No. 2 ·

The Oct 30 edition

Window
2026-10-29 07:30 → 2026-10-30 07:30 (America/Chicago, CDT). Older items that you can still act on are kept and labeled.
Method
discovery from aggregators, confirmation from official sources, rubric score >=5, no read-time cap
Rule
every topic, story, and mention is a clickable markdown link
  • 190 primary links
  • 9 sections
  • 19 min read
  • Still open3
  • News28
  • Update4
  • Context6

Labels: News = first seen in the window · Update = older story with a new number or artifact · Context = background you need to act · Still open = not new, but still worth trying

Snapshot#

  • News Mistral released open weights for Mistral Large 4 ("Le Chonk"): a 1T-parameter MoE with 49B active parameters, 1M context, under a modified-MIT license. On AA's Coding Agent Index it scores 48.3 with Mistral Vibe at $1.26 per task. That is about the same as opencode + GLM-5.3 (48.9), which costs $5.23 per task. Unsloth GGUFs start at 233 GB (1-bit).
  • News (security): Cursor says a malicious Anthropic.claude-code 2.1.313 was on the Cursor extension marketplace from about 04:27 to 15:35 CDT on 10-29. Anthropic did not build it, and up to 5,000 users were affected. It tried to plant a persistent SSH backdoor. The npm registry goes straight from 2.1.312 to 2.1.314.
  • News Claude Code v2.1.314 adds permissions.blockWritesOutsideWorkingDirectories. It also adds idle compaction for conversations over 200K tokens in VS Code and Desktop.
  • News Codex CLI 0.166.0 adds transcript search and selection, codex queue --resume, and codex doctor --network-only. It also adds GPT-6.1 Terra to the bundled, Vertex AI, and Bedrock catalogs.
  • News opencode v2.2.3 lets you drive Claude Code from OpenCode: run claude auth login, then pick Claude Agent in /connect.
  • News Google is winding down Jules on Dec 31. Sessions go read-only after Nov 11, and an export is available now.
  • News An OpenRouter test found Sonnet 5.5 and Haiku 5.5 passed every coding task that Opus 5.5 did. Median Claude Code bug-fix session cost was $0.098 for Opus, $0.062 for Sonnet and $0.0052 for Haiku. The tasks were small and the test was run by a vendor.
  • Lighter day for new models beyond Mistral: no new OpenRouter model IDs in the window, and the AA articles list hasn't changed since 10-28.
  • Do this today: if you use Cursor, run the svc-telemetry check from Cursor's notice (grep svc-telemetry ~/.ssh/authorized_keys). If it matches, follow Cursor's remediation scripts and rotate credentials. Everyone on Claude Code should update to v2.1.314.

Frontier and open-weight models#

Mistral Large 4 "Le Chonk" (open weights) — News (2026-10-29)#

  • What: Mistral's largest model. It is a natively multimodal MoE with about 1T total parameters and 49B active (52B counting embeddings and output layers), using 128 routed experts with 4 active plus 1 shared. It has 1M context, takes text and images in, and outputs text. Reasoning effort can be set to low, medium or high (high is the default). Mistral recommends temperature 1.0 and top_p 0.95 (HF card). An embedded DSpark speculative-decoding draft head ships with the weights.

  • License / serving:

  • Local: unsloth/Mistral-Large-4-1T-A52B-GGUF lists these sizes:

    • 1-bit: 233 GB (fits a 256 GB RAM box)
    • 2-bit: 302–378 GB
    • Q8: 1.1 TB

    Unsloth says the DSpark draft (--spec-type draft-dspark) gives about 2x decode at batch 1. See the Unsloth guide.

  • Independent numbers (AA):

  • Vendor-reported numbers (Mistral post): Terminal-Bench 4.0 41% (+13 pp vs preview), FrontierCode 2.0 19.7%, BrowseComp 84.5%. Mistral says it is "effectively tied" with GLM-5.3 on the Coding Agent Index while using 47% fewer tokens.

  • Builder take: This is the strongest Western open-weight model on AA's indexes, but Chinese open models still score higher, as the HF card itself says. The practical win is cost per agent task: GLM-5.3-level Coding Agent Index at about a quarter of the per-task cost. Check the MaaS clause before you resell it.

