Issue No. 7

The Nov 4 edition

  • 21-minute read
  • 182 primary links
  • 9 sections
  • News23
  • Update5
  • Context7
  • Still open7

Snapshot#

A fairly quiet shipping day. One new model from a major lab, some re-scored leaderboards, and several agent-auth and agent-security items.

  • News (Chinese model, weights promised): Tencent Hy4 launched on 11-04. Tencent calls it one of the strongest open-weight models, but the weights and system card are only promised "in the coming days" (Tencent press release). You can use it today on OpenRouter as tencent/hy4 with a 1M context window at $0.417 in / $1.2505 out per 1M tokens. That is exactly half of Tencent's own list price of $0.834 / $2.501. No independent index has scored it yet.
  • Update (leaderboard): AA Intelligence Index v5.0.1 regraded every model on CritPt v2 using the CritPt authors' grading fixes. Most scores moved up by 0.1–0.7 points, and the order did not change: Gemini 4 Carbon (High) is still #1 at 60.0. AA also moved its six Capability Indexes to v2.0, giving more weight to non-hallucination and agentic tool use (AA on X).
  • News (agent security): Postman Desktop auto-installed Claude Code and Codex plugins, plus a startup hook, a remote MCP connection and a CLI. Its installer also tried to mark its own Codex hook as trusted. The issue was #1 on HN with 1,368 points. If you have Postman installed, audit your agent configs.
  • News (MCP auth): Claude can now act as the identity issuer for enterprise-managed MCP auth (Claude Help Center, beta, Team/Enterprise). Supabase and WorkOS MCP Auth already accept it. The catch: disabling a user in your identity provider no longer revokes their access.
  • News (coding CLI): Claude Code v2.1.318 adds per-fork effort for mods and a sign-in expiry field in claude -p JSON output. It also fixes a --continue regression and MCP config writes that skipped the lock. GitHub Copilot now runs inside JetBrains Air over ACP.
  • News (local hardware): on RTX Spark laptops running native Linux, GPU clocks drop to about 780 MHz and GPT-OSS-120B falls from 65.4 to 13.9 tok/s (MindStudio write-up, HN). WSL avoids it. This affects the RTX Spark laptop, not the DGX Spark.
  • Still open: GitHub's first pull_request_target brownout is Monday 11-09, 10:00–12:00 CST. Kimi K3.1 Pro and LongCat-2.5 still have no weights on Hugging Face.

Frontier and open-weight models#

News#

  • Tencent Hy4 (11-04; press release; transparency center lists "V1.0 Hy4, November 04, 2026").
    • New capabilities: vision input, plus computer use in both CLIs and GUIs. It is Tencent's first model with either.
    • Benchmarks: Tencent reports big gains over Hy4 preview, but only on internal or partner benchmarks: +19.1% GameDev, +9.8% GameLogic, and an internal search hallucination rate down from 15.76% to 5.45%. On a 579-task finance study it ranks first among open-weight models (Hy4 3.04, DeepSeek-V4.1-Flash-1017 2.97, Kimi K3.1 Pro 2.68, GLM-5.3 2.68). None of this is independently verified.
    • Gaps: Tencent says it still trails top closed models on research math, long-horizon coding and advanced visual QA.
    • Weights: promised but not released. tencent/Hy4 on HF is not public yet. Tencent says it is working with vLLM (Inferact) and SGLang so both engines support the model at release. No parameter count or license has been published.
    • Safety notes: Tencent says it limited cyber training data to defensive material and filtered out exploit development and reverse engineering. It reports a biosecurity score of 30.4/100 on its own Hy-BioExperimentBench.
    • Access today:
    • Price: three different figures:
      • Tencent's research page: $0.834 in / $2.501 out / $0.042 cached per 1M.
      • Press release: "approximately" $0.90 / $2.70 / $0.045.
      • OpenRouter: $0.417 / $1.2505 / $0.021, exactly half the research-page price.
      • Hy4 preview on OpenRouter is still at $0.834 / $2.501. Treat the OpenRouter half price as a possible promo.
    • Usage signal: Hy4 preview was already #4 on OpenRouter's weekly ranking at 9.48T tokens (+6%).
  • Crouton Alpha (stealth model on OpenRouter, created 11-02 19:38 CST).
    • Free, 1M context. Accepts text, image, audio and video. Reasoning is always on.
    • It is #2 in OpenRouter's new-and-trending list at 25.1B tokens.
    • The vendor is unknown. Don't send it private data. Watch it as a free test target.

