Issue No. 4

The Nov 1 edition

  • 20-minute read
  • 163 primary links
  • 9 sections
  • News12
  • Update5
  • Context3
  • Still open10

Snapshot#

  • Quiet weekend day.
  • News DeepSeek open-sourced dsh-superagent-prototype (MIT, 5,634★, created 10-31 04:32 CDT).
    • It is a proactive coding agent for one GitHub repo, packaged as a DeepSeek Harness bundle.
    • It listens to GitHub (and optionally Feishu), opens a work item per request, works in its own session and directory, and answers with a PR.
    • DeepSeek labels it a research prototype, "not an official DeepSeek product".
  • News (plan economics): OpenAI's ChatGPT and Codex usage limits help page says that starting in December:
    • Plus, Pro and Business allowances become one shared pool across ChatGPT and Codex, measured in tokens.
    • Instant chat does not draw from the pool.
    • Combined usage shows under Settings → Usage, and you can buy credits.
    • HN: 187 points.
  • Update (session cost data): OpenRouter published median real Claude Code session costs by model for the 30 days to Oct 18 (dated 10-30, first seen today). A 2–9 turn session cost:
    • Opus 5.5: $0.305
    • Sonnet 5.5: $0.155
    • Kimi K3: $0.117
    • GLM-5.3: $0.044
    • DeepSeek V4.1 Flash: $0.009
    • GPT-6 Luna: $0.005
    • Fable 5.1: $1.049
  • News (agent stack):
  • News (open-weight serving): NVIDIA posted DeepSeek-V4.1-Flash-1017-NVFP4 (MIT).
    • Size: about 400 GiB, down from about 475 GiB for the source.
    • Accuracy stays within ±0.7 points of the source on six benchmarks in NVIDIA's own runs.
    • Serving: Blackwell only, and it needs a vLLM build with PR #61530.
  • News (tool): Memdoor's team-solve workflow has five Qwen3.8-27B roles share a problem through files on disk.
    • It scored a mean 9.8/10 on 10 LiveCodeBench Hard problems (5 passes), against 9.6 for the original paper's harness and 8.0 for one-call Fable 5.
    • Median cost was $0.39 per run.
    • The authors say plainly that the sample is small, contamination is possible, and they are claiming parity, not a gain (HN, 300 points).
  • Do this today:

Frontier and open-weight models#

NVIDIA DeepSeek-V4.1-Flash-1017-NVFP4 — Update (HF repo created 10-30 12:37 CDT; first seen today)#

  • What: an NVFP4 quantization of DeepSeek-V4.1-Flash-1017 made with NVIDIA Model Optimizer. MIT license, the same as the base model.
  • Size: about 400 GiB, against about 475 GiB for the source checkpoint.
  • Numbers (source → NVFP4, from the NVIDIA model card):
    • GPQA: 92.77 → 92.93
    • AA-LCR: 80.41 → 80.19
    • SciCode: 57.58 → 57.32
    • IFBench: 76.33 → 76.47
    • MMMU-Pro: 78.38 → 78.73
    • Terminal-Bench 2.1: 81.81 → 81.11
    • NVIDIA ran these itself; there is no independent check yet.
  • Serving path:
    • Requires a vLLM build that includes PR #61530 (MXFP4 Engram).
    • Tested on the vllm-openai v0.31.0 image on 4× GB300.
    • Blackwell only. SGLang is not validated.
  • Builder take: the ~75 GiB saving makes a single 4-GPU Blackwell node comfortable for the most-used open model on OpenRouter. It does nothing for Hopper or consumer cards. On those, stay with the GGUF routes listed in section 7.

Still open#

  • Mistral Large 4 (open weights): still the Western open-weight datapoint.
  • Kimi K3.1 Pro: AA still says "Weights currently not published, Moonshot is expected to release the weights soon after launch."
    • The Moonshot HF org still tops out at Kimi-K3 (2026-06).
    • The Agentic Digest says the weights are promised by Nov 4. I found no Moonshot primary for that date, so treat it as unconfirmed.
  • Gemini 4 Carbon: still the closed frontier datapoint, at 59.4 and $7.74 per task with Antigravity (Coding Agent Index). Access is still limited.
  • GPT-6.1 Terra: $1/$5 on OpenRouter. It still holds the cheap end of the Codex rows (section 3).

Pareto frontier#

Coding Agent Index (Artificial Analysis, v2.0: FrontierCode v1.1 + Terminal-Bench 4.0 + SWE-Atlas-QnA-AA).

