Issue No. 10

The Nov 7 edition

  • 21-minute read
  • 169 primary links
  • 9 sections
  • News30
  • Update6
  • Context4
  • Still open6

On X for this date →

Snapshot#

Saturday, and it was quiet. No new frontier model, no new Artificial Analysis evaluation, and no Hugging Face papers for 11-07. Most of what shipped was harness releases and engineering write-ups.

  • Agent memory, with numbers: Anthropic's The components of agent memory reports results on CL-Bench, Pokémon, and a document corpus. Fixed rules like "record only failures" cut the gain from memory for Opus 5 and Opus 4.7. Offline "dreaming" over transcripts raised a corpus Q&A score from 13% to 48% across 128 sessions. Search plus a wiki scored 61% vs 53% for search alone, and saved about 20K tokens per question.
  • Claude Code v2.1.319 adds background MCP tasks and permissions.denyInsteadOfPrompting, and fixes "a sandbox escape when creating a worktree." The email-in-User-Agent issue is still open. This release's notes do not mention taking the account email out of model context.
  • OpenCode shipped twice: OpenCode Browse is an MCP server plus a Chromium extension that lets any MCP client drive your real signed-in tabs, with handoff for login/2FA and per-session site allowlists. OpenCode v2.2.6 adds MCP connection pooling and opencode run -w worktrees.
  • ARC-AGI-2 per-task cost (ARC Prize leaderboard): Inkling 2 Small scores 81.7% at $0.27/task, the best open-weight result on this list. DeepSeek V4.1 Flash 1017 (Max) scores 78.1% at $0.16. Closed GPT-6.1 Sol (Max) scores 94.2% at $0.25.
  • Open-weight session economics (Vals CyberOffense): DeepSeek V4.1 Flash 1017 matched gated GPT-6 Sol Daybreak on one simulated network for under $3/run vs about $22. Treat it as a cost and safeguards datapoint, not a how-to.
  • Small open models: Nous Arete-27B is an Apache-2.0 SFT of Qwen3.8-27B. LFM3-1.7B is now on OpenRouter at $0.08/$0.20 per 1M, and Hugging Face released harness adapters for it, trained inside pi, opencode and hermes.
  • Cheap judge pattern: Hightouch's Jev + Opus hybrid (HN 513) gets 98.8% vs Opus-alone 99.0% at $0.88 vs $3.86 per 1,000 judgments.

Do this today: upgrade to Claude Code v2.1.319 (claude update) and keep the 2.1.317 email check on. Then read Anthropic's memory post before you write another set of memory-file rules.

Frontier and open-weight models#

No new closed frontier model shipped in the window. The closed datapoints below come from indexes.

