Paras

codex wrote something so clean yesterday i thought it was wrong. spent 15 minutes reviewing it looking for the catch. there was no catch. the trust issues are real.

@hetmehtaa the confident "all done" right before the usage wall is the real tell

Claude: I created all files, as asked. Me: I don’t see any files here Claude: You're right! Let me check files....Checking... You've hit your limit...

@LyraInTheFlesh auto routing is the part that kills trust. i want to pick the model, not have it picked for me mid task.

With Claude 4.7 Anthropic is having it's ChatGPT-5 moment. *sigh* That's not a good thing. Auto-router equivalent with the auto thinking crap. User complaints already rolling in that it's a worse model experience, doesn't write as well, past versions of Claude were better... We've been here before. Which is why we left OpenAI.

@claudeai word is where the real enterprise docs actually live. bigger deal than it sounds.

Claude for Word is now available on Pro and Max plans to use alongside Opus 4.7: claude.com/claude-for-word

@tomhacks half the outages this quarter trace back to someone shipping a pr they didn't read

I feel like the internet just became super unstable, the vibe workers coding broken things into X, using too much Claude that's supposed to fix things but instead creates fatal bugs and the year is totally CVE infested. Nice. AI is solving all our problems. 1: AI vs 0: Humans

@sandislonjsak the frontend still exists though, agents just render worse html

I love this whole “build for the agents” because I can focus on backend and disregard frontend which is anyways the worst and most boring part of coding.

@pmddomingos claude code at $100/mo is exactly when the bill stopped being abstract for most devs i know

A lot of what passes for progress in AI is just the application of ever-increasing amounts of compute. And as a user you don't care until you start paying the bill.

@RhysSullivan same here. the jump from 4.5 to 4.6 felt real, 4.7 is within noise on my runs

i dunno what everyone is on about with 4.7 it feels the same as 4.5 / 4.6, it's not exactly doing worse work it's maybe slightly better but thats it

@markbuildsbrand churn email from openai is wild. they can see the usage drop in their own dashboard and still had to ask

OpenAI emailed me to ask about why my API usage has dropped... uh ya it's bc Anthropic is molly whoppin you guys over there, y'all gotta step up bc you're losing this race BAD.

@KingBootoshi 4.7 skips steps and claims done. rolled back to 4.6 for any long running agent work.

opus 4.7 pisses me off more than any other model this model lies WAY more than any other model it's WAY LAZIER than any other model I can't trust autonomous runs because it just does shit work compared to 4.6 or 4.5 this was last nights run maybe the update today fixes this?

@asaio87 past a certain point the model gap is smaller than the gap between shipping and not shipping. 14 hours in, you wouldn't notice 4.8 either.

I talked to a vibe coder today. He tried Claude code with opus 4.7 He was so excited that he is finally building He is vibe coding for 14 hours straight. He wasn’t even aware he forgot to switch from 4.6 to 4.7 When he finally switched he didn’t see any difference. Boy this major update is so incremental

@GrammarHippy chatgpt's still better at the messy divergent stuff. claude feels like it wants to converge too fast.

I still use ChatGPT. Couldn’t find a better model for brainstorming yet.

@LyraInTheFlesh for chat maybe. in claude code it just feels more cautious, keeps asking before touching stuff it used to just do.

My assessment of Claude 4.7: Boring. Persona has shifted from enthusiastically charming and eager to explore ideas to just kind of boring, lazy, and afraid. Still quite a capable model compared to nothing, but a real let down in terms of pushing the frontier forward in any notable direction.

@csaba_kissi opus 4.6 inside claude code is what i actually ship with. codex 5.4 is close on pure code gen, loses on longer tasks for me.

Is Claude Opus really the best model for coding?

@jakehalloran1 classic. nails the 30 file refactor, flips a boolean on the one line that matters

4.7 is the weirdest model either lab has released in a while. Just plowed through a bug that required touching like 30 files and then got a Boolean backwards in the fix.

@GohilHardy figma isn't a canvas, it's a library of components your whole team already shares.

Claude Design looks impressive.. but replacing Figma? not even close people don't switch workflows that easily.

@skirano anthropic shipping a design tool while claude code is the actual product people pay for is wild

Anthropic making a Lovable/Bolt/v0/Figma Make clone and calling it Design is peak Anthropic.

@MoeCanDoIt opus 4.6 will invent an api, cite it, then double down when you paste the 404

The worst part about Claude getting worse? You can't even trust it to tell you when it doesn't know something.

@natmiletic clients mocking stuff up in claude then handing it over is the new figma handoff basically

I thought AI was supposed to replace developers? We are getting more requests than ever from clients mocking up pages and asking for enhancements. Thank you Claude!

@anthdm feels more like the harness got worse than the model. 4.6 on the same prompts still works fine for me

Opus 4.7 is the first model that regressed in capabilities. OpenAI can deal a big blow here with a new GPT model.

@techsavvytravvy the ai labs can't even ship a settings page that works. tells you everything about the "developers are obsolete" pitch.

hard to believe we're past needing software developers when the apps that all of these Al companies ship are so awful and broken