Update / Context#

  • Axion 1.0 (ByteDance with Duke, UMD and Buffalo; code, Apache-2.0; 38 votes on HF Papers 11-04).
    • Agent models fine-tuned from Qwen bases with two SFT stages plus RL. The data and the full training recipe are released.
    • Axion 1.0-122B (122B total / 10B active, from Qwen3.5-122B-A10B) reports 59.0 on full HLE and 87.0 on FrontierScience Olympiad. These are the authors' numbers.
    • Official checkpoints are not released. The only weights are a community 35B reproduction, which the README says is not an official checkpoint.
  • Mage-VL-Cowork (Microsoft, project page). A 4B hybrid (GUI plus code) computer-use agent with a full training recipe. It is #28 on HF trending.
  • HF trending: no new frontier-class weights. The closest thing is a new community quant, ISTA-DASLab's DeepSeek-V4.1-Flash-1017 GSQ coder GGUF (created 11-02). DeepSeek V4.1 Flash 1017 and Mistral Large 4 remain #3 and #4.

Still open#

  • Kimi K3.1 Pro: weights are still not on HF; the newest Moonshot model there is still Kimi-K3. AA's new capability indexes still rank it the top open-weights model in all six domains (AA on X).
  • LongCat-2.5: still API-only (Vercel AI Gateway, $0.30 / $1.20). The newest model in Meituan's HF org is still from July.
  • Gemini 4 Carbon (Low) at $4 / $20 is still the cheapest way to use the AA #1 model family (covered yesterday).

Pareto frontier#

AA Intelligence Index v5.0.1 (Update; AA changelog "Methodology updated · 3 Nov").

  • What changed: "Incorporated grading robustness improvements from the CritPt authors and regraded all models on CritPt v2." Nothing else. The nine evals and their weights are the same as in v5.0.
  • Effect: scores move by up to +0.7 points, and the rank order among these models stays the same.
  • Table: compares the AA home payload from yesterday and today. The $/task column is AA's cost per task, which did not change.
Model v5.0 (11-03) v5.0.1 (11-04) $/task (AA) Weights
Gemini 4 Carbon (High) 59.43 60.03 $7.76 Closed
Claude Opus 5.5 (Max) 58.41 59.11 $10.04 Closed
GPT-6 Astra (Max) 56.85 56.85 $6.24 Closed
GPT-6.1 Sol (Max) 54.51 54.61 $1.27 Closed
Claude Sonnet 5.5 54.24 54.54 $10.02 Closed
GPT-6.1 Terra (Max) 50.98 51.33 $0.96 Closed
Kimi K3.1 Pro 46.00 46.45 $3.68 Weights not released
DeepSeek V4.1 Flash 1017 (Max) 32.49 32.59 $0.60 Open
GLM-5.3 (Max) 31.65 31.65 $5.01 Open
MiMo-V2.6-Pro 29.89 29.89 $0.26 Open
MiniMax-M3.1-Flash 28.91 29.06 $0.49 Open
GPT-6 Luna 27.76 27.86 $0.15 Closed
Mistral Large 4 (High) 26.02 26.12 $2.06 Open
  • How to read it: GPT-6.1 Sol at $1.27/task is still the knee of the closed curve: 1.4 points behind Sonnet 5.5 at about an eighth of the cost. Among open weights, MiMo-V2.6-Pro ($0.26) and MiniMax-M3.1-Flash ($0.49) are the cheap points. Hy4 has no AA row yet.
  • Leaderboards disagree: OpenRouter's benchmarks view shows "Qwen3.8 Max" at 53.4, linked to the 0803 snapshot. AA's home page lists Qwen3.8 Max (0902) at 33.65. These are different snapshots, so don't mix the two numbers.