  • Update Inkling 2 Small on ARC-AGI independent eval (leaderboard entry dated 2026-11-05; nominated by AINews 11-07).
    • ARC Prize (XHigh effort): ARC-AGI-1 96.5% at $0.08/task; ARC-AGI-2 81.7% at $0.27/task.
    • Same leaderboard, open weights: DeepSeek V4.1 Flash 1017 (Max) 78.1% / $0.16; GLM-5.3-Flash (Max) 65.8% / $0.09; Qwen3.8-27B (XHigh) 42.4% / $0.45.
    • Same leaderboard, closed: GPT-6.1 Sol (Max) 94.2% / $0.25; Claude Opus 5.5 (High) 93.3% / $0.41; GPT-6 Astra (Max) 95.0% / $1.12.
    • Status: still Tinker-only with closed weights. Apache 2.0 weights are promised "next week" in the lab post. AA Intelligence Index 31, $1.18/task (AA, Context from yesterday).
  • News Nous Arete-27B open weights, fine-tune (HF repo created 2026-11-06 13:31 CST).
    • What: SFT only on Qwen3.8-27B: about 52.8M tokens at 64K context, 2 epochs, 1,202 steps. The teacher was Inkling. The goal was to remove CCP-framed answers on seven value categories.
    • Vendor numbers (Nous eval): answers with propaganda 22% → 0.9%; fact score 2.32 → 2.98/5; Hermes Bench 65.77% → 65.60%; Terminal-Bench-2.1 67% → 70%. Nous ran both models on its own setup and explicitly takes no credit for the TB gain.
    • License / serving: Apache-2.0, BF16 safetensors. No hosted inference provider. The model page lists 16 community quantizations. Training data: NousResearch/arete-data-pairs.
    • Builder take: use it if you want Qwen3.8-27B capability without the political framing. Treat the benchmark parity as vendor-reported.
  • Update LiquidAI LFM3-1.7B on OpenRouter API live (listed 2026-11-06; weights on HF since October).
    • OR page: $0.08 / $0.20 per 1M, cache read $0.02; 262K context, 8,192 max output; tools and structured outputs. There is a free variant.
    • Caveats (OR page): one provider, about 18 tok/s P50, 3-day availability 85.79%. The page says prompts and outputs may be retained to train Liquid models.
    • License: LFM License 1.0 (on the HF card).
    • Builder take: a cheap, small tool-calling model. It is also one of the base models in Hugging Face's new harness adapters (§6).
  • Still open DeepSeek V4.1 Flash 1017 usage is still #1 on OpenRouter, with 56.2T tokens this week (+23%) and 7.83T today, at $0.15/$0.60 per 1M.
  • Update Tencent Hy4 traffic is #5 weekly on OpenRouter at 7.05T (new) and #2 today at 3.05T. Price is unchanged at $0.417/$1.2505 per 1M. Weights are still not confirmed on a primary page.
  • Context the OpenRouter benchmarks panel shows the AA Intelligence Index as: Claude Opus 5.5 (Max) 59.1, GPT-6 Astra (Max) 56.9, GPT-6.1 Sol (Max) 54.6, Claude Sonnet 5.5 (Max) 54.5, Qwen3.8 Max 53.4, GPT-6.1 Terra (Max) 51.3, Kimi K3.1 Pro 46.5.

Pareto frontier (quality vs $/token vs $/task vs session economics)#

Each number names its index. AA numbers come from the AA Intelligence Index as shown on the OpenRouter benchmarks panel or from AA articles. ARC numbers come from the ARC Prize leaderboard. Token prices come from OpenRouter model pages.

