Issue No. 9
The Nov 6 edition
Snapshot#
- Thinking Machines shipped Inkling 2 Small on Tinker today (closed weights now; Apache 2.0 planned next week). AA scores it 31 at $0.30/$1.00 per 1M and $1.18 per index task — 17B active / 594B total MoE.
- GPT-6 Cyber (Daybreak Red) is OpenAI's gated cyber-specialized GPT-6 for authorized vulnerability research. Official ExploitGym: 48.5% vs GPT-5.6 Cyber 32.1%. Access is Daybreak Red customers via API/Codex only — not a general coding daily driver.
- Claude Code privacy incident: issue 107757 (1,305 HN points) — signed-in email leaked into outbound
User-Agentduring web/curl tool use. Anthropic is shipping a privacy mod and plans to remove account email from model context. Community hook: whoami-guard. - OpenRouter movers: official Tencent Hy4 (AA 34.6 on the OR page, $0.417/$1.251 with 50% off, Terminal-Bench 4.0 39.4%); free stealth Crouton Alpha (1M context; Claude Code already routed 40.7B tokens to it); and Auto Code, a coding-agent router shipped Nov 5.
- Open weights for local/laptop: Nanbeige4.3-33B hits AA 20 (Apache, 33B dense, Terminal-Bench 4.0 30% vs Qwen3.8 27B 6%). Linum P-JiT is a ~$21K Apache image model on HN (255 pts).
- Harnesses: Codex CLI 0.167.0 adds an agents command center and Vertex setup wizard. Early Rust harness Synaps CLI (56 stars; author claims unverified).
- Still open DeepSeek V4.1 Flash 1017 remains OR usage king at 56.5T tokens this week (+33%). Conductor Cloud goes read-only 11-13 and shuts down 11-20.
Do this today: upgrade Claude Code and install a privacy guard (issue / whoami-guard), then try Inkling 2 Small on Tinker or route a coding session through OpenRouter Auto Code.
Frontier and open-weight models#
- News Inkling 2 Small open weights pending from Thinking Machines (HN, 235 points; AA article).
- Architecture: 594B total / 17B active MoE; vision + variable effort.
- Access now: Tinker with closed weights. Apache 2.0 weights planned “next week” after safety testing.
- Independent (AA): Intelligence Index 31; $0.30 in / $1.00 out per 1M ($0.03 cached); $1.18 per index task.
- Other AA numbers: FrontierCode 2.0 25.4%; DeepSWE v1.1 72.1%; AA-Briefcase Elo 1527.
- Vendor training notes (lab post): H1 pretrain 8.4T tokens in 68 hours on GB300 NVL72 native NVFP4; RL 17M+ rollouts / 9 days / 2048 GPUs.
- AA framing: highest-scoring open-weights model outside of China upon weights release (weights not public yet).
- News GPT-6 Cyber (Daybreak Red) gated frontier (confirmed via OpenAI + Agentic Digest).
- What it is: cyber-specialized GPT-6 for authorized vulnerability research; Critical preparedness cyber classification; gated Daybreak Red.
- Official eval: ExploitGym 48.5% vs GPT-5.6 Cyber 32.1%; out-of-scope avoidance 99.93% / 100% with deployment prompt.
- Access: select Daybreak Red customers via API/Codex. Not a general coding model.
- Scope note: OpenAI also published defender-oriented cyber eval numbers from their research pipeline. Exploit construction details are deliberately omitted here.
- Update Tencent Hy4 API live — official release on OpenRouter (released Nov 4; was preview-scored yesterday).
- OR page: context 1.0M; multimodal text+image in; price with 50% off $0.417 / $1.251 per 1M (list before discount $0.834/$2.501).
- AA on the OR page: Intelligence Index 34.6 (yesterday’s AA homepage teaser used 35 — indexes/surfaces disagree slightly; number named by source).
- OR page benchmarks: Terminal-Bench 4.0 39.4%; Vibe Code Bench 81.0%.
- Traffic: new on OR weekly trending; #2 today by tokens (+110% to 2.72T). Top apps include Claude Code, Strix, Hermes, pi.