AA Capability Indexes v2.0 (News, announced 11-03 18:47 CST).

  • What they are: six job-weighted indexes: Finance & Accounting, Strategy & Ops, Legal, Healthcare, Engineering and Economics.
  • Changes:
    • More weight on non-hallucination (AA-Omniscience v1.1) and agentic tool use (AutomationBench-AA).
    • GDPval-AA and AA-LCR are removed, and HLE is replaced by HLE-Diamond.
    • Engineering adds a Science component (CritPt v2 Hard plus Terminal-Bench-Science) and FrontierCode 2.0.
  • Results (per AA's post, as relayed by AGI Hunt): Opus 5.5 (Max) leads all six, and Kimi K3.1 Pro is the best open-weights model in each.
  • Builder take: if you pick models by domain, the Finance and Legal weights now punish hallucination harder. Check the per-index page rather than the headline index.

AA Coding Agent Index v2.0 (Still open). Today's payload is identical to yesterday's, with no new rows. Key points ($/task = mean USD per task):

Harness + model Index $/task
Claude Code + Opus 5.5 (Max) 61.7 $11.42
GitHub Copilot + Copilot Fusion (max) 59.9 $5.59
Codex + GPT-6.1 Sol (Max) 54.1 $1.50
Opencode + GLM-5.3 (Max) 48.9 $5.23
Mistral Vibe + Mistral Large 4 (High) 48.3 $1.26
DeepSeek Harness + V4.1 Flash 1017 (Max) 46.3 $0.30
Codex + GPT-6 Luna (Max) 31.1 $0.16

Harness economics (News): two write-ups this window show harness design moving cost as much as model choice.

  • IBM's same-model comparison. With GLM-5.3-flash and the same 90 AppWorld tasks:
    • CUGA completed 81.1% (73/90), Hermes 46.7% and Deep Agents 30.0%.
    • Counting only scenarios where all three variants passed: CUGA 66.7%, Hermes 26.7%, Deep Agents 6.7%.
    • IBM ran the test and tuned CUGA's AppWorld configuration itself, so read these as vendor numbers.
  • Framer Harness 2 (11-02). The agent writes JavaScript, and Framer runs it while the model is still streaming the code.
    • Framer reports up to 33% lower cost and 20% faster runs.
    • Half as many tool steps, and 39% fewer cached input tokens per message, with no measured quality loss.

Usage signal (OpenRouter rankings, data through 11-03).

  • Weekly tokens: DeepSeek V4.1 Flash 1017 leads at 54.8T (+47%), followed by GLM 5.3 Flash at 16.4T and Hy4 preview at 9.48T. MiniMax M3.1 Flash rose 181% to 6.5T.
  • Market share: DeepSeek held 29.8% in the week of 10-26.
  • Apps: Hermes Agent 1.89T, Claude Code 1.68T, Kilo Code 727B, pi 613B, DeepSeek Harness 368B and Codex 204B.

