# Work I used to give juniors now takes Opus 5.5 minutes. I don't think we've clocked how far past GDPval we are — retained record

> Retained reddit record from reddit:singularity.

Canonical: https://highsignal.app/data/records/219fea5f525f73fd

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:singularity)  published 2026-09-30 11:22 ingested 2026-10-01 02:41

[original source ↗](https://www.reddit.com/r/singularity/comments/1wu2bxh/work_i_used_to_give_juniors_now_takes_opus_55/)

## retained text



```
Back in December, GPT-5.2 won or tied against industry professionals on 71% of tasks. By April, GPT-5.5 had hit 85% and Opus 4.7 80%. Then OpenAI quietly stopped publishing its own human-graded number and we just moved on. (GDPval-AA v2.1 published by artificial analysis is AI-graded so i'm leaving that out)

I know these are well-scoped tasks and not whole jobs. But they're a good chunk of my working day, and the newer models have only pushed further since. With Opus 5.5, there are tasks I'd previously have delegated to junior colleagues that now get done in minutes, with minimal errors. 

I barely build things from scratch anymore. I describe what I want, the model produces it, and I review. Going by my reviews, it gets it right 95%+ of the time.

Looking forward to Fable 5.5 and other subsequent models. I imagine more and harder domains will reach a similar point.

It feels like normalcy bias to me. Each model is only a step up from the last, so nobody feels how far we've actually come.

How are you all finding it on your day-to-day office work? Are you finding yourselves delegating to AI more and reviewing, as opposed to a lot of doing? Keen to hear thoughts
```



document fetched 2026-10-01 00:34 · canonical https://reddit.com/r/singularity/comments/1wu2bxh/work_i_used_to_give_juniors_now_takes_opus_55