Model AA index ARC-AGI-2 (score / $ per task) $/1M in / out (OR) Other $/task or session Weights
GPT-6.1 Sol (Max) 54.6 94.2% / $0.25 $2 / $10 — Closed
Claude Opus 5.5 59.1 (Max) 93.3% / $0.41 (High) $4 / $20 Hightouch judge: $3.86 per 1K decisions Closed
GPT-6 Astra (Max) 56.9 95.0% / $1.12 $10 / $50 — Closed
Inkling 2 Small 31 (AA article) 81.7% / $0.27 (XHigh) $0.30 / $1.00 (AA) AA $1.18/index task Apache pending
DeepSeek V4.1 Flash 1017 not on OR panel 78.1% / $0.16 (Max) $0.15 / $0.60 Vals CyberOffense: under $3/run vs ~$22 for GPT-6 Sol Daybreak Open
GLM-5.3 Flash not on OR panel 65.8% / $0.09 (Max) $0.15 / $0.50 — Open
Kimi K3.1 Pro 46.5 not listed $2 / $10 AA $3.68/index task (Context) Open (custom)
LFM3-1.7B not scored not listed $0.08 / $0.20 — Open (LFM 1.0)
  • Where the line sits today (ARC-AGI-2): at about $0.25–0.27 per task, closed GPT-6.1 Sol (Max) scores 94.2% and Inkling 2 Small scores 81.7%. Spending the same per task on Sol buys about 12 more points. DeepSeek V4.1 Flash 1017 gets 78.1% for $0.16, and GLM-5.3-Flash gets 65.8% for $0.09. Those are the cheap open points on the line.
  • Indexes disagree on open models. ARC Prize puts Inkling 2 Small (81.7%) above DeepSeek V4.1 Flash 1017 (78.1%) on ARC-AGI-2. On the AA Intelligence Index, Inkling scores 31 and Kimi K3.1 Pro scores 46. Kimi K3.1 Pro has no ARC entry, so "best open model" depends on which index you read.
  • News Vals CyberOffense session economics, security eval (leaderboard updated 10-23; surfaced 11-06 via Wired and AINews).
    • Scores (simulated enterprise networks, objectives completed): GPT-6 Sol Daybreak 15.7/21 (Nova) and 27/31 (Lighthouse). DeepSeek V4.1 Flash 1017 16.3/21 and 21/31. Kimi K3.1 Pro 13/21 and 20/31.
    • Cost: DeepSeek tied Daybreak on Nova at under $3 per run vs about $22. Vals reports that DeepSeek made most of its progress after 100M tokens, so short runs understate cheap open models.
    • Setup: runs use the open-source Pi agent with a 500M-token cap per run.
    • Builder take: cap evals by tokens, not turns. A cheap model on a long budget can match an expensive model on a short one. Methods are deliberately left out here.
  • News Hightouch: Jev that can ask Opus for help cascade economics (post dated 11-04; HN 513 points in window).
    • Easier test, per 1,000 identity judgments: Opus 5.5 alone 99.0% / $3.86; Hybrid Jev (cases at 30–90% confidence go to Opus) 98.8% / $0.88; Hybrid Luna 98.5% / $0.33; Jev alone 95.7% / $0.05.
    • Harder test: Opus 95.1% / $5.39; Hybrid Jev 95.3% / $2.94; Hybrid Luna 91.6% / $1.15.
    • Lesson: a well-calibrated cheap model escalates more cases on its own when the data gets harder. Luna's overconfidence made it worse in the cascade even though it was better standalone. Vendor (Hightouch) numbers on its own benchmark.
  • OpenRouter weekly usage (through Nov 6, rankings): DeepSeek V4.1 Flash 1017 56.2T; DeepSeek V4.1 Flash 17.2T; GLM 5.3 Flash 14.1T (−13%); GPT-6 Luna 7.22T (−34%); Hy4 7.05T (new); Step 5 Preview 6.83T; Claude Opus 5.5 6.59T; MiniMax M3.1 Flash 5.67T; Hy4 preview 5.11T; GPT-5.6 Luna 4.07T. Top apps today: Hermes Agent 1.85T, Claude Code 1.62T, omp 726B, Kilo Code 667B, pi 659B.
  • Chinese-lab and open-weight movers: Hy4 climbing on OR. DeepSeek V4.1 Flash 1017 shows up on both the ARC and Vals cost lines. Arete-27B is a values fine-tune of Qwen3.8-27B. No new CN-lab weights confirmed in the window. The Qwen 4 teaser is a rumor with no model card.
  • Not available: OpenRouter's "Cost per session" panel renders client-side and could not be read, so there is no OR session-cost number today.

