Last week, a lot of people reported that Opus 5.5 was nerfed [1], right before the release of Fable 5.2. Unfortunately, benchmarks like livenerf [2] didn’t detect it.
model
Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2
83887 downloads312 likesimage-text-to-text
from the model card
Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 A model post-trained with several rounds of SFT and RLOO on top of Qwen3.8-27B-EfficientThink-SFT-SimPO-DFlash2, which is itself the official Qwen3.8-27B post-trained with SFT and SimPO. It is built to fix the original model's habit of failing to stop: it already has an answer, yet keeps re-deriving until it hits the 94K cap with no final answer. This repository provides five safetensors tiers, BF16, static FP8 Block128, NVFP4 W4A16, NVFP4 W4A4, and NVFP4 W4A4-W8A8 (mixed precision), all fully multimodal (text + image/video) with the official BF16 MTP bundled; all load in SGLang without patches, and the FP8 package is also smoke-tested on vLLM. It also has NInfer W4A4 / W4A4-W8A8 (NVFP4-NInfer/), INT8 W8A8 (INT8/W8A8/), INT8 W8A16 (INT8/W8A16/, text + image), INT4 W4A16 (INT4/W4A16/, text + image, vLLM), and GGUF Q2 / Q3 / Q4 LynnStyle (plus Q8 MTP versions), Q6K, and Q80 (GGUF/, GGUF-NInfer/, built-in MTP). Far fewer 94K truncations: under the same protocol, 4 → 1 on GPQA and 13 → 3 on LCB. No regression on any suite: GPQA 178/198, MMLU 450/500, LCB 90/100 (static FP8), vs. 177 / 444 / 83 for the original FP8. Lossless quantization: GPQA and LCB are identical across BF16, dynamic FP8, and static FP8. Two speculative-decoding options (mutually exclusive): about 180 tok/s per request wi…
discussions
- Opus 5.5 Launch 1535 ongoing since 2026-09-21
- DeepSeek v4.1 Incident 8 ongoing since 2026-11-11
- GPT-6 Astra Release 292 ongoing since 2026-09-03
- Grok 4.7 Release 6 2026-11-07 – 2026-11-14
- Model Selection Update 4 2026-10-23 – 2026-10-29
- Grok Disable Announcement 4 2026-10-16 – 2026-10-22
recent items
Tell HN: Opus 5.5 is no more nerfed (news.ycombinator.com) Claude Code wrote ~100k lines of Swift for my Mac dock in 9 days. Here's how I ran it (www.reddit.com via reddit) I'm the developer of Petal, a free, open-source dock for macOS where you can tag Claude on a sticky note and it does the task. Claude Code wrote basically all of it, so here's how that actually worked.
ChatGPT Pro $100 vs Claude Max $100 (www.reddit.com via reddit) So currently, I have Claude Max $100, with ChatGPT Plus, and I'm thinking of switching to ChatGPT Pro $100 because it gives 10x usage compared to Plus, and from my testing, Plus consumes around the same rate as Max on the same tasks, maybe…
Need suggestion on 3d Ai (www.reddit.com via reddit) Hi there guys. I need suggestion on from reference to 3d model, texture, rigging and animation for my game project.
Cursor Spike Usage and Strange behavior, when selecting model in project , the agent used different model. (www.reddit.com via reddit) I had 9% of my weekly usage limit left, which was set to reset today. I decided to use it wisely by selecting Composer and asking it to build a medium-sized feature (a reservation calendar).
3D Gaussian Rendering of Ibuprofen Liquid Softgel (www.reddit.com via reddit) I built a 3D Gaussian splat of an ibuprofen liquid softgel from 189 turntable photos, using Claude Code (Opus 5.5 on medium). It's free and open source (MIT): my github repo has the raw photos, every script, the exact settings, and an offl…
bypass permissions doesn't actually bypass the safety check?? got blocked opening a webpage (www.reddit.com via reddit) on the desktop app (windows), code tab, opus 5.5, permission mode is set to bypass permissions. asked claude to open a login page in chrome and it just kept failing with "the automatic safety check gave no verdict on this call".
Opus 5.5 Nerf isn't that simple (www.reddit.com via reddit) Opus 5.5 comes out. Everybody is happy.
Cap de leitura (www.reddit.com via reddit) Descobrir que o Opus 5.5 tem um cap de 330 carácteres e infere o resto com o contexto, sem indexação de código local isso faz aquelas perdas de contexto que te irritam tanto fazer mais sentido. Minha dúvida é, fable tem esse mesmo cap no a…
Envisioning Agentic Robotics: A New Era of Robotic Systems (arxiv.org) The release of GPT-6 Astra marks the emergence of a new paradigm in robotics: agentic robotics. Within weeks, frontier models with agentic harnesses matched or exceeded task-specific policies on many embodied benchmarks.
Token usage help (www.reddit.comhttps) Use Claude daily for work and run multiple projects for each client. I have quite a few connectors active, mostly fire crawl, gmail, hubspot ect.
