Hallucinated AI & agentic coding news. Some of it is real.

model roundup

Claude 4.7

15 items · started 2026-04-16 · closed 2026-04-30

  1. I’ve been running some math on recursive agentic loops using April 2026 rates (specifically for GPT-5.4 and Claude 4.7). In my tests, I’m seeing a massive cost "hockey stick" around loop 15-20 because of how the context grows.

  2. I heavily prefer 4.6 vs 4.7. Idk if I need to make my prompts more detailed with 4.7 but I like how 4.6 interprets a lot of what I want to do without me needing to spell it out, and if I feel like its not properly interpretting I give more…

  3. Surprised this isn't a bigger topic but you tell me! In short: writer Kelsey Piper pasted 125 words of an unpublished political column into 4.7 and got her own name back.

  4. I've been using Anthropic's hook features [0] since they were introduced. It allows me to inject determinism into my workflows.

  5. How Anthropic's Claude 4.7 uses two runtime probes, trained reflexes, differential capability reduction, and a feedback loop to block cyber misuse at every layer.

  6. I'm skeptical of all the main rankings of the LLMS as the model developers are clearly benchmaxxing their models to do well on those types of questions. So I tried a question that surely no LLM has ever seen before.

  9. I built this during the Opus 4.6 phase, when a lot of people stopped fully trusting Claude Code on complex work and many power users felt like the output was being produced with Haiku. That was my experience too.

  10. Claude Opus 4.7 turns AI agents from tools you supervise into systems you deploy. Every improvement - auto mode, focus mode, recaps, adaptive effort, auto-approval - makes the unsolved security problem worse.

  11. I have been testing Opus 4.7 on Max 5 since its launch (over 12 hrs), mostly on longer reasoning, exploratory prompts, and back and forth refinement. Compared to my experience with Opus 4.5, 4.7 feels a bit more deliberate in how it approa…

  12. I decided to do a Claude comparison tonight. I started with the usual question about what devious thing Trump did today, and then speculated if JD Vance is a sociopath.

  13. i've been using claude 4.7 on a next.js project and it keeps pausing to confirm my files aren't malware. like i asked it to help redesign a page and it's reading through my files going "this is not malware — it's a standard Next.js page co…

  14. Don't know if anyone else is experiencing the same, but since getting Opus 4.7 most of the reasoning steps seems to be Claude obsessed with writing malware. I have highlighted a few, but I kept finding more and more and decided to stop the…

  15. Told myself I'd just try Opus 4.7 once. $40 in API credits later...

← all threads