- Weights: “coming soon” was claimed in Code4U digest only — not confirmed on the OR page.
- News Nanbeige4.3-33B open weights (model page).
- Model: dense 33B, Apache-2.0, 524K context, laptop-sized; no confirmed public API.
- AA: Intelligence Index 20 (ties Qwen3.8 27B xhigh).
- Vs Qwen3.8 27B (AA article): Terminal-Bench 4.0 30% vs 6%; CritPt Hard 16% vs 4%; AutomationBench-AA 53% vs 48%; FrontierCode 2.0 11% vs 15%; Omniscience hallucination 18% vs 29%; ~158K output tokens/task vs ~110K.
- News Crouton Alpha stealth, free on OpenRouter (released Nov 3).
- Price: Free during preview. Context 1.0M. Multimodal (text/images/video/files/audio → text); tool calling; ~33 tok/s P50.
- Usage signal: Claude Code already routing 40.7B tokens to it. Anonymous provider — fine for experiments, caution for sensitive workloads.
- News Gemini Omni 1.2 Flash multimodal video (Google blog, Nov 5).
- What: native multimodal video gen/edit; up to 15s clips (was 10s); duration 3–15s in 1s steps; multi-turn conversational edits; SynthID watermark; GA on Gemini API / Cloud.
- Price: 30% cut to $0.07/s at 720p.
- News Linum P-JiT open image model (Show HN, 255 points).
- Model: ~2.18B DiT params; Qwen3.5-4B text encoder; Apache 2.0; pixel-space Pyramid-JiT.
- Train cost (field notes): pretrain $20,348 + post $690 at $2.50/H100-h (~$21K / 660M samples).
- Limits (authors): no rendered text, weak multi-entity.
- News NVIDIA Cosmos 3 Nano interactive checkpoints open world sims (NVIDIA tech blog).
- What: four open-weight 16B post-trained checkpoints (Sim-Bimanual, Sim-Transfer, Sim-Auto, Transfer-Auto); OpenMDW1.1; recipes for vLLM-Omni / SGLang / Framework.
- Vendor note: NIM ~2–3× speculative-decoding speedups on H200.
- Still open DeepSeek V4.1 Flash 1017 usage — #1 weekly on OpenRouter at 56.5T tokens (+33%); #1 today 6.9T. No new model-card open this window; price not rechecked.
- Context Kimi K3.1 Pro weights Pareto baseline (AA 46) and Ripple-65B Preview remain yesterday’s open-weight anchors; still on HF trending / OR new&trending.
Pareto frontier (quality vs $/token vs $/task vs session economics)#
All index numbers are from AA unless named otherwise. Hy4’s 34.6 is the number on the OpenRouter model page; yesterday’s AA teaser used 35.
| Model | AA index | $/1M in / out | $/index task | Notes | Weights |
|---|---|---|---|---|---|
| Inkling 2 Small | 31 | $0.30 / $1.00 ($0.03 cached) | $1.18 | Tinker now; Apache next week | Open pending |
| Tencent Hy4 | 34.6 (OR page) | $0.417 / $1.251 (50% off) | not on OR page today | TB4.0 39.4%; #2 OR today | Weights unconfirmed |
| Nanbeige4.3-33B | 20 | no public API | n/a | TB4.0 30% vs Qwen3.8 27B 6% | Apache-2.0 |
| Crouton Alpha | not scored | Free | n/a | 1M ctx; 40.7B tokens via Claude Code | Stealth / closed |
| DeepSeek V4.1 Flash 1017 | (on AA since 10-30; not reopened) | not rechecked this window | n/a | OR week 56.5T (+33%) | Open |
| Kimi K3.1 Pro (Context) | 46 | $2 / $10 | $3.68 | 36 min/task | Open (custom license) |
- Takeaway: Upon Apache release, Inkling 2 Small is the AA-framed #1 open-weights model outside China at 31 / $1.18 per task — cheaper per task than Kimi (46 / $3.68) at 15 points lower. Hy4 sits near Inkling on index (34.6 vs 31) at a similar discounted OR token price; OR page does not show $/task today. Crouton Alpha is the free session experiment slot. Nanbeige is the laptop-dense pick when you care about Terminal-Bench relative to size. DeepSeek Flash remains the high-volume default by OR share.