  • The data was re-materialized on 2026-10-30 at 13:04 CDT. I compared all 34 rows field by field with yesterday's read, and every score, $/task, cache-hit rate and token count is unchanged.
  • Cost is mean USD per task.
Harness + model Index $/task Cache hit · tokens/task Note
Claude Code + Opus 5.5 (Max) 61.7 $11.42 95.3% · 14.5M Top score
Devin Fusion + 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 + Gemini 4 Carbon (High) 59.4 $7.74 Limited access
Claude Code + Sonnet 5.5 (Max) 59.3 $13.30 · 24.1M Costs more per task than Opus here
Codex + GPT-6 Astra (Max) 54.7 $7.04
Codex + GPT-6.1 Sol (Max) 54.1 $1.50 Strong mid-tier $/point
Opencode + GLM-5.3 (Max) 48.9 $5.23 CN open-weight
Mistral Vibe + Mistral Large 4 (High) 48.3 $1.26 91.2% · 8.5M Open weights
Codex + GPT-6.1 Terra (Xhigh) 48.2 $0.69
Codex + DeepSeek V4.1 Flash (Max) 47.3 $0.39 98.3% · 27.2M CN open-weight
DeepSeek Harness + V4.1 Flash 1017 (Max) 46.3 $0.30 96.8% · 16.7M Cheapest row above 45
Codex + GPT-6.1 Terra (Medium) 42.0 $0.39
Kimi Code CLI + Kimi K3 41.4 $5.17 CN open-weight, expensive per point
Mistral Vibe + Mistral Large 4 (Medium) 35.4 $0.78
Codex + GPT-6 Luna (Max) 31.1 $0.16 Floor

Read it like this: the tiers haven't moved.

  • Frontier tier (≥59): Copilot Fusion costs about half of what Claude Code + Opus costs per task, for 1.8 fewer points.
  • Value tier (~46–48): DeepSeek Harness + V4.1 Flash 1017 at $0.30, Codex + DeepSeek V4.1 Flash at $0.39, and Codex + Terra Xhigh at $0.69 all sit within about 2 points of Mistral Large 4 High at $1.26.

Session economics (today)

  • Real Claude Code session costs (OpenRouter, median, API-key sessions on OpenRouter, 30 days to Oct 18):
Model 1 turn 2–9 turns 10–49 turns
Claude Fable 5.1 $0.342 $1.049 $3.734
Claude Opus 5.5 $0.136 $0.305 $1.390
Claude Sonnet 5.5 $0.071 $0.155 $0.596
Claude Haiku 5.5 $0.002 $0.007 $0.053
Kimi K3 $0.040 $0.117 $0.825
GLM-5.3 $0.010 $0.044 $0.363
DeepSeek V4.1 Flash $0.003 $0.009 $0.047
GPT-6 Luna $0.002 $0.005 $0.024
  • OpenRouter says most of the cost is Claude Code's own context. It is written to the prompt cache on the first request and read back on every turn.
  • The raw numbers are in OpenRouter's session-cost dataset.
  • Index disagreement:
    • On AA's index, Sonnet 5.5 (Max) costs more per task than Opus 5.5 ($13.30 vs $11.42).
    • On OpenRouter's real sessions, Sonnet costs about half as much as Opus at every length.
    • The difference is workload: AA runs long benchmark tasks at Max effort, while OpenRouter's medians come from mostly short everyday sessions. Pick the view that matches how you work.
  • ChatGPT + Codex pooled limits (OpenAI help):
    • From December, one token-measured allowance covers both ChatGPT and Codex on Plus, Pro and Business.
    • Heavy Codex weeks will cut into chat, and the other way round. Instant chat is excluded.
    • Extra credits can be bought.
  • Memdoor team-solve (README, author-run):
    • Median $0.39 and 40 minutes per problem; 50 runs cost $35.79.
    • Models were Qwen3.8-27B on OpenRouter.
    • This gives a concrete $/solve for a multi-agent small-model loop.
  • OpenRouter usage (rankings, data through Oct 31):
    • Top apps by tokens: Hermes Agent 1.61T, Claude Code 1.12T, Kilo Code 681B, OpenClaw 608B, pi 425B, Codex 247B.
    • New entrants: omp (242B), DeepSeek Harness (219B) and Strix (214B).
    • Weekly model leader: DeepSeek V4.1 Flash 1017 with 47.8T tokens (+55%).
    • MiniMax M3.1 Flash went to 5.67T (>999%), and Step 5 Preview to 6.3T (+122%).
  • List-price churn: several OpenRouter list prices moved overnight. For example, z-ai/glm-5.3 completion price is now $3.60/M. These are routing-level prices that swing daily, so check your provider route before you budget.

Coding CLIs, IDEs and harnesses#

DeepSeek dsh-superagent-prototype — News (10-31 04:32 CDT)#

  • What: a reference "Super-agent" design built on DeepSeek Harness. It works on one GitHub repo, takes requests from GitHub (and Feishu), keeps per-work-item sessions and directories, replies with PRs, and has a web panel that shows what needs a human.
  • Status: a research prototype. DeepSeek says much of it was written by coding agents and it is "not identical to our internal version." It differs from the Super-agent design article: some ideas there are future directions that this code does not implement.
  • Try it: the README's setup is to point your coding agent at its SETUP.md. Run it against a scratch repo with a dedicated, least-privilege GitHub token, not your main account.
  • Builder take: the idea worth taking is a resident router with isolated worker sessions that are triggered by events, instead of one chat session per task. You can copy that pattern on any harness.
  • HN threads: 1, 2.