Coding CLIs, IDEs, harnesses#

  • News: Claude Code v2.1.318 (11-03 18:01 CST).
    • Cost control: a mod's $.model.fork now takes an effort option, so side jobs can run at low effort whatever the session uses.
    • Sign-in expiry: claude auth status now shows loginExpiresAt, and claude -p JSON includes credential_expires_at. Text output prints a stderr warning starting 10 days before expiry. Useful for headless bots.
    • Gateway: assume_role on mantle upstreams for cross-account AWS and per-developer cost attribution.
    • Fixes: --continue no longer loses history after a malformed tool call (a regression from 2.1.314). claude mcp add/remove now take the config lock. Idle compaction no longer replaces a conversation with a non-summary. Headless -p no longer reports success when an API error cut off a reply. Empty or bad NODE_EXTRA_CA_CERTS no longer breaks every connection.
  • News: GitHub Copilot in JetBrains Air (11-03).
    • Copilot appears in Air's agent picker over the Agent Client Protocol (ACP), runs parallel sessions with in-IDE review, and can run cloud tasks through Air Teams.
    • Cloud tasks bill against your Copilot plan, not JetBrains AI credits.
  • News: Copilot usage API reports BYOK (11-03).
    • New fields used_local_byok and totals_by_local_byok separate requests that go straight from Copilot CLI or the app to your own key or local model.
    • Enterprise custom models are reported too.
  • News: llm 0.37.1 (11-03 15:41 CST). A fix release:
    • llm openai endpoint add no longer writes invalid YAML when the endpoint list is empty.
    • The system prompt is no longer dropped when you pass both system= and messages=.
    • --no-stream no longer crashes when an OpenAI-compatible server omits usage.
  • News (demo): Simon Willison: Qwen 3.8 27B on a laptop can interpret model weights, too (HN, 664 points).
    • Using yesterday's weight-space prompts, the local 17 GB Qwen 3.8 27B correctly named Pac-Man, Pong and Breakout from raw policy weights with no tools. Each answer took 46–65 minutes and up to about 43.5K reasoning tokens.
    • Via OpenRouter, Kimi K3.1 Pro named Breakout in 9m 41s.
    • Reproduce with llm -f <prompt URL> -m lmstudio/qwen/qwen3.8-27b. Raise LMSTUDIO_TIMEOUT and the context window first.
  • Pre-releases only elsewhere:
  • News (industry, Codex): OpenAI to acquire Pierre Computer (Axios exclusive, OpenAI confirmed to Axios; terms not disclosed).
    • OpenAI plans to fold Pierre's Git-based Code Storage, which Codex already uses, into Codex.
    • Pierre's open-source projects will stay open source.
    • Builder take: if you build on Code Storage, watch for API changes. Otherwise this is a Codex-vs-GitHub positioning signal.

Usable tools and open projects#

Repo or tool What it is Stars / +24h Runnable today? Why a tinkerer should care
cuga-project/cuga-agent IBM's open agent harness for APIs, web and MCP, with harness-enforced policies 1,069 / n/a Runnable (Apache-2.0, any OpenAI-compatible endpoint) IBM reports 81% vs 47% for Hermes and 30% for Deep Agents on AppWorld with the same open model (blog)
microsoft/apm Agent Package Manager: versioned skills, prompts and instructions via apm.yml 4,203 / n/a Runnable (MIT) Dependabot now opens version-update PRs for it (11-03)
llama.app (macOS, Windows) Desktop/tray app from the llama.cpp team and HF; Linux gets a CLI install 1,730 (macOS repo) / n/a Runnable (Context: macOS 0.40.0 on 10-28, Windows v0.12.0 on 10-30, MIT) Curated local models list (Qwen 3.8 27B, GLM 5.3, LFM 3 Omni) with an OpenAI-compatible local server. Resurfaced in an HN post
floci-io/floci Free local AWS emulator 29,275 / +1,082 Runnable (2.4.0, 10-31) A safe AWS target for agent infra tests without real credentials
D4Vinci/Scrapling Web-scraping framework 95,845 / +2,006 Runnable (v0.5, BSD-3) Fetch layer for research agents; trending again
NVIDIA/auto-ontology Semantic/ontology layer served over MCP for text-to-SQL agents 3,854 / +2,053 Watch (Update: was 1,855 yesterday; no releases) Stars doubled in a day; needs Postgres plus NVIDIA NIM
davide-troiani/moeshift Memory-efficient MoE inference for Apple Silicon n/a Watch (first release 11-02; Show HN, 2 points) Another expert-offload engine to compare against colibri
FareedKhan-dev/kimi-k3-in-c Pure-C99 Kimi K3 (2.78T) inference that streams the 1.56 TB checkpoint from disk 9,998 / +351 Watch (Apache-2.0; last push 10-02) The author measured 26.5 s/token on 8 GB RAM and 5.6 s/token with 128 GB+. Good for learning, not for real use
Muvon/octomind Single-binary, MCP-native coding agent runtime 429 / n/a Watch (Apache-2.0) Small; another any-model terminal/CI agent
max-sixty/worktrunk Git worktree CLI for parallel agents 10,812 / +524 Runnable (Still open, v0.81.0) Still the easiest way to give parallel agents separate worktrees
vercel-labs/agent-desktop Computer-use CLI using OS accessibility APIs 14,147 / +2,463 Runnable (Still open, v1.1.0) Covered yesterday; still climbing
thedotmack/claude-mem Cross-session memory for coding agents 112,472 / +1,165 Runnable (Still open, v15.9.0) No new release since yesterday
mattpocock/skills, garrytan/gstack Skill/prompt packs +1,376 / +285 Skip Prompt packs with no runtime