- News OpenRouter Auto Code session economics / routing (created Nov 5).
- How it works: coding-agent router that treats every request as part of a session. Chooses by coding-task pass rate, completion time, and cost per solved task. Tracks edits/tests/errors; keeps the current model while work goes well; tries more effort before switching.
- Billing: underlying model prices; context up to 2M depending on routed model.
- News AA VIALS-AA vision+tools $/task (Nov 6).
- What: scientific image reasoning with tools; 161 tasks × 3.
- Leaders: Gemini 4 Carbon (high) 64% pass^3; GPT-6.1 Sol (max) 59% at $0.06/task vs Claude Opus ~30× cost for similar score.
- Cost range: $0.02–$1.77/task across the board.
- Chinese-lab and open-weight movers today:
- Hy4 (official OR, AA 34.6 on page).
- Nanbeige4.3-33B (Apache 33B).
- DeepSeek V4.1 Flash 1017 (usage king).
- Inkling 2 Small (US lab; Apache weights pending).
- OpenRouter weekly usage (through Nov 5, from rankings): DeepSeek V4.1 Flash 1017 56.5T; DeepSeek V4.1 Flash 16.5T; GLM 5.3 Flash 15T; GPT-6 Luna 7.26T; Step 5 Preview 6.67T; Hy4 preview 6.58T; Claude Opus 5.5 6.53T; MiniMax M3.1 Flash 6.04T; GPT-5.6 Luna 4.99T; Hy4 4.01T (new). Author share (week of Oct 26): DeepSeek 29.8%, OpenAI 18.8%, Google 14.4%, Z.ai 11.9%, Anthropic 5.7%.
Coding CLIs/IDEs/harnesses#
- News Claude Code email-in-User-Agent privacy / harness hygiene (1,305 HN points).
- What happened: on v2.1.318, signed-in email could leak into outbound
User-Agentduring web/curl tool use. Anthropic staff: email was added to context for @author tags; models sometimes put it in UA. - Fix path:
cc-plugin-privacydefault-on in ≥2.1.317 (users report gaps); next release will remove account email from context. - Immediate action: upgrade Claude Code; add an outbound hook such as whoami-guard. Do not treat the issue thread as an exploit guide — patch and guard.
- What happened: on v2.1.318, signed-in email could leak into outbound
- Update: Codex CLI 0.167.0 (changelog entry 2026-11-06).
- Install:
npm install -g @openai/codex@0.167.0. - New: agents command center (
codex agents --no-daemon); unread markers; GCP Vertex setup viagcp_setup_wizard;/experimentalsystem proxy; image thumbnails; compaction accounting fix.
- Install:
- News: OpenRouter Auto Code — drop-in coding-agent router (see §3). Pair with Claude Code / Codex / OpenCode session traffic.
- News Synaps CLI early harness (Show HN; 56 stars via API).
- Claim (author, unverified): Rust terminal-native agent runtime / Claude Code alternative; daemon attach 7ms; MCP/skills/subagents; local or any API.
- Caveat: treat perf numbers as author-claimed until independently verified.
- Still open Conductor joins Meta deadline — Conductor Cloud goes read-only on 11-13 and shuts down on 11-20. Local app continues until early 2027. Worktree alternative: worktrunk.
- Context free Crouton Alpha is already eating Claude Code sub-task traffic (40.7B tokens) — useful as a cheap router target while you evaluate Auto Code.