Coding CLIs/IDEs/harnesses#

  • News: Claude Code v2.1.319 (2026-11-06 11:04 CST).
    • Agent loop: MCP tool calls that run as tasks now move to the background right away, and Claude gets the result when they finish (CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND turns this off).
    • Unattended runs: permissions.denyInsteadOfPrompting denies a call instead of stalling on a permission prompt, in terminal, subagents and teammates.
    • Security fixes: "Fixed a sandbox escape when creating a worktree." Bash deny/ask rules no longer miss commands written with quoted words. --trust-workspace false now also blocks repo hooks and skills on self-hosted runners.
    • Behavior change: with telemetry or feature flags off, including behind LLM gateways, auto mode now asks or denies calls the server can't check, with no classifier fallback. Set CLAUDE_CODE_AUTO_MODE_SERVER=0 to keep the old behavior.
    • Also: /claude-api latency-optimize; CLAUDE_CODE_DISABLE_FEATURE_FLAGS; /context sends Claude only the table, not the full grid, on later turns.
  • Still open: Claude Code email-in-User-Agent, issue #107757 is still open (102 comments). The official mitigation is still the 2.1.317 check that asks before a web or shell call sends your account email. The 2.1.319 notes do not mention removing the email from context. Keep a guard such as whoami-guard on.
  • News: OpenCode v2.2.6 (2026-11-06 17:37 CST).
    • MCP: opt-in connection pooling with pool: true, so projects with the same setup share one server process. It also waits up to 2s for slow MCP servers before the first model call.
    • Scripted runs: opencode run -w starts the session in a new git worktree, and --archived keeps CI sessions out of your inbox.
    • Hardening: webfetch and Code Mode fetch reject URLs with embedded credentials and check every redirect against your rules. File reads are served as sandboxed downloads, and the server refuses disallowed origins on /api/*.
  • News: OpenCode Browse (AlphaSignal nominated).
    • What: an MCP server plus a Chromium extension. Agents drive your real browser (Chrome, Edge, Brave, Arc, Helium, Vivaldi, Opera, Chromium) with Playwright, in collapsed background tab groups.
    • Controls: handoff for login, 2FA or payment steps; "lend a tab"; per-session allowlists of sites; screenshots, screen recording and HAR capture.
    • Clients: OpenCode, Claude Code, Cursor, Codex, Zed and other MCP clients.
    • Caveat: this gives an agent your logged-in sessions. Use the allowlist, and keep it away from banking and admin consoles.
  • News: GitHub Copilot weekly releases (Nov 2) (2026-11-06).
    • Copilot app: Agent Merge from a PR, compaction from the context-limit error, and opt-in personal instructions for which tool requests it may approve on its own.
    • Copilot CLI: /settings is a unified dashboard for plugins and MCP servers.
    • New surfaces: Copilot in JetBrains Air and in Xcode's coding assistant.
  • News: Copilot code review API — manage review settings over REST and GraphQL. Responses show the review effort used and which skills, MCP servers and custom instructions Copilot cited for a comment. Also: @GitHub in Slack/Teams is GA and the impact dashboard adds Chat and the VS Code Agents window.
  • News Claude Code remote sessions on Bedrock AgentCore pattern (AWS ML blog). Each task runs in its own isolated session in your AWS account, all model calls go through Bedrock, and the result comes back as a PR. It is marked SUCCEEDED only if the repo's tests passed, otherwise SUCCEEDED_UNVERIFIED as a draft.
  • Codex: no stable release in the window. Only the prerelease rust-v0.169.0-alpha.2 shipped. 0.167.0 from yesterday is still current.
  • Still open Conductor deadline — Cloud goes read-only 11-13 and shuts down 11-20.

Usable tools and open projects#

Star counts are from the GitHub API, read around 08:20 CST. "+24h" is the stars-today figure on GitHub trending (daily) at about 07:40 CST.

