Is Day 17 gone? Did Tibo fall asleep and forget?
Model intelligence feed
Model intelligence.
Follow the changes. Understand what comes next.
Editor’s pickOther providers
Valve's Answer to AI Slop and Clone Games on Steam: We Won't Police Developers' Tools
At Steam Indie Connect in Korea, Valve staff addressed AI slop and clone games head-on: no gatekeeping on AI tools, and the best defense against clones is community and steady updates.
Latest stories332
Newest published131415161718192021222324
Step 5 Preview Cache Hit Price Doubles Nov 6—Why Your Total Bill Might Not
Tencent Cloud's TokenHub & CloudBase notice raises Step 5 Preview (step-5-preview) cache-hit input from 2026-11-06 00:00:00 Asia/Shanghai: Guangzhou 0.35→0.70, Singapore 0.36→0.72 RMB/M tokens. Miss input and output stay unchanged (numbers not listed). As of 2026-11-04, still a preview—not yet live. A doubled cache-hit line ≠ a doubled total bill; impact tracks hit share. TokenHub/CloudBase only, not a StepFun all-channel hike. DigIntoAI has not checked bills or hit rates—figures follow the Tencent Cloud notice.
step-5-preview／TokenHub／CloudBase／缓存命中输入价Read story
Tencent Hy4 Hits TokenHub and OpenRouter—Vision In, 50% Off Through Year-End
TokenHub listed formal Hy4 (hy4) on 2026-11-04 12:13:30, adding image understanding vs Hy4 preview, with 1M / 960k / 64k, deep thinking, structured output, Function Calling, and Cache. List 6 / 18 / 0.3 RMB/M tokens; documented 50% through 2026-12-31 23:59:59 Beijing time → about 3 / 9 (estimate; bill unchecked). OpenRouter tencent/hy4 at 2026-11-04 12:20:31 Asia/Shanghai (created=1793766031): ~$0.417 / $1.251/M, 50% off, text+image→text. DigIntoAI has not called Hy4.
Hy4／hy4／Hy4 preview／TokenHub／OpenRouter tencent/hy4Read story
Claude Text Watermarks Now Cover Mythos 5, Opus 4.7, and Sonnet 4.6—Detection Still Private
On 2026-11-03, Claude Platform release notes state that text from Claude Mythos 5, Claude Opus 4.7, and Claude Sonnet 4.6 carries Anthropic’s text watermark—an expansion of coverage, not a brand-new marking stack. Help Center tables list those models (and others); Amazon Bedrock rollout completes by 2026-11-10; Sonnet 4.6 with non-default temperature／top_p／top_k may omit the mark. Watermark detection remains private preview for eligible organizations. DigIntoAI has not retested detection, sampling edges, or Bedrock region progress.
Claude Mythos 5／Claude Opus 4.7／Claude Sonnet 4.6／Text watermarkRead story
AWS Distills GLM-5.3 Into a Qwen3.5-4B Shopping Agent—Near the Teacher, With Tradeoffs
AWS Machine Learning Blog (02 NOV 2026; Zappia et al.) shows SeqKD of a shopping agent: Bedrock GLM-5.3 teacher → ShoppingBench rejection sampling → LoRA SFT on Qwen3.5-4B (SageMaker AI). Held-out 324 problems, single LoRA run: student ASR 0.541 vs greedy teacher 0.540 vs base 0.297; tool-only 0.415; IFEval ≈6–8 pts down. No per-task cost/latency published. AWS-reported; we haven't retested.
SageMaker AI／Bedrock／GLM-5.3／Qwen3.5-4B／ShoppingBench／SeqKDRead story
Claude Code 2.1.318 Adds Login Expiry Warnings and Stops Auto-Installing IDE Extensions
Anthropic pushed Claude Code 2.1.318 to npm as latest (same as next). Package time 2026-11-03T21:55:17.247Z (2026-11-04 05:55 Asia/Shanghai); modified and GitHub Release v2.1.318 land around 2026-11-04 08:01 Asia/Shanghai. stable stays 2.1.311. CHANGELOG: loginExpiresAt / credential_expires_at plus a 10-day stderr warning; IDE terminals no longer auto-install or update the IDE extension (/ide or the editor extensions view). DigIntoAI has not installed or exercised this build.