Usable tools and open projects#
Star counts are from GitHub trending (daily) and the GitHub API, read in the discovery window (~07:40 CST).
| Repo or tool | What it is | Stars / +24h | Runnable today? | Why a tinkerer should care |
|---|---|---|---|---|
| ChrystianSchutz/whoami-guard | Outbound hook to strip identity leaks from agent tool traffic | n/a | Runnable (News) | Pair with Claude Code until Anthropic’s next release lands |
| openai/codex CLI 0.167.0 | Agents command center + Vertex wizard | n/a | Runnable (Update) | npm install -g @openai/codex@0.167.0 |
| HaseebKhalid1507/SynapsCLI | Early Rust agent daemon harness | 56 / n/a | Watch (News; author claims unverified) | Interesting post–Claude privacy direction; early |
| openrouter/auto-code | Coding-agent session router | n/a | Runnable (News) | Pass-rate / time / $/solved-task routing |
| stealth/crouton-alpha | Free 1M-context stealth model | n/a | Runnable (News; caution on sensitive data) | Cheap Claude Code sub-task target (40.7B tokens already) |
| NVIDIA/auto-ontology | Semantic graph layer for text-to-SQL / agent workflows | 8,441 / +2,005 | Watch (Update; Apache-2.0; no releases called out) | Still climbing; needs Postgres + NIM |
| cloudflare/security-audit-skill | Multi-phase security audit skill for coding agents | 62,985 / +2,524 | Runnable (skill install; Update momentum) | Installable agent skill, not a runtime |
| google/skills | Agent skills for Google products | 22,657 / +532 | Runnable (skill install; Update) | Installable for coding agents |
| mattpocock/skills | Skills for Real Engineers | 316,229 / +2,058 | Runnable (skill install; Update) | Installable skill pack; no separate runtime |
| Linum P-JiT | ~$21K Apache text-to-image weights | n/a | Runnable (News; research limits) | Cheap open image stepping-stone |
| NVIDIA Cosmos 3 Nano interactive | Four 16B open world-sim checkpoints | n/a | Runnable (News; needs NVIDIA GPU + Framework) | Physical-AI / robotics builders |
| Gemini Omni 1.2 Flash | 15s video gen/edit API | n/a | Runnable (News; GA) | $0.07/s 720p |
Agent builder stack (protocols, runtimes, evals, context engineering, sandboxes, computer use)#
- News Claude Code UA email issue harness hygiene — treat outbound HTTP headers from agent tools as untrusted for PII. Default privacy plugin + whoami-guard; expect Anthropic to drop account email from context in the next release.
- News OpenRouter Auto Code routing — session-aware model switching by pass rate, time, and cost per solved task. Natural fit for multi-model coding harnesses.
- News AA VIALS-AA evals — scientific image reasoning with tools. Large cost/tool-call spreads ($0.02–$1.77/task); Gemini 4 Carbon (high) leads at 64% pass^3. Useful when picking a vision+tools stack, not just chat quality.
- News mattpocock/skills skills momentum (+2,058), cloudflare/security-audit-skill (+2,524), google/skills (+532), NVIDIA/auto-ontology (+2,005) — installable coding-agent skills / ontology layer, not prompt-only packs without a repo.
- News Synaps CLI early runtime — Rust daemon attach harness with MCP/skills/subagents. Author perf claims unverified.
- News GPT-6 Cyber gated cyber agents — Daybreak Red access only. Builder takeaway is access gating + official defender-oriented eval numbers, not general agent scaffolding advice from the post.
- News from HF Daily Papers 11-05 and 11-06.
- What Does Multi-Turn RL Really Improve in LLM Agents? (on the 11-05 list).
- Hybrid Computer Use (unified programmatic/semantic/visual interface) — on the 11-05 list; open the date page for the paper card.
- Harness-S (sibling harnesses for agent training) — on the 11-06 list.
- EvoMemState (personal agent memory) — on the 11-06 list.
- Context Rivet S3-backed Actors, AWS Bedrock AgentCore Web Answer MCP, Docker Sandboxes guidance, Vercel Schedules beta, Baseten stateful microVM sandboxes — flagged by Agentic Digest; open official docs before adopting.