  • Operator chatter: HN thread (1,546 points, 406 comments when fetched).

Baseten joey-1 — News (2026-10-29)#

  • What: A 2B "decision model" post-trained from Qwen3.5-2B. It returns calibrated probabilities for yes/no, choice and score questions in one forward pass and generates no text. It takes text and images, with 16K context.
  • Cost / license:
    • $0.01 per 1M input tokens, $0 output, on Baseten and OpenRouter baseten/joey-1.
    • It runs on OpenRouter's Decisions API, not the chat endpoint.
    • MIT weights on HF.
    • Baseten reports 20 ms P50 latency.
  • Builder take: A cheap gate for agent tool calls, routing and rubric checks. Run it on a labeled sample of your own approve/deny decisions before you trust its calibration.

Granite Switch 4.1.1 — News (2026-10-29)#

  • A patch release on the same base model. It adds an updated Granite Speech ASR adapter (10 languages) and an updated Decision adapter in granitelib-core, and is backward compatible with 4.1. The interesting part is shipping model updates as adapters. Watch, not urgent.

Still open from yesterday#

  • Gemini 4 Carbon: still the closed-model frontier datapoint. It is #1 on AA Intelligence Index v5 at 59. On the Coding Agent Index it scores 59.4 with Antigravity CLI at $7.74 per task. Access is still limited to Fairwind.
  • GPT-6.1 Terra: $1/$5 per 1M tokens. It is now in the Codex CLI catalogs as of 0.166.0 (see section 4).
  • Kimi K3.1 Pro: weights still aren't on Moonshot's HF org. The newest there is Kimi-K3 (2026-06).
  • DeepSeek V4.1 Flash 1017: still the most-liked trending model on HF (3,937 likes). See the Pareto table below.

Pareto frontier#

Coding Agent Index (Artificial Analysis), read about 07:40 CDT. Each row is a harness + model pair; cost is mean USD per task.

Harness + model Index $/task Note
Claude Code + Claude Opus 5.5 (Max) 61.7 $11.42 Top score
Devin Fusion CLI + Opus 5.5 Max / GPT-6 Luna High 61.5 $8.00
GitHub Copilot + Copilot Fusion (max) 59.9 $5.59 Best $/point near the top
Antigravity CLI + Gemini 4 Carbon (High) 59.4 $7.74 Not generally available
Codex + GPT-6.1 Sol (Max) 54.1 $1.50
Opencode + GLM-5.3 (Max) 48.9 $5.23 CN open-weight
Mistral Vibe + Mistral Large 4 (High) 48.3 $1.26 New today, open weights
Codex + GPT-6.1 Terra (Xhigh) 48.2 $0.69
Codex + DeepSeek V4.1 Flash (Max) 47.3 $0.39 CN open-weight
DeepSeek Harness + DeepSeek V4.1 Flash 1017 (Max) 46.3 $0.30 Cheapest in the ~46–48 band
Kimi Code CLI + Kimi K3 (Max) 41.4 $5.17
  • What moved: Mistral Large 4 sits on the open-weight Pareto curve. It matches GLM-5.3's score for about a quarter of GLM's per-task cost, but GPT-6.1 Terra (closed) and DeepSeek V4.1 Flash (open) give roughly the same score for less. It is a good choice if you need EU-trained open weights you can self-host. It isn't the cheapest option in its score band.
  • Index-level price check: the same model costs different amounts depending on which index you read. Mistral Large 4 costs $2.06 per task on AA's Intelligence Index but $1.26 per task on the Coding Agent Index. Carbon is $7.76 on the Intelligence Index and $7.74 on the Coding Agent Index.
  • Session economics:
    • OpenRouter's Claude tier test (published 10-30): all three 5.5 tiers passed 25/25 hidden-test function tasks and 3/3 Claude Code bug-fix runs.
      • Cost per hard task: Opus $0.0739, Sonnet $0.0380, Haiku $0.0027.
      • Median Claude Code session: Opus $0.098, Sonnet $0.062, Haiku $0.0052.
      • These are short tasks, so they can't rule out an Opus edge on long, ambiguous work. That's a vendor caveat, and the post says so itself.
    • Anthropic Usage and Cost Analytics API (beta, 10-29) has new /v1/organizations/analytics/usage and /cost endpoints. You can group by model, workspace, API key, or user/service account, which is useful for measuring your real $/session. The older Usage and Cost API is now legacy but still supported.
    • Routing floor: joey-1 at $0.01 per 1M input tokens is the cheapest yes/no gate listed today.

