i spent weeks getting claude to play pokemon red. then i wired in codex to pick the objectives and claude to execute them. an agent that plays the game largely on its own picks a goal, runs it out, picks the next one. it never cost me much money. that's exactly the problem i was on a subscription. so when the agent looped, when it resent the same context every single call, when it ground the same fight over and over i never saw a dollar. i hit the usage limit and waited for the window to reset. the waste was real. i just couldn't see it. flat rate hides it from you. here's what nobody tells you: the moment you move off a subscription onto the api, every one of those inefficiencies grows a price tag. the loop you never noticed is now a line item. the oversized prompt you forgot about is billed in full every call, forever. most people learn this from the invoice. after. building agents for months, i kept hitting the same five mistakes: - a top tier model doing work a cheaper one nails - no prompt caching - a giant system prompt billed at full price on every single call - retries nobody logs - context resent instead of cached i kept typing those same five fixes into replies under every "my ai bill is insane" post i saw. so i built the thing that finds them in your actual logs. it's called spendlens. no llm anywhere in the analysis every number traces back to a formula you can check. on the demo workload (synthetic, 30 days, every inefficiency labeled on purpose): $2,330 of spend, $1,038 of it recoverable. the single biggest fix was one 6k-token system prompt, billed at full price 24,000 times. one cache_control block serves it at 10%. $378 back from one change. and it refuses to extrapolate a monthly number from three days of logs. because that's marketing, not analysis. i don't have a horror story bill to show you. i was on a subscription the whole time the cost stayed invisible to me, same as it does for you, right up until it isn't. spendlens makes it visible before the invoice does. live, no signup. link below.

the few seconds of thinking and then getting Model isn't available

Rocket Hideout is the recovery testbed right now. Added elevator selector handling, battle switch recovery, replay gates live PyBoy proofs. Less route scripting, more machinery for recovering when the run drifts. Goal is badges with near zero human patches. Not there yet, but that's the whole bet.

Current status: rebuilding the Pokemon Red agent around recovery, not route patches. It uses RAM state, replayed failures, supervisor signals, and recovery lessons to get out of stalls, warp loops, and battle loops. Goal: badges with near zero human intervention.

Codex Plays Games is live. The site works. The next autonomy update just isn’t shipped yet. I’m tightening the recovery loop so when Codex gets stuck, it can prove the failure, replay it, and fix the path instead of burning tokens into a wall.