Infrastructure dash#
- News Cosmos 3 Nano interactive sims open world models — four 16B checkpoints; streaming causal distillation; agent skills under
.agents/skills; backends via vLLM-Omni / SGLang / Framework. Vendor NIM note: ~2–3× speculative-decoding on H200. - News VIALS-AA evals infra — new AA scientific image+tools eval with explicit $/task spreads.
- Update OpenRouter rankings serving demand through Nov 5 — DeepSeek Flash family still dominates weekly tokens; Hy4 and Crouton Alpha are the new/trending movers; top apps remain Claude Code, pi, omp, Kilo Code, OpenClaw, DeepSeek Harness, Strix.
- News Linum P-JiT ~$21K cheap open image train — useful cost baseline for small-lab DiT training ($2.50/H100-h assumed in their write-up).
- Engine blogs: not nominated (HN inference slice returned 0 hits; Cosmos mentions engines only as backends).
Builder actions#
- Patch Claude Code privacy now — read issue 107757, upgrade past the privacy-plugin baseline, and install whoami-guard (or equivalent) on outbound tool traffic.
- Try Inkling 2 Small on Tinker at $0.30/$1.00 and watch for Apache weights next week (AA write-up).
- Point a coding harness at OpenRouter Auto Code and compare pass rate / time / cost against a fixed-model baseline.
- Use free Crouton Alpha for non-sensitive sub-tasks; keep sensitive work off the anonymous provider.
- Kick Hy4 on agent/coding traffic at the discounted $0.417/$1.251 and compare Terminal-Bench-style tasks against your current open-weight pick.
- Upgrade to Codex CLI 0.167.0 (
npm install -g @openai/codex@0.167.0) and trycodex agents --no-daemonplus the Vertex wizard if you use GCP. - Laptop / local: pull Nanbeige4.3-33B if you want a 33B Apache reasoner with strong Terminal-Bench relative to size (no public API).
- Install one high-signal skill: cloudflare/security-audit-skill or google/skills; watch NVIDIA/auto-ontology if you do text-to-SQL agents.
- Still open — Conductor Cloud: export before 11-13 (announcement); shutdown 11-20. Local parallel worktrees: worktrunk.
- Still open — default cheap volume model: keep DeepSeek V4.1 Flash 1017 as the OR high-volume baseline (56.5T week) unless a session router beats it on $/solved task.
- If you have Daybreak Red access: read GPT-6 Cyber for access scope and official defender eval numbers only — do not use the post as an exploit cookbook.
- Physical-AI builders: try the Cosmos 3 Nano interactive checkpoints on Framework / vLLM-Omni / SGLang.
Checkmarks stay in your browser only.
Source log#
Show the full source logEvery source checked for this issue.
Primary discovery
| Source | Result | Notes |
|---|---|---|
| Artificial Analysis + articles | Hit | Homepage + Intelligence Index; VIALS-AA (Nov 6), Inkling 2 Small (Nov 5), Nanbeige4.3-33B (Nov 5) |
| OpenRouter rankings | Hit | Usage through Nov 5; weekly/today/new&trending; Hy4, Crouton, Auto Code model pages opened |
| Hacker News front_page API | Hit | Claude Code UA (1305), Linum (255), Inkling (235), Synaps Show HN |
| HN AI/LLM/agent last-24h (Algolia) | Partial | search_by_date numeric filter returned very few hits (4); front_page + manual AI picks used |
| HN inference slice | Miss | 0 hits for vllm/sglang/trt-llm/mooncake/"kv cache"/inference with same 24h filter |
| GitHub trending daily | Hit | mattpocock/skills, cloudflare/security-audit-skill, google/skills, NVIDIA/auto-ontology |
| HF papers 2026-11-05 + 11-06 | Hit | Multi-turn RL agents, Hybrid Computer Use, Harness-S, EvoMemState |
| HF models trending | Hit | Kimi K3.1 Pro, Ripple-65B, DeepSeek Flash, Qwen3.8-27B |
| pricepertoken.com/news (+ agents, coding, open-source, model-releases) | Hit | Feed updated ~1h before fetch; nominated Cosmos, Omni, Auto Code, Crouton |
| news.learnprompt.pro | Partial | SPA shell empty (“加载中…”) |
| learnprompt.github.io/ai-news-radar/ | Partial | Same SPA empty |
| LearnPrompt/ai-news-radar data JSON | Hit | data/latest-24h.json, data/daily-brief.json via raw.githubusercontent.com (generated_at 2026-11-06T12:57:11Z UTC) — Codex 0.167.0, Omni, Cosmos |
| radarai.top/en | Hit | Inkling, Hy4, Omni, Cosmos adjacent; Ant Ling-3.2-flash not primary-confirmed |
| The Agentic Digest | Hit | GPT-6 Cyber issue + Claude Code quick hit + digest-only infra noms |
| simonwillison.net 2026/Nov/5 | Hit | Zero AI nominations (bird sighting + Paul Ford quote) |
Dailies in rank order
- Latent Space AINews, “Inkling 2 Small…” (Nov 6): Hit. Confirmed Inkling; Nov 5 “It's Swapping Season” carried as Context.