Coding CLIs, IDEs, harnesses#

  • News (security) — Malicious Claude Code extension in the Cursor marketplace
    • Who was exposed: anyone who installed or updated Anthropic.claude-code 2.1.313 from the Cursor VS Code extension marketplace between about 2:27am and 1:35pm PT on 10-29 (04:27–15:35 CDT).
    • What it did: collected host info and tried to add an attacker SSH key labeled svc-telemetry and set up persistence.
    • What Cursor did: removed the version, shut down and rebuilt the marketplace, rotated keys, and had the attacker server taken offline. Cursor says it found no other tampered extensions.
    • What to do: run Cursor's check and remediation scripts. Cursor warns that uninstalling the extension alone may not be enough, and says to rotate credentials reachable from affected hosts.
    • Other write-ups: Phorion and OpenSourceMalware. Phorion says 2.1.313 was also pulled from npm and the VS Code registry. The npm registry currently has no 2.1.313.
    • HN: thread (258 points). Treat "Anthropic hacked" headlines as unconfirmed; the primary source names only the Cursor marketplace.
  • News — Claude Code v2.1.314 (2026-10-29 20:06 CDT)
    • permissions.blockWritesOutsideWorkingDirectories: Write, Edit and NotebookEdit refuse paths outside the working directories in every permission mode.
    • Gateway upstream groups and Bedrock guardrail_trace.
    • Idle compaction for conversations over 200K tokens in VS Code and the Desktop Code tab.
    • Fixes the effort-cap retry, and fixes claude -p exiting 0 with no output.
  • News — Codex CLI 0.166.0 (2026-10-29 14:47 CDT; changelog)
    • Keyboard-driven transcript selection and search.
    • codex queue --resume.
    • codex doctor --network-only.
    • Clearer safety-block explanations.
    • GPT-6.1 Terra in the bundled, Vertex AI and Bedrock catalogs.
    • Approval and permission settings are now kept when you resume or fork a thread.
  • News — opencode v2.2.3 (2026-10-30 04:55 CDT) and v2.2.2 (10-29 13:43 CDT)
    • Claude Code as a provider: run claude auth login, then choose Claude Agent in /connect.
    • Faster long sessions in the desktop app.
    • experimental_image_generate is now on by default where the provider supports it.
    • A worktree picker.
  • News — Jules turndown
    • You can finish sessions until Nov 11, 2026. After that Jules is read-only until it shuts down on Dec 31.
    • The export covers settings, env vars, scheduled tasks, memories, setup scripts and session logs with patches. Code and PRs stay in GitHub.
    • Treat the export as sensitive, since it can include env-var values.
  • Context — Android Studio Rabbit 2 stable with Bring Your Own Agent (10-28)
    • You can use 40+ coding agents that support the open Agent Client Protocol (ACP), including Claude Agent, Codex, Junie and Antigravity, inside Android Studio.
    • Agents get Android Studio tools: the Emulator, the build system, Logcat and Compose Preview.
    • Set it up under Settings > Tools > AI > Agents.
  • Context — Vercel CI beta (10-28): CI that runs only the tasks affected by a change, aimed at agent-heavy PR traffic. Vercel claims push-to-CI-pass is 2–6x faster on its own repos.
  • Update — JetBrains thinkrail v0.3.0 (10-29 13:26 CDT): the pi-based IDE pinned Bun 1.4.3, compiles its CLI with JSC bytecode, and added a read/unread workspace marker. It has 6,349 stars (+717 today).
  • Update — DeepSeek Harness
    • dsh-v0.2.1-rc.5 (10-29 21:12 CDT) only improves how thinking content is displayed.
    • The 10-28 alpha added an experimental subagent_oar plugin that hands tasks to a local Codex or Claude CLI.
    • AA lists it with V4.1 Flash 1017 at 46.3 / $0.30 per task.
    • Aggregators framing it as "newly open-sourced" are stale: the repo dates to 2026-08.
  • No movement: gemini-cli shipped nightlies only. pi is still on v1.4.0. aider has had no release.

