Sonnet 4.6

1. Claude Sonnet hits 100% comprehension on a data format it's never seen. Opus scores 96.2%. We tested 10 models across 3 providers. (www.reddit.com via reddit)

   15h haikusonnetopus

   I built a wire format called GCF and tested whether LLMs could read and write it without any prior training. I sent 10 models the same payload: 500 symbols, 200 edges.

2. Claude manually writing base64 burning tokens, drive connector. (www.reddit.com via reddit)

   1d sonnet

   When I prompt to create a .docx and upload it to Drive, Claude writes the output as Base64 manually instead of uploading it directly to Drive. Has anyone experienced something similar?

3. opus 4.8 vs sonnet 4.6 for the dashboard analytics engine. opus improved the trend analysis. sonnet still handles the routine summaries. the model split matters. (www.reddit.com via reddit)

   1d sonnetopus

   saas. 310 customers.

4. Sonnet 4.6 Max - unable to follow instructions to a T (www.reddit.com via reddit)

   2d sonnetclaude-code

   Even when given instructions like this: ``` CRITICAL REQUIREMENTS: Read EVERY section from start to finish—no sampling, no skimming If you cannot process all 234 sections in one response, STOP and tell me Process in batches of [X] sections…

5. Gemma 4 31B QAT Q4 vs standard Q4 — Top1 KLD benchmark results have me confused. Someone please explain or poke holes in this. (www.reddit.com via reddit)

   2d gemmasonnet

   I'll be upfront: I vibe-benched and vibe-reported this with Claude Sonnet 4.6, but I reviewed and edited everything before posting (too lazy to take out all the AI EM dash —), so hopefully nobody considers this AI slop. And more importantl…

← all threads