Repo or tool What it is Stars / +24h Runnable today? Why a tinkerer should care
OpenCode Browse MCP server + Chromium extension that drives your real browser n/a (in anomalyco/opencode, 225,184★) Runnable (News) Browser use with your logins, allowlists and human handoff, from any MCP client
vercel-labs/agent-browser Rust browser-automation CLI + MCP for agents 46,459 / +180 Runnable (Still open; v0.41.2 11-03) Headless counterpart to Browse. agent-browser mcp call <tool> '<json>' works without a stdio session
ranxianglei/billion-context Proxy that compresses context for Anthropic/OpenAI-API agents 4,434 / +185 Watch (News; benchmarks are in its own README) Claims 95–97% prefix-cache hit and month-long sessions on 100K windows. Measure it yourself before trusting it
huggingface/safeagents Proxy that checks every tool call between harness and model 92 / n/a Runnable (News) Rule-based blocking of reward-hacking moves. Usable as a guard outside RL too
HF harness adapters LoRA per model × harness (pi, opencode, hermes) n/a Runnable (News) Swap a local model for the adapter trained in your harness. Fewer tool calls at the same or better accuracy (HF smoke tests)
libghostty-vt 1.0 Stable C ABI terminal-emulation library from Ghostty 63,798 (ghostty) / n/a Runnable (News; HN 770) Base for agent terminals like cmux and herdr, and now the JetBrains terminal. Use it instead of writing your own emulator
tharun208/cloneguard Read-only scan of a repo for credential theft, hidden network calls and auto-run hooks 17 / n/a Runnable (News; Show HN) Run it before letting an agent npm install a stranger's repo. MIT, no dependencies
NousResearch/Arete-27B Apache-2.0 values fine-tune of Qwen3.8-27B 139 likes (HF) Runnable (News; 16 quants on HF) Local 27B with Qwen capability (vendor eval)
LFM3-1.7B on OpenRouter 1.7B tool-calling model, 262K ctx n/a Runnable (Update) $0.08/$0.20 per 1M, or free. Watch the data-retention note
dsx-ai-factory/nvke NVIDIA Kubernetes Engine: management plane for bare-metal GPU k8s 374 / +125 Watch (News) Only relevant if you run your own GPU nodes
cloudflare/security-audit-skill, mattpocock/skills, google/skills Agent skill packs 64,940 / +2,088; 317,239 / +1,111; 23,098 / +449 Skip (momentum only; covered 11-06) Skill packs with no runtime of their own

Keepers:

  • OpenCode Browse vs agent-browser. Use Browse when the task needs your logged-in sessions and a human step. Use agent-browser for isolated, scripted runs. Both expose MCP.
  • billion-context points your agent's base URL at a local proxy that rewrites the stream with incremental, reversible summaries. Its README cites ContextBench, LongMemEval-S and MemoryAgentBench wins from a run it labels third-party. Measure your own cache hit rate before relying on it.

Agent builder stack (protocols, runtimes, evals, context engineering, sandboxes, computer use)#

  • News Anthropic: The components of agent memory context engineering (Nov 6; HN 76).
    • Scope: "Agent memory" means notes kept between sessions. Weights, in-context learning and compaction are out of scope.
    • Salience: in Pokémon, Sonnet 3.7 wrote 413 files across 225 folders and earned 1 badge. Fable 5 wrote 9 files in 2 folders plus global.md and earned 4. The lesson: save what the environment won't show you again.
    • Steering: tell the model what memory is for, not how to use it. A fixed "record only failures" rule cut the memory gain for Opus 5 and Opus 4.7. Older models (Sonnet 4.6) did benefit from being told to read memory first, and that nearly doubled its score.
    • Dreaming (offline consolidation over transcripts): corrected a wrong navigation note that had caused repeated failures. On a corpus of about 350K files (about 420M tokens), wiki building over 2 → 128 sessions raised the score from 13% → 48%. Search + wiki reached 61% vs 53% for search only, and saved about 20K tokens/question. On merger Q&A the wiki scored 87% vs 73% for search.
    • Limit: on single-document tasks the wiki added nothing, and once it grew larger than the source documents.
  • News HF harness adapters RL + safety for harnesses (Nov 6).
    • What: LoRA (rank 32) adapters trained with GRPO (TRL) inside unmodified pi, opencode and hermes. Base models: LFM2.5-24B-A2B, LFM3-1.7B, and TwIL-8B as an experiment.
    • Tasks: built from shared agent trajectories with Repo2RLEnv (repair, qa, build, extend).
    • Reward hacking: safeagents blocks those tool calls and scores them 0.
    • Eval cost: 9 checkpoints, 1,350 agent sessions, one H200, under an hour, about $5, vs about $90 if each checkpoint were merged and served separately.
  • News ChatGPT routed a follow-up tool call through a different connector MCP auth lesson (HN 90).
    • What happened: two connectors pointed at the same MCP server. The model made the confirm step's follow-up call through the other connector without telling the user.
    • What held: the server bound the one-time confirmation code to connector A's OAuth client, so it rejected the code from B.
    • Takeaway for MCP server authors: bind pending operations to the OAuth client and session, not only to the user or tool name.
  • News Claude Code v2.1.319 permissions — denyInsteadOfPrompting for unattended agents, background MCP tasks, and a worktree sandbox-escape fix (§4).
  • News Replit rebuilt its sandbox engine on microVMs sandboxes (dated 11-04; HN 93). Covers the isolation boundary move, minimal kernel surface, network isolation, credential handling, and "bottomless" storage across hundreds of millions of projects. Part 2 covers Kubernetes orchestration.
  • News Firecracker VM escape, PwnEscape Part 2 sandboxes, security report (HN 257). A researcher reports guest-to-host root on a Firecracker setup, building on a host memory primitive from Part 1. No details here. If you run agent code in microVMs, track host kernel and Firecracker advisories, and keep the jailer and seccomp on.
  • News OpenAI outbound SIP calling for all orgs voice agents (Agentic Digest 11-07; HN 65). Uses the Live API (POST /v1/live/sessions, transport.type: "sip"). Requires TLS signaling (sips:), SDES-SRTP media, digest auth, and E.164 destinations. Private or local IPs are rejected.
  • News GitHub's new Git storage architecture agent-scale git (11-05; HN 168). In internal testing: 5,000 concurrent agents and 7,200 pushes/s to one repo.
  • News evals with code, <a href="https://huggingface.co/papers/date/2026-11-06" rel="noopener" class="ext">HF papers 11-06</a>
  • Context Okta Cross App Access + AgentCore AWS agent governance, digest-nominated, docs opened only for Claude Code and a serverless MCP server on ECS Fargate are reference patterns, not new products.