Totals for cuga-agent, apm, Scrapling, kimi-k3-in-c, octomind and Llama-macOS come from the GitHub API at about 07:45 CST. The other totals, and all +24h figures, come from GitHub trending at about 07:35 CST.

Agent builder stack#

  • News (supply chain): Postman Desktop auto-installed agent plugins (issue filed 11-02; HN #1, 1,368 points).
    • What the reporter found after opening Postman 12.33.4:
      • a user-wide Codex plugin with a SessionStart hook that tells the agent to load Postman skills;
      • a remote MCP connection;
      • a Claude Code plugin (postman@claude-plugins-official, user scope);
      • a Postman CLI in ~/.local/bin.
    • Postman's own log shows "source":"auto-existing-upgrade". The installer source calls a function that writes a trusted hash for its own hook.
    • No Postman response appears on the issue yet.
    • Builder take: any desktop app can now write to your agent configs. Diff ~/.codex and ~/.claude (settings, installed plugins, hooks) after installing or updating tools.
  • News (MCP auth): Claude-issued enterprise-managed auth (beta; Team/Enterprise).
    • Claude signs the ID-JAG identity assertion itself, so connectors auto-connect without Okta or another identity provider that supports the extension.
    • The Claude-issued beta works with Amplitude, Apollo.io, Atlassian, Datadog, Dropbox, Figma, Granola, HubSpot, Linear, Mixpanel, Notion, PagerDuty, Semrush and Supabase.
    • Supabase (GA on its side, 11-03) spells out the tradeoffs:
      • Claude checks only Claude-org membership and email. It does not check SSO, MFA or Supabase's own sign-in rules.
      • Disabling a user in your identity provider no longer revokes access; remove them from both orgs.
      • Supabase trusts one issuer per SSO provider, so switching to Claude breaks assertions from your identity provider for other AI tools.
    • WorkOS MCP Auth added support the same day (early access).
    • Builder take: if you run an MCP server, supporting the Enterprise-Managed Authorization extension is now the path to zero-click installs.
  • News (agent security, GitHub): GitHub Next Enclaves (11-02, surfaced by TLDR AI 11-03).
    • An experimental primitive in GitHub Agentic Workflows for cross-repo work.
    • The main agent launches a single-use sandbox in another repo with a response schema. A broker caps how many bits can come back, with one bit reserved for errors.
    • Private files never enter the main agent's context. This addresses the private-to-public leak class ("GitLost") behind the Noma Labs research the post cites.
  • News (platform): Siri AI open to all developers (11-03, Developer Preview in iOS/iPadOS/macOS/visionOS 27.3 betas).
    • Any app can expose content and actions to Siri AI through App Intents.
    • New "skills" let you combine several actions into one experience.
    • Siri AI can now surface any entity in the Spotlight semantic index, not just ones that fit an existing schema.
  • News (harness design): Framer Harness 2 (11-02).
    • The model writes JavaScript, which Framer compiles into cells and runs while the model is still streaming.
    • If an assumption fails mid-stream, Framer stops generation to save output tokens and returns the error so the model can retry.
    • It is a concrete pattern for any agent that makes large numbers of repetitive edits.
  • News (agents in CI): Dependabot supports Agent Package Manager (11-03).
    • Dependabot opens version-update PRs for git dependencies in apm.yml and regenerates apm.lock.yaml.
    • It covers version updates only, not security updates.
  • News (papers with code) from HF Daily Papers 11-04 and 11-03:
    • Caper, "Shall We Approve?" (code, Apache-2.0). Runs shell commands in a throwaway sandbox first, so an approver sees the actual effects instead of judging the command text. Relevant to anyone building auto-approve modes.
    • LEANGUARD (code, MIT). Blocks tool calls that violate a policy proven in Lean and can repair them. The authors report fewer unauthorized executions and lower prompt-injection success with no loss on benign tasks.
    • ToolSpace (code, 31 votes). Builds replayable multi-environment tool-use tasks from mined execution traces.
    • Switch-SWE (code). Coding agents lose ground when they switch between unfinished tasks. A context-management module helps them recover. Relevant to parallel-session workflows.
    • Route Anything (code, MIT, 87 votes, top paper). A study of routing tasks across models and policies.
    • WebHarbor (code). 153 dockerized real-website environments for GUI-agent evals and RL.
    • ParetoSaddler (Microsoft). Optimizes a harness for quality and cost at the same time. No code, so Context only.
  • Context (eval): StateBench (Ai2; code). Frontier models including Opus 5.5 and GPT-6 Sol still fail harder state-tracking chains in long inputs. Recurrent and hybrid models generalize further when trained on the task. Worth reading if your agents track variables across long logs.
  • Context (still free until 11-10): Vercel Security Review (public beta since 11-02, Pro/Enterprise).
    • Scans each push's diff for issues like missing authorization checks, then posts findings to PRs, the dashboard and the CLI.
    • Reviews are free until 2026-11-10. Fixes made with Vercel Agent bill at provider token rates plus $0.25 per 1M tokens.
  • Still open (CI): pull_request_target blocked by default from 2026-12-02. Brownouts are 11-09 and 11-16, 10:00–12:00 CST.
  • Still open (MCP SDK OAuth): upgrade to mcp ≥1.30.0 / ≥2.2.0 and @modelcontextprotocol/sdk ≥1.31.0 (yesterday's brief).