Usable tools and open projects#

Stars come from the GitHub API, and "+24h" from GitHub trending (daily), both read about 07:40 CDT.

Repo What Stars / +24h Runnable / Watch / Skip Why
herdrdev/herdr-app Desktop app for herdr: spaces, tabs and panes for terminals, plus a sidebar showing which agents are Working, Blocked or Done 598 / — Runnable: Apache-2.0; Linux (Vulkan), macOS Apple Silicon, Windows. Read the installer before running it One view of many CLI agents across machines (HN, 249 points)
cisco-ai-defense/project-quarry A detection rule language and engine for agent sessions: messages, tool calls, results, event order, optional model judges — Runnable: Apache-2.0, uv sync --frozen, alpha DSL; PyPI v0.1.0 Snort/YARA-style rules over agent traces; ships a Claude Code/Codex plugin (Cisco blog)
milind-soni/OpenMausBot Open-source bot-with-a-VM agent 7,981 / +899 Watch: Apache-2.0; v0.1.134 shipped 10-29 19:37 CDT Self-hosted computer-use agent you can read end to end
lightpanda-io/browser Headless browser designed for AI and automation 37,900 / +559 Runnable: AGPL-3.0, 1.0.2 (10-23) A lighter Chrome replacement for agent scraping and test loops
pingdotgg/t3code Coding-agent GUI 32,302 / +920 Watch: MIT; the 10-30 nightly adds secret-provider credentials and force-stop for agents Shows UI patterns for managing multiple agents
JetBrains/thinkrail pi-native IDE 6,349 / +717 Watch: Apache-2.0, v0.3.0 JetBrains iterating fast on a pi-based IDE
thedotmack/claude-mem Cross-session agent memory 105,393 / +1,145 Watch: Apache-2.0, v15.6.0 Still trending; measure it against plain RAG (yesterday's Memory Tax paper)
HKUDS/DeepTutor Agentic tutoring system 51,297 / +2,282 Watch: Apache-2.0, v1.7.0 (10-28) A big open multi-agent app to read through
CopilotKit/OpenBot Self-hosted "AI coworkers" (AG-UI) 8,473 / +707 Watch: MIT Covered yesterday; still climbing
macro-inc/macro Workspace app 5,705 / +379 Skip for agent work: AGPL-3.0 Not agent infra
SheepSeb/tinyworldformer Tiny world model that runs games on a Raspberry Pi Pico 2 W 249 / — Runnable (learn-from) A small codebase to learn world models from (HN, 234 points)
pbakaus/impeccable, expo/skills, mattpocock/skills, obra/superpowers, anthropics/skills Skill and prompt packs +2,299 / +233 / +882 / +1,255 / +287 Skip Viral skill/prompt packs, per the rubric

Also on the daily trending list and skipped: a jailbreak collection (not linked) and an astrology repo.

Open weights trending on HF

Agent builder stack#

Runtime safety and monitoring

  • News — Cisco Project Quarry (open-sourced 10-28, alpha)
    • Rules combine symbolic checks (tool-call order, strings, RE2) with optional LLM/classifier signals. Each rule compiles to a finite-state automaton, and findings come back with witness evidence and common secret formats masked.
    • Three uses: runtime guardrails, after-the-fact scans of stored traces, and replaying a newly found pattern over old traces during incident response.
    • The Register coverage.
  • News — OpenAI: How we detect frontier model misalignment (10-29)
    • Architecture: a layered monitor. Activation probes first, then a model reviewer, an investigative agent, and finally telemetry correlation. It covers all internal Codex traffic.
    • Applies to you if you build on the Responses API with the most capable models. External monitoring covers tool-using Responses API requests, and can auto-pause a conversation only when you use supported conversation tracking. Without tracking you get a webhook alert, but the conversation continues. Chat Completions is not covered.
  • News — Claude Code blockWritesOutsideWorkingDirectories: a cheap way to keep agent writes inside the working directory, even in permissive modes.

