for SuperHeavy power users… keep weekly for the other subs.
@mattyp@Cloudflare@bot@mattyp Set up about 10 of these scheduled jobs this weekend, all read only, each ends in a brief. What made me stop trusting the report: an agent in my stack marked a batch of document pulls done and had written 200-byte placeholder files. I open the real output first now.