Claude Code／2.1.318Read story
Claude Pricing (November 2026)
Claude.ai subscription map for November 2026 (USD, ex-tax): Free $0; Pro $20/mo or $17/mo with $200 annual prepay; Max $100/mo; Ultra $200/mo. Comparison table: Free Limited/Limited on 5-hour and weekly windows; Pro 1x/1x; Max 5x/5x; Ultra no 5-hour limit for verified accounts / 10x weekly. Max includes $100/mo API credits; Ultra $200/mo. Team Standard $20/$25 per seat; Premium $100/$125; seats mixable for 2–150 people; pooled Team API credits cap $500/mo. Enterprise US$20/seat/mo annual with 20-seat minimum plus usage at API rates. Separate from the Claude API model wall. For reference only.
Claude 订阅档（Free／Pro／Max／Ultra／Team／Enterprise）Read story
Why Does the Same DeepSeek Score So Differently Across ARC-AGI-3 Harnesses?
On the ARC Prize official results page for DeepSeek V4.1 Flash 1017 (page label Oct 17, 2026 · 5 harness configurations), the same model at the same max reasoning setting scores 0.2% on ARC-AGI-3 under Standard and 5.8% under Provider Adapter. The gap is the evaluation harness—kept notes versus opaque state plus compaction—not a new weight drop. Same-page Semi-Private peaks: v1 high 94.5%, v2 max 78.1%. Figures are ARC Prize self-reports; we haven't retested.
DeepSeek V4.1 Flash 1017；ARC-AGI-3 Standard／Provider Adapter harnessRead story
How Do You Prove an API Didn't Swap Models? Anthropic Opens FlashProve
On 2026-11-02, Anthropic’s Frontier Red Team published *Progress on provable inference for large language models* and open-sourced FlashProve: a zkSNARK research prototype for Commit→Prove→Verify. A provider can commit to weights without publishing them, then attach a checkable proof that a reply came from those committed weights. The proof covers model identity and the inference path—not factual correctness. Official Maverick figures are self-reported; we have not retested. This is not a Claude API feature already live for all users.
FlashProve（研究原型）Read story
Perplexity Opens Multimodal Retrieval Embeddings—Why Document Chunks Need Context
On 2026-11-03 09:09:18 Asia/Shanghai, pplx-announcements posted open-weight vision-language embeddings pplx-embed-v2-context (9B / 0.6B). Contextual mode encodes chunks with surrounding document context; both sizes share a 1024-d MRL prefix. Paper arXiv:2610.38854 was submitted 2026-10-28—not a same-day first appearance. Ordinary vs contextual chunking and encode_queries vs encode are laid out below. Card SOTA lines are self-reported; we have not retested recall, VRAM, or multilingual retrieval.
pplx-embed-v2-context／pplx-embed-v2-context-9b／pplx-embed-v2-context-0.6bRead story
Can SenseNova 6.8 Flash Deliver Editable Dynamic PPT? A From-Generate-to-Revise Checklist
SenseTime’s SenseNova site shows SenseNova 6.8 Flash as a formal release for lasting office work and dynamic PPT, posters, pages, and 3D. IT Home ties the announce to the evening of 2026-11-02. The detail page sells narrative-timed PPTX motion in one pass, but office case captions cite Xiaohuanxiong’s internal R&D build—not every-account proactive office. This is an untested generate-to-revise checklist; we have not measured delivery or revise quality, and have not re-run vendor or press scores.
SenseNova 6.8 Flash／SenseNova Studio／SenseNova Token Plan／商汤小浣熊／SenseNova SkillsRead story
Deep Principle Opens Hydrogen-1.0 on GLM-5.3, Open Weights Still Need a Runtime
Deep Principle’s blog (2026-11-02) shipped Hydrogen-1.0: scientific post-training on GLM-5.3, open BF16/MIT weights—not a Zhipu flagship. HF main is gated=auto; ModelScope mirrors weights; NeuroSpark/Mira are separate. Cards show SGLang/quant deploys, but hubs ≠ hosted chat, and full data/recipes are not shipped. Blog scores = self-reported; we have not re-run them.
Hydrogen-1.0／Hydrogen-1.0-NVFP4／Hydrogen-1.0-MXFP4／GLM-5.3／deep-principle/Hydrogen-1.0Read story
OpenAI Plans Next Security Sprint and Raises Bug Bounty Rewards
On 2026-11-03 ~10:29 Asia/Shanghai, OpenAI plans a next internal security sprint with Cathedral Systems and raises Security Bug Bounty rewards. Blogs (November 2, 2026): midpoints for medium/high/critical more than quadrupling on average; ChatGPT/Codex/API critical up to $30,000; exceptional eligible up to $172,000 with bonuses. Make-ups for awaiting awards and past-30-day payouts. DigIntoAI hasn't re-run the bounty math or reproduced vulnerabilities.
Security Bug Bounty／ChatGPT／Codex／API PlatformRead story
From the Webmaster
Personal experiences, thoughts and notes.