Infrastructure dash#

  • News (local hardware): Linux on the RTX Spark laptop (MindStudio, 11-01; HN, 221 points).
    • On wall power under native Linux, the GPU locks to about 780 MHz, versus about 2.2 GHz on Windows.
    • The author blames Windows-only NVIDIA Platform Controllers and Framework software, which can decrypt the firmware's power table.
    • Author-measured GPT-OSS-120B: Windows 65.4 tok/s, WSL 65.1, DGX Spark 61, native Linux 13.9.
    • The 96 GB GPU memory reservation can only be changed from Windows.
    • HN commenters note the headline confuses this laptop with the DGX Spark, which is unaffected.
  • Context (engine tuning, nominated via LearnPrompt/MarkTechPost): NVIDIA: 3 ways to boost MoE inference on DGX Spark (10-27). Three SGLang settings for Ling-3.0-flash (124B total / 5.1B active):
    • swap the FlashInfer MXFP4 MoE backend for the Humming NVRTC kernels;
    • set SGLANG_ENABLE_FP8_LM_HEAD=1;
    • turn on MTP speculative decoding with the model's NEXTN layer.
    • Combined result on full GSM8K: +42.5% decode throughput and −24.5% mean latency, with accuracy flat (96.89% vs 96.82%).
  • Update (measurement): AA Performance Benchmarking v3.0.0 (dated 10-30; it appears in AA's 11-03 changelog). Speed numbers on AA are now measured differently:
    • Workloads now include multi-turn agentic runs with tool calls from AA's coding and knowledge-work evals.
    • Headline speed is end-to-end (including time to first token) and averaged over 5 days.
    • Output tokens are counted as each provider's API reports them.
    • Endpoint errors count as 0 tok/s, and runs happen every 6 hours.
    • Don't compare AA speed charts from before 10-30 with current ones.
  • News (papers): SpecPilot (code). An agent that searches for speculative-decoding algorithms and serving settings tuned to your model, drafter and serving setup. Also NVIDIA's "You don't need long context" (project), which argues that frequent compaction preserves accuracy on long-horizon tasks.
  • Hy4 serving: Tencent says it is working with vLLM and SGLang so both support the model when weights ship (Tencent). Watch both engines' release notes.
  • Engine blogs: vLLM, SGLang and TRT-LLM blogs were not nominated today. The NVIDIA post above is the only engine-tuning item.

