Claude API raised max request size for Messages and Token Counting from 32MB to 128MB. Image-heavy requests like long computer-use sessions and big PDFs now have ~4× more room.
Key points
Max request size raised from 32MB to 128MB.
Image-heavy requests like long computer-use sessions and big PDFs were previously hitting 32MB with plenty of context window left; now you get ~4× more room.
Sources
Grok 4.7 improves Sincere, reduces Slop and Sycophantic
Grok 4.7 has Slop 6, Sycophantic 1, and Sincere 90. Across Grok 4.5-4.7, Slop fell 62 percentile points, Sycophantic fell 17, and Sincere rose 64.
Key points
Slop: Grok 4.5 68, Grok 4.6 38, Grok 4.7 6
Sincere: Grok 4.5 26, Grok 4.6 63, Grok 4.7 90
Sources
LFM3-Omni-2B runs locally in browser at 280 tok/s
LFM3-Omni-2B runs 100% locally in a browser, achieving ~280 tokens per second on an M4 Max. It uses custom WebGPU kernels and handles visual, audio, or text input, outputting interleaved text and audio tokens.
Key points
Performance: ~280 tok/s on M4 Max, 90 ms to first token, 140 ms to first audio.
Capabilities: first true omni model running end-to-end locally, accepts visual, audio, or text and produces interleaved text and audio tokens.
Sources
Claude Opus 5.5 hits 1629 elo on DeepRSI
Claude Opus 5.5 hits 1629 elo on DeepRSI, a new industry standard for measuring recursive self-improvement. It outperforms Fable 5.1 with roughly 25% lower cost.
Key points
DeepRSI score: 1629 elo for Claude Opus 5.5
Cost: about 25% less than Fable 5.1
DeepRSI task duration: agents work 48 hours on ML tasks from ongoing proprietary research
Sources
NVIDIAAI xToken recipe boosts Nemotron 3.5 pass@1
NVIDIAAI xToken recipe boosts Nemotron 3.5 Super VL pass@1 from 6.67% to 11.83% in NeMo Gym. Qwen3.5-9B also improves from 4.40% to 8.20% on the same metric.
Key points
Nemotron 3.5 Super VL: pass@1 rises from 6.67% to 11.83% in NeMo Gym.
Qwen3.5-9B: pass@1 rises from 4.40% to 8.20% in NeMo Gym.
Sources
IREN AI Cloud demo shows sovereign token flow
IREN AI Cloud demo shows sovereign token flow, using k0rdent control plane and NVIDIA DSX OS recipes. It includes automated cluster provisioning, private model serving with secure API token generation, and full-stack observability.
Key points
k0rdent control plane and NVIDIA DSX OS recipes power the demo.
Provides automated managed cluster provisioning, private model serving & secure API token generation, and full-stack observability.
Sources
Vibecode adds multi-model app building and in-chat preview
Vibecode now builds iPhone or web apps from ChatGPT, Claude, Codex or Cursor and tests them. In ChatGPT you can tap through the finished app directly in the chat.
Key points
Multi-model support: builds apps from ChatGPT, Claude, Codex or Cursor
In-chat preview: tap through the finished app inside ChatGPT
Sources
OWASP Global AppSec USA schedule next week with session times
OWASP Global AppSec USA lists four sessions next week. The schedule includes a Zero Day Off community night on Nov 4 at 5 pm PT and three talks on Nov 5 and Nov 6 with specific times and speakers.
Key points
Nov 4, 5 pm PT: Zero Day Off community night
Nov 5, 10: 30 am: @insiderphd on reluctant devs
Nov 5, 11: 30 am: Diptendu Kar on A04 crypto failures
Nov 6, 9 am: Katie debates auto-remediation
Sources
- Original post
- semgrep.dev - Linked in the post