- AlphaSignal: Hit. Confirmed Inkling. Dropped aggregator-only: Claude bank accounts, MiniMax Music 3, Hermes Agent, Cursor Build Live Apps, METR (no official primary opened).
- code4u.app/digest: Hit (Nov 5). Confirmed GPT-6 Cyber pipeline, Hy4. Weights-“coming soon” for Hy4 kept as digest-only / not on OR page.
- TLDR AI: Hit (Nov 5 issue). Mostly prior-day items (Images 2.6, Conductor, SWE-2 Flash) — deduped / not rehashed.
- The Neuron: Hit. Nov 5 GPT-6 Cyber; Nov 4 Images/Conductor/Kimi — deduped.
- AI Daily Brief: Hit. Nov 5 data-center slowdown / Nov 4 politics — no extra builder items promoted.
Nominated → confirmed or dropped
- Confirmed on primary pages: Inkling 2 Small (lab + AA); Nanbeige4.3-33B (AA); GPT-6 Cyber (OpenAI + Agentic Digest); Hy4 (OpenRouter); Crouton Alpha (OpenRouter); Auto Code (OpenRouter); Codex CLI 0.167.0 (OpenAI changelog); Claude Code UA email (GitHub issue); Cosmos 3 interactive sims (HF blog + NVIDIA blog); Gemini Omni 1.2 Flash (Google blog); Linum P-JiT (field notes); VIALS-AA (AA); DeepSeek Flash usage (OR rankings); trending skills/ontology (trending + API stars).
- Partially confirmed: Synaps CLI (HN + repo; author claims unverified); Hy4 weights timing (digest-only).
- Dropped: AlphaSignal bank-accounts / Music 3 / Hermes / Cursor Live Apps / METR (no primary); RadarAI Ant Ling-3.2-flash (no primary); GPT-6 Cyber API prices from search snippets (not on OpenAI post body).
Confirm opens
| Event | Status |
|---|---|
| Inkling 2 Small | nominated → confirmed (lab + AA) |
| Nanbeige4.3-33B | nominated → confirmed (AA) |
| GPT-6 Cyber | nominated → confirmed (OpenAI + Agentic Digest) |
| Hy4 | nominated → confirmed (OpenRouter) |
| Crouton Alpha | nominated → confirmed (OpenRouter) |
| Auto Code | nominated → confirmed (OpenRouter) |
| Codex CLI 0.167.0 | nominated → confirmed (OpenAI changelog) |
| Claude Code UA email | nominated → confirmed (GitHub issue) |
| Cosmos 3 interactive sims | nominated → confirmed (HF blog + NVIDIA blog) |
| Gemini Omni 1.2 Flash | nominated → confirmed (Google blog) |
| Linum P-JiT | nominated → confirmed (field notes) |
| VIALS-AA | nominated → confirmed (AA) |
| Trending skills/ontology | nominated (trending + API stars; no deep release notes) |
| Synaps CLI | nominated (HN + repo; author claims unverified) |
| Aggregator-only items without primary | nominated, no primary opened / not promoted |
Engine blogs: not nominated