Platform plumbing

Agent evals (HF Daily Papers)

  • News — Your Benchmark Is an Empirical Measure Over Difficulty (136 upvotes; code): looks at 25 benchmarks for difficulty-distribution mismatch and coverage gaps. Use it before you trust a point delta on your own eval set.
  • News — Co-Design Bench (code): 273 tasks drawn from real coding sessions, graded by judge-then-refute.
  • News — PersonaCUABench (code): 74 tasks and 44 personas. Task completion drops 36% across 9 backbones when user personas vary (3,330 runs).
  • News — WebAnything (code): 244 resettable website environments for web-agent RL and evals.
  • News — WorkFactory (code): 112,686 tasks across 612 occupations, plus WorkPilot Qwen fine-tunes (4B, 27B and 35B-A3B).
  • News — Meta JiTTesting (paper): LLM-generated tests that are designed to fail against risky diffs before they land.
    • Engineers reported 60% of catches as true positives.
    • 74 serious privacy bugs caught at the diff stage in H1 2026.
    • A pattern to copy for reviewing agent PRs.

Infrastructure dash#

  • News — SkyPilot: Grace Blackwell needs a new scheduler (10-29)
    • Measured NVL72 unavailability is about 2%. At that rate, NVIDIA's model gives 16-node slices 27% more usable capacity than 18-node slices (2,037 vs 1,602 nodes).
    • SkyPilot adds topology-aware placement, in-place recovery, and automatic defragmentation that preempts only checkpoint-safe jobs.
    • Its headline claim is +14% NVL72 training throughput. That figure is vendor-reported.
    • Relevant to multi-node inference on NVL72 too.
  • News — Mistral Large 4 NVFP4: an official Blackwell-native FP4 checkpoint with vLLM and SGLang tags. It is the realistic way to serve a 1T model on a single node.
  • Context — Marketplace supply chain: the Cursor incident is a reminder to pin extension versions and watch authorized_keys changes on dev machines.
  • Engine blogs: not nominated (the HN inference slice for vLLM, SGLang, TRT-LLM, Mooncake and "kv cache" returned nothing builder-relevant in the window).

Builder actions#

  1. Check Cursor machines for the backdoor. Run grep svc-telemetry ~/.ssh/authorized_keys (Cursor's notice has a PowerShell variant). If it matches, run Cursor's remediation script and rotate credentials from a clean machine.
  2. Upgrade Claude Code and fence writes. Install v2.1.314 (npm i -g @anthropic-ai/claude-code@2.1.314) and set permissions.blockWritesOutsideWorkingDirectories: true in your settings.
  3. Price-check Mistral Large 4 on your own tasks. Run the same 10 agent tasks through mistralai/mistral-large-4-0, deepseek/deepseek-v4.1-flash-1017 and openai/gpt-6.1-terra. Compare $/task against AA's Coding Agent Index ($1.26 / $0.30 / $0.69).
  4. Try it locally if you have 256 GB+ RAM. Pull the 1-bit Unsloth GGUF and test the DSpark draft flags from the Unsloth guide.
  5. Run a Claude tier bake-off. Use Ori Eval, or set ANTHROPIC_MODEL per run as in OpenRouter's test, to compare Opus, Sonnet and Haiku 5.5 on three of your own repo tasks.
  6. Meter real session cost. Call the new Usage and Cost Analytics API (beta header analytics-usage-cost-2026-08-24) and group by API key or service account.
  7. Write one agent-detection rule. Clone Project Quarry, run uv run --frozen quarry scan --config examples/config.yaml --target examples/data/benign.json, then write a rule for "agent reads a secrets file, then makes an outbound call", and replay it over your stored traces.
  8. Export from Jules before Nov 11. Follow the turndown notice and store the archive as a secret, since it can include env values.
  9. Gate tool calls cheaply. Send 100 labeled approve/deny tool calls through baseten/joey-1 on the Decisions API and pick a probability threshold.
  10. Try opencode + Claude Agent. Upgrade to opencode v2.2.3, run claude auth login, then use /connect → Claude Agent.
  11. Check eval coverage. Run benchmark-difficulty-measure on your internal golden set to see whether it covers hard cases at all.

Checkmarks stay in your browser only.

Source log#

Show the full source logEvery discovery source checked, hits and misses, and every confirm open.