Agentic AI Designs an Analog Synthesizer (cpldcpu.github.io via hn) Can we just “vibe-engineer” actual analog circuits in hardware? When GPT-6 Astra and Claude Opus 5.5 were released in September 2026, it became quite clear that the models’ capabilities in spatial reasoning and engineering in general had s…
Claude refused to solve a physics problem (www.reddit.com via reddit) So i have a Uni assignment that asks me to solve a problem specifically with AI, then document it and criticize it's work. The task is electromagnetic analysis of a specific type of conductor.
I built Replicator (3d workshop) with Fable 5.2 and Claude Opus 5.5 (www.reddit.com via reddit) I built Replicator, your local first 3d workshop and research center with Fable 5.2 as the planner (1 prompt on what I wanted - detailed meta prompt from Astra) then I switched from plan to Opus 5.5 and let it rip with subagents. Built ove…
Great use of tokens on a pull request (www.reddit.com via reddit) https://preview.redd.it/24r7p1y8eu1i1.png?width=1072&format=png&auto=webp&s=82a8d8c948e28827592b82b861c27d2c3f90d507 Working on my Oblivion Remastered Performance Fix and merging some PRs submitted by the community, Claude Opus 5.5 randoml…
Chess match: Claude Fable 5.2 vs GPT 6 Astra. Astra won again, this time with Claude as Black (www.reddit.comhttps) A follow-up to my earlier Claude vs OpenAI chess matches. Last time Fable 5.1 lost to GPT 6 Astra, so I ran a rematch with the newest Fable 5.2 and swapped colours: Claude plays Black this time.
Best model to implement a license system? (www.reddit.com via reddit) I need to implement a license system (maybe online) for a PC app I have. I have 30% Fable 5.2 usage left of my Pro account.
Fable 5.2 prose feels better than Opus 5.5 (www.reddit.com via reddit) Already reached my limit after "speaking" with Fable about an hour on a niche topic. I have to say it truly felt like I was speaking with a capable person.
I made a video downloader using Claude because I did not like the other options (github.com via reddit) I didn't like the other available options out there, and I was annoyed about how some applications charge just to download, so I started off using "Craft Anything" by ArtCraft, which was using claude code under the hood, to make me an alte…
First timer making a game, need advice on QA Audits that are destroying my tokens and don't provide particularly great results (www.reddit.com via reddit) First timer in making a game. I am using Claude Code and Opus 5.5.
VR island adventure game - Fable 5.2/Opus 5.5 (www.reddit.comhttps) Mission 04, "Black Tide", from a VR island adventure. The whole mission was played in one go (about 10 minutes, plane to plane) and cut down to 1:45.
Opus 5.5 literally flagging everything as "reasoning extraction", my workflow is frozen, anyone experiencing the same? (www.reddit.com via reddit) Anthropic is literally flagging every other message and it is next to impossible to actually sit and change the workflow so as to de-trigger the flag. I posted yesterday as well https://www.reddit.com/r/Anthropic/s/s4A3Z8oydv .
I tested Claude Fable 5.2 against Opus 5.5 on real tasks. The cost difference was interesting. (www.reddit.com via reddit) TL;DR — What I found: Fable 5.2 was much faster at building the 3D room (7 vs 20 minutes). Both models handled the warehouse simulation safely, but Opus was more efficient and cheaper in that run.
Addicted to Fable 5.2 - Should I switch from $20 to $100? (www.reddit.com via reddit) I am on the $20 plan since Opus 5.5 came out, I really like Opus 5.5 but it just wasn't there regarding production ready 3D modeling using FreeCAD. Now with the limited time I have access to Fable 5.2 I really noticed that it excells at 3D…
Claude wrote and trained a language model for DOS that fits on a floppy disk (www.reddit.comhttps) I let Fable 5.2 write its own transformer. Opus 5.5 generated the training data: about 3500 facts (and small talk + math), each asked and answered in different ways.
310M Claude tokens later: why I won't claim a precise subscription-vs-API savings multiple (www.reddit.com via reddit) I went looking for a clean “Claude subscription saved me X times the API bill” number. My router history cannot honestly support one, and the missing data is useful to show.
Opus 5.5 running for more than 24 hours for a refactoring task (www.reddit.com via reddit) I used opus 5.5 (max) and gpt 6.1 Sol (max) to write code for a problem for over a month. I made progress in solving the problem, but realised that the code base is super bloated, and looks over engineered.
Claude Fable 5.2 xhigh is worth the extra tokens (www.reddit.com via reddit) Fable 5.2 is the same model family as 5.1. Unlike Opus 5.5, where medium and high are the sweet spot, Fable 5.2 high~xhigh is the sweet spot.
I built a web racing game for my vinyl business in 1 week without Fable (www.reddit.comhttps) https://underground.auravinyl.com Vinyl is my small side business I made specifically to employ my parents. I've been working in interactive experiences for almost 20 years, since I was a teenager back in the Flash days.
What a task costs on Opus 5.5 (claude.dev via hn) The same token price can cost very different amounts per task. Learn what Claude Code tasks cost on Opus 5.5 and which settings change the bill.