Builder actions#

  1. If Postman Desktop is on any machine you use with agents, list what it added. Check ~/.claude/plugins/installed_plugins.json, ~/.claude/settings.json, your Codex plugin and hook config, and ~/.local/bin/postman. Remove what you didn't opt into (issue #14370).
  2. Try Hy4 on one real agent task through OpenRouter tencent/hy4 while it's at $0.417 / $1.25. Compare it with DeepSeek V4.1 Flash 1017 on the same task. Don't plan around open weights until they actually appear on HF.
  3. Upgrade Claude Code to v2.1.318. In headless bots, read credential_expires_at from claude -p --output-format json and alert before it expires.
  4. If you run an MCP server for teams, read Supabase's write-up and the Claude Help Center page, then decide whether to accept Claude as an ID-JAG issuer. Document the two-org offboarding rule.
  5. Re-run your model shortlist against AA v5.0.1 and the relevant Capability Index. The order didn't change, but cached screenshots of v5.0 scores are now stale.
  6. If you keep agent skills in a repo, try APM plus a Dependabot apm entry (changelog), so skill updates arrive as reviewable PRs.
  7. Before the 11-09 10:00 CST brownout, search public repos' .github/workflows for pull_request_target (GitHub).
  8. Use Vercel Security Review on a Pro team before 11-10 while it's free (Vercel).
  9. Reproduce Simon's weight-reading test on your local model with llm -f. Set a long context and timeout, and budget about an hour per prompt.

Checkmarks stay in your browser only.

Source log#

Show the full source logEvery source checked for this issue.

Primary discovery

Source Result Note
Artificial Analysis + articles Hit Newest article still 10-28. Changelog "Methodology updated · 3 Nov": Intelligence v5.0.1 and Performance v3.0.0. Coding Agent Index unchanged. No Hy4 row
OpenRouter models API + rankings Hit New IDs: tencent/hy4, stealth/crouton-alpha. Inkling :free variants removed. Several list-price changes (e.g. gpt-oss-120b, llama-3.3-70b) not printed
HF trending models Hit No new frontier weights; ISTA-DASLab DeepSeek GGUF quant; Mage-VL-Cowork #28
HF Daily Papers 11-04 + 11-03 Hit Route Anything, Axion 1.0, ToolSpace, SpecPilot, Caper, Switch-SWE, WebHarbor, LEANGUARD kept. LoCo and MAServeBench scored below threshold. SWE-Trail and Democratizing Robot-Use not printed
HN front page + Algolia 24h + inference slice Hit Postman, Simon/Qwen, RTX Spark, Hy4 (via Techmeme), Siri, Show HNs (MoEShift, llama.app, Octomind). Inference-slice queries ("vllm", "kv cache", "sglang", "mooncake", "tensorrt") returned noise or nothing
GitHub trending Hit auto-ontology, agent-desktop, Scrapling, claude-mem, floci, worktrunk, kimi-k3-in-c
Price Per Token news + agents/coding/open-source/model-releases slices Hit Hy4 (today), Crouton Alpha, StateBench. AgentCore posts and the Fal funding round not printed
LearnPrompt AI News Radar Miss (live) / Hit (fallback) Live site and Pages rendered empty. The repo's latest-24h.json (generated 06:59 CST) was used: GitHub changelogs, Claude Code 2.1.318, Techmeme Hy4 and Pierre, NVIDIA DGX Spark, IBM CUGA
RadarAI Hit AA Capability Indexes v2.0 confirmed. Reward hacking across multi-agent coding systems, BEMULA and Hindsight memory models: no primary found, dropped
The Agentic Digest Hit Supabase/Claude MCP identity (lead), Vercel Security Review; AWS EDB semantic layer not printed
Simon Willison 2026-11-03 Hit Qwen 3.8 27B weights post kept
X (narrow, via x tool) Partial One query: AA Capability Indexes post confirmed

