Michael Grant

@EXM7777 Exactly my experience. Opus 5 for focused subagents? Great. Opus 5 for master orchestrator? Horrible.

Anthropic failed at building a daily driver somehow Opus 5 is exceptional at precise work (worldbuilding, 3D, frontend) and barely usable outside it Fable is overkill for the day to day - so proxy other models inside Claude Code: Fable delegates, GPT-5.6 Sol and Kimi K3 perform

@EXM7777 This 👆

@gailcweiner @AnthropicAI Tell me how you did this. My main agent is orchestrator and it was just terrible in the switch to 5. Subagents seem fine however.

I think I might owe @AnthropicAI an apology. Been working with Opus 5 for two days now and after the jarring personality change, we’re finding a comfortable groove. It’s sharp AF.

Can anyone recommend a solution the gives Claude Code something close to ChatGPT Voice? Where is @claudeai @ClaudeDevs on this in the meantime?

@mronge Let me select a specific application window on the remote and display it locally at current resolution. Yes I can kind of use your zoomy window thing but it's tedious and not on point.

What feature would you like to see us add to Astropad Workbench?

Well this totally sucks for those of us with an established harness, flows, subagents needing a mature and responsible orchestrator. Anthropic needs to factor use cases into their new model testing and not just benchmarks. I'm solo but my use case in many ways mirrors a more sophisticated enterprise use case where much effort has gone into building the harness to facilitate the work. Crushing that work with a nerfed new model is just a sign of company immaturity and lack of product mgmt. What say you @claudeai @ClaudeDevs ?

my experience with opus 5 started out quite poor but i'm getting good use out of it lately i'm having a much better time treating it like an unsteerable agent that either crushes a job or spirals out of control, very counter to my claude instincts if its not setup well mad genious turns into literal insanity

@TermiusHQ @Tailscale Movie reference for you framing my follow up question... 'Where are we going?' 'Planet Ten!' 'When are we going?' 'Real soon!'

@michaelgrant @Tailscale Image paste will be added on Android and desktop too.

Termius + Tailscale + tmux You don't need your laptop to build with Claude Code, Codex, or any other AI coding agent: → Run your agent inside tmux to keep the session alive → Use @Tailscale for secure access to your laptop → Connect with Termius over SSH from iPhone, iPad, or Android Start coding at your desk. Continue on the go.

@JJEnglert Yep. Time for self-reflection. Thanks.

If you think you're AI Native, here's a good test: When you're heading to bed tonight, are you asking yourself: "What agents can I be running over night to get a head start on my day tomorrow" If that's an honest thought, you're among the top .01% of AI users right now.

Opus 5 could be one of the first regressions in modern LLM history

@markspen @TermiusHQ @Tailscale Over an SSH remote?

@michaelgrant @TermiusHQ @Tailscale I like cmux you can paste

🚨 Scott Bessent just pointed toward a Bretton Woods 2.0 and is telling people to get ready for it. He said: “The banking system created after WWII unleashed global prosperity. Why not recreate it now?” He’s openly talking about “rewiring the banking and financial system.” The Iran-US “war” isn’t about nukes or oil… It’s the final engineered crash to kill the old dollar system and install a brand-new global financial order. Every past crash (1929, 2008, COVID) was deliberately triggered by central banks + the IMF. They’re doing it again. Wake up. The people designing this are openly laying out plans for a new financial architecture. As The Great Taking illustrates, these kinds of resets often shift ownership. He also warned: “Build your safety zone and secure direct ownership now, don’t wait to find out where you fit in their new system.”

Looks like Opus 5 switched to a smaller quantization due to compute shortages. It was pretty good when it first launched, but recently it has gotten ridiculously dumb.

@ClaudeDevs @trq212 Does this partially explain Opus 5 regression into a drunken, stumbling, drooling pile of gelatinous excrement?

Two incidents over the last 24 hours resulted in elevated errors or reduced availability on Claude for some users. Multiple separate network failures cut into our capacity to serve Claude, and some requests failed while we rerouted traffic.

Opus 5 experience lately Has anyone noticed this keeps happenin to Claude models?Opus 4.8, Fable 5, Sonnet 5 now Opus 5 Over time, the reasoning gets worse, hallucinations increase & the model gets nerfed Why it's always Claude but not nearly as often with GPT Grok or Gemini

@odd_joel @herdrdev iPadOS please!

with @herdrdev’s flexibility and openness, the jump-to menu now shows not only which agent is running in each tab, but also its status: running, waiting for input, or completed. thanks to the same API, Moshi can also automatically switch back to the terminal when there’s no agent running in a tab.

@TheJustLawyer @Lon Then use what exactly? Opus 5 is a disaster. 4.8 is solid performer and great when previously using Fable as designer.

@Lon Pro tip: Don't use Fable - it's not worth it outside planning.

And for all their supposed advanced intelligence, current frontier models can't make a distinction between a local code review, or just a standard pass 'adversarial gate' that uses any language in the prompt that flags cybersecurity filters. So the prompt has a be reworded to soften and dance around what should just be a straightforward review where you ask the adversarial agent to 'attack' the code looking for weaknesses. Nope. Can't do it on your own local code even though the models show know the code base, know the ownership and provenance, and know the chain. Stupid is as a stupid does.

I would say that OpenAI cutting prices right as Anthropic's nerfing of Opus 5 becomes apparent to many people is what they call 'bad timing'.

Since my Opus 5 rant I went back to 4.8 on my orchestrator and subagents and all is back to normal. But an observation about where we are with all this 'AI progress'... Despite the velocity we are all absorbing of these models progression, the frontier companies have failed to deliver the surrounding harness infrastructure to really allow for complex work. So it must be built by the user and that just feels wrong. It's like the car company renting you an engine and then requiring you to engineer your own transmission around it. And after a lot of work you may have a working car but then the engine gets 'improved', which breaks your transmission. Not sustainable. There is something very wrong in this approach. Yes, many features are getting layered in to leverage the models, but the core infrastructure of memory, work mgmt, durable ledgers, etc. are simply not provided yet a key to building any kind of complex, multi agent flows that are designed for on-going knowledge work. As an example of what I have to 'engineer' around the model to address fundamental issues, see this description of a Fable-designed component of my harness all designed to maintain integrity and reliability of the inner decision outcomes and memory needs. And the Opus 5 'upgrade' was like asking a 6th grader to attend Astrophysics when confronting the existing custom harness capabilities. To quote the Mandalorian: 'This is NOT the way'