@robinebers same. burned an ungodly amount of tokens. so productive. really interesting how this plays out. Was this just spite on the part of US Dept of War from convos a few months ago? Do OpenAI release a monster model now? Would they get the same treatment? Does Opus get secretly better?
@andyhennie@robinebers Hmm it does for me too. Didn’t configure anything - just saw threads in app that I had initially started in CLI. A double take moment as it was unexpected!
Ironically, this is the first hidden backdoor I found while intentionally hunting for bad actors in the WordPress plugin repo. It might also be my last big find. With AI, bad actors won't be able to hide for long.
https://t.co/A2X8WP9vUv
@robinebers yeah was getting obscene Opus stupidity last week... seems to have sorted itself out this week though. Codex yesterday had an hour of nonsense too...
@marcelpociot@tomvance94@AnthropicAI have been unclear on this recently, but the usage policy does explicitly state that the Agent SDK cannot be used with a Max plan. I think people are getting away with it for now, but I can't risk my Max plan! Really weighing up an OpenAI pro sub...!
@alexoakdev Yes... very noticeable on $100 plan. I hit my limit once in 4 months, but this week hit 5-hourly limit twice in a row + weekly + used $40 in extra usage. Even using Sonnet 4.6 rather than Opus to try and reduce usage. But I think Sonnet is token-hungry. Have had to upgrade
@clifgriffin my vote is yes - ditch the bird. also the movement draws eyes there (though that's a micro optimisation). can you offer a completely free tier (y'know, cos wordpress)? free trial doesn't quite feel compatible with a review collection product as data could get "stuck".
@TheCraigHewitt@Shpigford Gemini Flash 3.0 was a no go for me. It would say “let me check and come back to you” and then never did. Setting its own config just didn’t work but it asserted it had been done. And got 400 errors with tool calls.
@benhylak yes had the exact same journey/experience. I was very underwhelmed at first given hype, but upped my prompt game and switched on gpt-5-high and it one-shotted an extremely complex solution that gpt-5 and opus 4.1 was making slow, painful progress on over 2 days. really amazing
@ericzakariasson@cursor_ai REEAALLY good output having used for 2 hours. Fast and not missing a beat with complex stuff thrown at it. A slight niggle is when I first started using it, changes were applied to my files by the agent, but later I then have to click "apply to xxx.js".