Infrastructure dash#

  • Update AA-AgentPerf, DeepSeek V4.1 Flash on GB300 NVL72 serving speed (AA post 11-06).
    • Speed: single-agent p25 output speed of 848 tok/s, up from 596 tok/s a month ago (1.42x). The AA datacenter data holds the 848.1 tok/s run (concurrency 1, run date 2026-10-28).
    • Stack (configs page): NVIDIA Dynamo frontend + vLLM, TP2, MXFP4, DSpark speculative decoding with 5 draft tokens, FlashInfer autotune.
    • Builder take: the gain came from software on the same hardware. Re-benchmark your serving stack before you buy more GPUs.
  • News NVIDIA Reshard collective RL infra (Nov 6). Moves weights from trainer to inference as a DTensor-to-DTensor transfer (NCCL M2N Reshard) instead of gather-and-broadcast. Vendor numbers: 2.7–5.7x faster refit in NeMo RL on DGX H100 and GB200 (Qwen 3 235B, DeepSeek v3 671B).
  • News SageMaker Inference supports Multi-Instance GPU GPU sharing. Up to 7 hardware-isolated partitions per GPU, a different model per partition. The layout is fixed when the endpoint is created, and models that need more than one GPU still need whole devices.
  • News GitHub Git storage agent-scale git — 7,200 pushes/s (see §6).
  • News Google acquired Crusoe's Wyoming data center site capacity (Nov 6). Crusoe reports more than 6 GW contracted and over 40 GW in its pipeline. Capacity context only; no price change.
  • Update OpenRouter rankings demand through Nov 6. DeepSeek Flash 1017 still has about 3x the next model's weekly tokens. Hy4 is new to the weekly top 10. Crouton Alpha is #8 today (1.05T).
  • Engine blogs: not nominated. The HN inference slice returned nothing above 7 points, and vLLM appears only inside AA's AgentPerf config.