Primary discovery

Source Result Note
Artificial Analysis home Hit Changelog: Mistral Large 4 high/medium evals (10-29), AA-Robotics v0.1
AA articles Partial No new articles since 10-28
AA Coding Agent Index Hit Mistral Vibe + Large 4 at 48.3 / $1.26; full table parsed
OpenRouter rankings Partial HTML fetched; top-apps table not re-parsed
OpenRouter models API Miss No new IDs in the window (newest particle-1, 10-28); joey-1 is on the Decisions API, not this list
HN front page Hit Mistral Large 4, Herdr, Claude Code extension backdoor, tinyworldformer, SkyPilot
HN Algolia last 24h (AI/LLM/agent keywords) Hit The > filter needed URL-encoding (first call 400); re-run as per-keyword queries and merged
HN Algolia inference slice (vllm, sglang, trt-llm, tensorrt-llm, mooncake, "kv cache", inference) Miss Nothing engine-specific in the window
GitHub trending daily Hit Stars-today checked against GitHub API totals
HF Daily Papers 2026-10-30 Hit 30 papers; benchmark difficulty, WorkFactory, Co-Design Bench, WebAnything
HF Daily Papers 2026-10-29 Hit PersonaCUABench, RealCLI (no code, not printed)
HF models trending Hit Via API
Price Per Token news + /agents, /coding, /open-source, /model-releases Hit Noisy; nominated Mistral, OpenAI monitoring, Meta JiTTesting, Deepgram/Daily
LearnPrompt/ai-news-radar data JSON (data/latest-24h.json, master branch) Hit Generated 2026-10-30 06:24 CDT; main branch path returned 404
RadarAI Hit Claude Code extension malware, DeepSeek Harness, AIE CODE
The Agentic Digest Hit 2026-10-29 issue: Android Studio BYOA, Project Quarry, Vercel AI Gateway virtual model IDs
Simon Willison 2026-10-29 Hit AI-pilled company link, API-credits follow-up; low builder action, not printed
Simon Willison 2026-10-30 Miss Day archive 404, no posts yet
X (narrow) Not used Optional

Dailies in rank order

  1. Latent Space AINews: Partial. Titles only (latest: Building AIE CODE); body not read.
  2. AlphaSignal: Partial. Headlines, some Pro-gated. Nominated Claude API 128 MB requests, Mistral, the Cursor extension, and OpenAI's monitoring.
  3. Code4U AI Digest: Partial. Latest date is still 2026-10-28; nothing new.
  4. TLDR AI archives: Hit. Opened the 2026-10-29 issue. Nominated Vercel CI, Copilot Automations, ElevenLabs Cue, and others that were X-only.
  5. TLDR flagship: Not checked. Only used if TLDR AI is blocked, and it wasn't.
  6. The Neuron: Partial. Titles: Mistral 1T open model, Deepgram buying Daily.
  7. AI Daily Brief archive: Partial. Latest 2026-10-29 ("The AI Rankings Just Got Reshuffled"); nothing printed.

Confirm opens: nominated → confirmed

Dropped

  • Codex usage limits doubled for Pro and Business Premium (HN via X): nominated, no primary (only an X post; nothing in the Codex changelog), dropped.
  • "Anthropic hacked" (HN → doublepulsar.com): Blocked (403). The primary source (Cursor) names only the Cursor marketplace, so the framing was dropped.
  • "Claude flags its own misuse at under 2% extra cost" (AlphaSignal): nominated, no primary found, dropped.
  • Vercel AI Gateway configurable virtual model IDs (Agentic Digest): not on Vercel changelog page 1. Nominated, no primary, dropped.
  • Antigravity plugin marketplace (third-party blog alongside Jules coverage): nominated, no primary, dropped.
  • Samaya Sterling, Decagon inference stack, Mirendil Lor (MarkTechPost / X-only): nominated, no primary, dropped.
  • Deepgram acquires Daily/Pipecat (PPT, Neuron): only TechCrunch opened; no company primary, dropped.
  • Epoch open-weight patch-gap report, Martin Fowler "LLM Voice", Claude CarPlay, Claude Duet, Kimi Gmail, Accenture DOD agent: scored below 5 (commentary, consumer, or policy).
  • Yesterday's items (Gemini 4 Carbon, Copilot Fusion, gh-aw, Particle-1, LFM3, Cloudflare SGLang) repeated in several dailies. They appear above only as Still open or as Pareto rows.

Engine blogs: not nominated.