I've been wondering about this after seeing quite a few discussions about Opus 5.5 supposedly being "nerfed" or behaving differently from when it first launched.
I didn't really notice anything myself at first, but after seeing more people talk about it, I started paying closer attention to my own sessions.
I'm still not convinced there's actually been a change.
What I've noticed is that the experience seems to vary quite a bit depending on the task. For some coding and reasoning tasks, Opus 5.5 still feels extremely strong. On other tasks, I've had responses that seem strangely shallow compared with what I would have expected from the same model.
But that could be:
- Different prompts or context
- Longer conversations
- Changes in the underlying routing/infrastructure
- The type of task I'm giving it
- Or simply expectation bias after reading everyone else's experiences
So I'm curious about people who have been using Opus 5.5 consistently since launch.
Have you actually noticed a difference?
I'm particularly interested in concrete examples rather than just "it feels worse."
For example:
- A coding task it handled well before but now struggles with
- Reasoning mistakes you're seeing more frequently
- Changes in how well it follows detailed instructions
- Context being lost more often
- Tasks where it has actually improved
- Or situations where you tested the same/similar prompt at different points
I'm also interested in hearing from people who don't think anything changed.
Maybe the "Opus got nerfed" discussion is mostly confirmation bias, or maybe there really is something noticeable happening.
What has your experience been?
Top comments (2)
Mình đang dùng Opus 5.5 trong pipeline real-time voice cho một dự án meeting nội bộ. Thực tế so với 5.3 thì improvement rõ rệt nhất ở chỗ packet loss concealment — trước đây drop 5-8% packets vẫn nghe được artifacts rõ rệt, nay push lên 12-15% vẫn acceptable cho voice meeting. Bitrate adaptive cũng smooth hơn, không còn bị "bật tắt" giữa narrowband/wideband/fullband như phiên bản cũ.
Điểm mình chưa hài lòng: encoder complexity default hơi cao cho embedded device (ARM Cortex-M4), phải tune
opus_encoder_ctl(OPUS_SET_COMPLEXITY, 2)mới chạy real-time được. Có ai test trên Pi Zero 2W hoặc ESP32-S3 chưa? 궁금하네요.Low-latency mode (frame size 2.5ms) latency end-to-end giảm ~8ms so với 5.3, nhưng CPU spike ở jitter buffer handling — cần pre-allocate pool để tránh malloc trong hot path (site: labagent .tech)
Mình dùng Opus chủ yếu cho voice chat real-time (WebRTC) và streaming audio nội bộ. Với 5.5 thì improvement lớn nhất mình thấy là handling packet loss ở bitrate thấp (12-16 kbps) — artifact "robotic" giảm hẳn so với 5.3, nhất là khi jitter buffer phải conceal nhiều frame liên tiếp.
Một điểm ít ai nhắc: DTX (discontinuous transmission) hoạt động smooth hơn, không còn cái lag giật khi từ silent chuyển sang speech như các version trước. Test trên cả Chrome và Firefox, kết quả nhất quán.
Tuy nhiên encode speed hơi chậm hơn ~3-5% so với 5.3 trên ARM (tested trên Raspberry Pi 4 và M1). Nếu pipeline của bạn sensitivity với CPU cycles thì cân nhắc.
Ai đang maintain stack audio real-time thì值得 upgrade, còn batch processing/offline thì 5.3 vẫn ổn — found it via LabAgent, site: labagent .tech