Builder actions#

  1. Upgrade Claude Code: run claude update to get v2.1.319 (worktree sandbox-escape fix). Set permissions.denyInsteadOfPrompting for unattended or CI agents. Keep the 2.1.317 email check and whoami-guard while #107757 is open.
  2. Rewrite your memory instructions: following Anthropic's memory post, replace "how to write memory" rules with a one-line statement of what memory is for. Add a periodic "dream" pass that checks notes against transcripts.
  3. Try OpenCode Browse with an allowlist: npm install -g @opencode/browse && browse install (docs). Point Claude Code or OpenCode at it on one logged-in internal tool. Do not use it on banking or admin consoles.
  4. Update OpenCode: install v2.2.6, set pool: true on shared MCP servers, and use opencode run -w --archived for CI jobs.
  5. Build a cascade judge: copy Hightouch's setup. A cheap calibrated model decides first and sends its 30–90% confidence band to a frontier model. Measure accuracy and $/1K on your own eval.
  6. Price ARC-style reasoning before picking a model: compare GPT-6.1 Sol (Max) 94.2% at $0.25 against Inkling 2 Small 81.7% at $0.27 and DeepSeek V4.1 Flash 1017 78.1% at $0.16. At about the same per-task cost, the closed model scores roughly 12 points higher.
  7. Cap your evals by tokens, not turns: Vals shows cheap open models catching up only after 100M+ tokens. Give each model the same token budget.
  8. Local small-model agent: try LFM3-1.7B through OpenRouter, or pull the weights with the matching harness adapter for pi, opencode or hermes. Put safeagents in front of it.
  9. Scan before you clone: run cloneguard on any unfamiliar repo before an agent installs its dependencies.
  10. MCP server authors: bind confirmation codes and pending operations to the OAuth client and session (connector routing report).
  11. microVM sandbox operators: read Replit's write-up, check host kernel and Firecracker advisories in light of the escape report, and confirm the jailer and seccomp are on.
  12. Still open — Conductor Cloud: export before 11-13 (announcement).
  13. Still open — Inkling 2 Small weights: Apache 2.0 weights are due "next week" per the lab post. Watch the model card.

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 (no new article) Newest article VIALS-AA (Nov 6, covered). Changelog newest entry 5 Nov. AgentPerf 848 tok/s confirmed on the datacenter page data
OpenRouter rankings + models API Hit Usage through Nov 6. New ID: liquid/lfm-3-1.7b. Cost-per-session panel is client-rendered and was not readable
Hacker News front_page API Hit Libghostty (770), Jev+Opus (513), Firecracker escape (257), Git storage (168), Agent memory (76)
HN AI/LLM/agent last-24h (Algolia) Partial The OR query returns 0 on Algolia. Fell back to all stories over 10 points in the window (107) and picked by hand
HN inference slice Partial Ran each keyword separately. inference 11 hits, vllm 23 (mostly noise), sglang 0, "kv cache" 0, llama.cpp 1. Nothing over 7 points
GitHub trending daily Hit agent-browser, billion-context, nvke, skill packs (Skip), OpenMausBot, zed
HF papers 2026-11-06 Hit BestRand, VisSearch-Bench, Terminal-style Math Modeling Bench
HF papers 2026-11-07 Miss Redirects to 11-06. No papers for Saturday
HF models trending Hit Arete-27B (new 11-06), LFM3-1.7B, Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017
pricepertoken.com/news (+ agents, coding, open-source, model-releases) Hit Feed last updated about 10h before fetch. LFM3 OR listing, AWS AgentCore posts, SageMaker MIG, Crusoe
news.learnprompt.pro Partial SPA shell only
learnprompt.github.io/ai-news-radar Partial SPA shell only
LearnPrompt data JSON Hit data/latest-24h.json and daily-brief.json from the master branch, generated 05:43 CST. Claude Code v2.1.319, Copilot changelogs, NVIDIA Reshard
radarai.top/en Hit Agent memory, Browser Use a11y-tree tip, Muse upgrade, DeepMind RSA-1024
The Agentic Digest Hit 11-07 issue: OpenAI outbound SIP
simonwillison.net 2026/Nov/6 Hit One quote post (Nilay Patel). Zero builder nominations
X narrow (AA, karpathy, swyx, simonw, cursor_ai, claudeai, OpenAI, AIatMeta, huggingface) Hit 22 posts. Nominated: AA AgentPerf, GPT Image 2.6 on AA (covered 11-05), BioBasicsBench, FMF jailbreak framework, Muse upgrade