Ranked dailies (in order)

Rank Daily Result Note
1 Latent Space AINews Partial Index only; newest is the 11-03 weight-space issue (covered yesterday); no 11-04 issue yet
2 AlphaSignal Partial Headlines only: Claude MCP auto-connect (confirmed via Claude Help Center), AA Capability Indexes (confirmed). Sarvam Prism, MiniMax H3, MedGemma local and Prime Intellect benchmark audits not opened (no primary named), dropped
3 Code4U AI Digest Hit 11-03: watermarking, RAISE Act, CrowdStrike, S-1 (not builder items)
4 TLDR AI archives Hit 11-03 issue: Framer Harness 2 and GitHub Next Enclaves confirmed. "GPT-6.2 Astra spotted in Codex" is a rumor, dropped. GPU weight teleportation and Copilot Cowork Rust lessons not printed
5 TLDR flagship Not needed TLDR AI loaded
6 The Neuron Hit 11-03 issue; nothing printable
7 AI Daily Brief Hit 11-03 episode on recursive self-improvement; nothing printable

Confirm opens

  • Tencent Hy4: Techmeme (via LearnPrompt), PPT, OpenRouter → confirmed (Tencent research page, press release). HF weights not public
  • AA v5.0.1: AA changelog → confirmed (methodology version history). Capability Indexes v2.0: RadarAI / AlphaSignal → confirmed (AA methodology, AA X post)
  • Postman plugins: HN → confirmed (the GitHub issue itself, with installer source links)
  • Claude auto-connect / Supabase: Agentic Digest / AlphaSignal → confirmed (Supabase, Claude Help Center, WorkOS)
  • Claude Code v2.1.318, GitHub Copilot JetBrains Air / usage-API BYOK / Dependabot APM: LearnPrompt → confirmed (release page and GitHub changelog)
  • RTX Spark Linux: HN → confirmed (MindStudio post; author-measured)
  • NVIDIA DGX Spark MoE: LearnPrompt (MarkTechPost) → confirmed (NVIDIA technical blog)
  • OpenAI / Pierre: Techmeme → confirmed (Axios, OpenAI statement to Axios). No OpenAI blog post found
  • Siri AI for developers: HN → confirmed (Apple Developer News)
  • IBM CUGA harness comparison, StateBench: LearnPrompt / PPT → confirmed (HF blogs)
  • Framer Harness 2, GitHub Next Enclaves: TLDR AI → confirmed
  • Vercel Security Review: Agentic Digest → confirmed (Vercel blog, 11-02)
  • llama.app: HN → confirmed (site + ggml-org release pages); not new, labeled Context
  • "OpenAI silently re-enabling training on Codex chats" (HN, 497 points): a user report with no OpenAI or vendor primary → nominated, no primary, dropped
  • Hy4 free WorkBuddy/CodeBuddy window: X post → confirmed in the Tencent press release
  • Carryovers re-checked: Kimi K3.1 Pro weights → still not published; LongCat-2.5 weights → still not published; Coding Agent Index → unchanged
  • Coding CLI releases checked: claude-code, codex, gemini-cli, opencode, pi, aider, llm (GitHub releases API)
  • Engine blogs (vLLM / SGLang / TRT-LLM): not nominated
  • Blocked: none. The GitHub API rate limit was hit partway through, so github.com release pages were used for the rest