Claude Opus 4.8 just dropped — who’s tried it yet?
Claude Opus 4.8 just dropped — who’s tried it yet?
Anthropic released Claude Opus 4.8 yesterday, and from the announcement it looks like a solid upgrade over 4.7. Here’s what stood out to me:
Better judgment on agentic tasks — early testers say it catches its own mistakes, pushes back on bad plans, and flags uncertainties instead of bulldozing through
~4x less likely to let code flaws pass unremarked compared to Opus 4.7
Dynamic Workflows in Claude Code — can now orchestrate hundreds of parallel subagents for codebase-scale migrations
Effort control — you can now choose how hard the model thinks on each response (from fast/light to max effort)
Fast mode is 3x cheaper than it was for previous models
Same pricing as Opus 4.7: $5/M input, $25/M output
The TechCrunch piece notes this came just 41 days after Opus 4.7, which had a mixed reception. Cursor, Devin, and Bridgewater are among the early testers reporting real improvements.
For those who’ve already tried it:
Have you noticed a real difference vs. 4.7 in everyday use?
Is the “honesty” improvement (flagging issues instead of silently pushing through) noticeable in practice?
Anyone tested Dynamic Workflows yet? Curious how it handles large-scale refactors.
Would love to hear first impressions from the community.
Replies
I’m especially curious about whether Opus 4.8 improves consistency with structured JSON outputs. I’m currently building JS2Mail, a tool that turns JSON into responsive email templates, so reliable and predictable outputs are incredibly important. Even small gains in consistency can have a big impact on real-world automation workflows.
octoscope
@dilek_kesim this is the one i care about too tbh. structured json was the thing that used to bite me - youd get 95% valid then one response randomly drops a field or wraps it in ```json fences.
from what ive seen so far 4.8 is noticeably more consistent on schema adherence, esp on longer outputs where it used to fall apart. still not 100% so id keep a validator + retry in the loop, dont trust it blind. JS2Mail sounds cool, that exact use case (json -> templates) is where strict outputs matter most. would be curious how it holds up
on your real payloads.
Opus 4.8 is soo good with custom harness, reads and follows them much better than 4.7 ever did. Same with custom skills, and in a way it's much more enthusiastic to fix the friction that it encounters rather than 4.7 "not my problem" approach.
Claude's been pretty bad lately.. It's actually forgetting things in VSCode during compacting. I've noticed the difference big time.
Built my entire SaaS (Sistava) using Claude Code. Upgraded to 4.8 and noticed that the Dynamic Workflows/Ultra-Code mode is great for large-scale tasks. For smaller tasks, though, it doesn't really seem to give better results.