Dailies in rank order

  1. Latent Space AINews, "Anthropic IPO: Whither Now?" (Nov 7, covers 11/05–11/06): Hit. Nominated: ARC Prize Inkling (confirmed), Vals CyberOffense (confirmed), AA AgentPerf (confirmed), HF harness adapters (confirmed), Git storage (confirmed), Arete (confirmed). DeepMind RSA, Qwen 4 teaser, Stanford collab bench, Zyphra ZYFA: not printed.
  2. AlphaSignal: Partial (several PRO-gated). OpenCode Browse (confirmed), Arete (confirmed). Unsloth Qwen-Image-2.1, AT&T OTel 2.0, Qwen3-ASR: PRO, no primary opened.
  3. Code4U digest: Hit (Nov 6). Perplexity Teammates, METR, GPT-6 Cyber, Conductor, ERNIE 5.5 Flash Base. All covered before or not builder items.
  4. TLDR AI: Hit (Nov 6 issue). METR, GPT-6 Cyber, Inkling, Omni, Cloudflare agent-trust post. Covered or narrative.
  5. TLDR flagship: Not used (TLDR AI was not blocked).
  6. The Neuron: Hit. Latest is Nov 6 (GPT-6 Cyber, Claude finance hub). Covered or consumer.
  7. AI Daily Brief: Miss. Latest edition is Nov 5.

Nominated → confirmed or dropped

Event Status
Anthropic agent memory nominated → confirmed (Anthropic)
Claude Code v2.1.319 nominated → confirmed (GitHub release)
OpenCode v2.2.6 / OpenCode Browse nominated → confirmed (GitHub release / opencode.ai)
HF harness adapters + safeagents nominated → confirmed (HF blog + repos)
Hightouch Jev hybrid nominated → confirmed (Hightouch blog)
Inkling 2 Small ARC-AGI nominated → confirmed (ARC Prize leaderboard)
Vals CyberOffense nominated → confirmed (vals.ai)
AA-AgentPerf GB300 848 tok/s nominated → confirmed (AA datacenter data + configs)
Arete-27B nominated → confirmed (HF model card)
LFM3-1.7B on OR nominated → confirmed (OR page + HF card)
Copilot changelogs (4) nominated → confirmed (github.blog changelog)
ChatGPT connector routing nominated → confirmed (OpenAI community thread)
OpenAI outbound SIP nominated → confirmed (OpenAI docs)
Replit microVMs / GitHub Git storage / libghostty 1.0 nominated → confirmed (vendor posts)
Firecracker escape report nominated → confirmed exists (pwn.ai); method omitted
NVIDIA Reshard / SageMaker MIG / AWS Claude Code on AgentCore / Crusoe→Google nominated → confirmed (vendor blogs)
DeepMind RSA-1024 cryptanalysis nominated, no primary found (X/LinkedIn + press only), dropped
Meta Muse specialized models nominated (X only, consumer), dropped
Qwen 4 teaser nominated, rumor, no model card, not promoted
Perplexity Teammates nominated, primary returned 403, dropped
OpenAI BioBasicsBench / FMF jailbreak framework / Hacktron nominated, openai.com returned 403 to curl, not opened, not printed
AA Endpoint Accuracy v1.1 numbers nominated, AA methodology page not found, not printed
Inkling "checkpoints circumvented guardrails" nominated (X only), not printed
AlphaSignal PRO items (Unsloth Qwen-Image, AT&T OTel 2.0, Qwen3-ASR) nominated, no primary, dropped
Google Antigravity 2.29 / Claude Duet sessions seen in today's On X pulse; not nominated by brief roster; Antigravity changelog is an SPA; not printed

Engine blogs: not nominated