OpenRouter Responses API server-side tools: AI code reviewer tutorial — News (11-01)#

  • A standard-library Python reviewer that uses three beta server-side tools on OpenRouter's Responses API:
    • the Files API, which uploads the changed files plus the diff;
    • a Shell tool that runs a hosted Linux sandbox, so the model can prove a bug by running a check;
    • Apply Patch, which returns fixes as V4A diffs that OpenRouter syntax-checks.
  • Numbers (OpenRouter's run, one five-bug change):
    • A review takes 1–4 minutes and costs $0.015–$0.045.
    • Haiku 5.5 averaged $0.0153 per review.
    • Sonnet 5.5 fixed all 5 bugs in all 3 runs, at $0.0447 and about 60 s.
  • Caveat (from the post): your changed files go to OpenRouter and to the model provider. Only run it on code you are allowed to share, and keep secrets out of the uploaded files.

Gemini API CLI (gemini-api) beta — Context (10-27, first seen today)#

  • The Gemini API changelog lists it as a stateless, zero-dependency command-line client for coding agents, scripts and CI, built on the Interactions API.
  • It is a scripting primitive, not a coding agent. Don't confuse it with gemini-cli.

OpenIDEA — News (HN)#

  • An Apache-2.0 IDE in Rust that copies the IntelliJ workflow. It covers 60+ languages and includes a debugger, Git, a terminal, and DB/HTTP clients.
  • Author benchmark (opening freqtrade, warm caches, median of 5 runs):
    • Ready in 1.1 s vs 13.2 s for IntelliJ IDEA 2026.2.3.
    • 169 MB vs 2,011 MB of memory.
    • The README notes that IDEA analyses Python more deeply, so "ready" doesn't mean the same work on both sides.
  • 2,547★ (HN).

No change#

  • Claude Code is still v2.1.316. Remember that Opus 4.8 fast mode is removed Nov 30 (see Still open in section 8).
  • Codex CLI is still 0.166.0.
  • gemini-cli shipped only nightly v0.67.0-nightly.20261101; stable is still v0.66.0.
  • opencode is still v2.2.5.
  • pi is still v1.4.0.
  • aider is still v0.86.0.
  • DeepSeek Harness is still dsh-v0.2.1-rc.5.

Usable tools and open projects#

Repo What Stars / +24h Verdict Why
guregodevo/memdoor Single-binary coding-agent CLI with YAML workflows; the team-solve example runs plan → 2 ideators → solve → test → review → repair → verify through files on disk. Apache-2.0 129 / n/a Runnable Works with OpenRouter, Anthropic, OpenAI, Gemini or local (Ollama/llama.cpp/vLLM/LM Studio). Their result: 9.8/10 vs 9.6 for the paper's harness, $0.39 median per run. Read install.sh before you pipe it to bash
anomalyco/hex Local-first voice dictation in Rust, from the opencode team; on-device transcription that pastes into the active app. MIT 2,650 / +1,329 Runnable v3.1.0 and v3.1.1 shipped overnight. macOS stable; Windows and Linux are beta. Useful for talking to your coding agent
datasette/datasette-llm Datasette plugin; 0.1a9 now uses LLM's default model 20 / n/a Runnable uvx --with datasette-agent datasette content.db --internal internal.db (Simon Willison)
nolanlawson/agent-framework-benchmark Benchmark harness: Claude Code builds the same app in each JS framework, and the harness measures time, tokens, LOC and Lighthouse 6 / n/a Runnable (learn-from) Write-up: React Router's median run cost about $1 more than Hono+htmx, and Hono+htmx needed about 70% of Angular's code. Runs are very noisy (HN, 144 points)
deepseek-ai/dsh-superagent-prototype Proactive GitHub coding agent (DeepSeek Harness bundle). MIT 5,634 / n/a Watch A research prototype; use a scratch repo (section 4)
FlashML-org/FreeToken Edge MoE serving engine: CPU–GPU co-execution, expert caching, semantic KV checkpoints for agent context edits; Anthropic/OpenAI-compatible APIs. Apache-2.0 15,132 / +232 Watch uv pip install "freetoken[accel]". It claims 290B+ MoE models on RTX 30/40/50 at interactive speed, which I have not verified (paper)
l0ng-ai/openidea Rust IntelliJ-style IDE. Apache-2.0 2,547 / n/a Watch Fast in the author's benchmark; no built-in agent story yet (section 4)
divagr1925/QwenPU A CPU compiled into Qwen weights, from the Qwenputer author. Apache-2.0 4 likes Watch (learn-from) A follow-up to nesformer (covered 10-31)
CopilotKit/OpenBot Self-hosted "AI coworkers", each with its own computer. MIT 9,263 / +178 Still open · Watch Covered 10-29
thedotmack/claude-mem Persistent cross-session memory for agents 108,574 / +1,527 Still open · Runnable Covered 10-30
pingdotgg/t3code Coding-agent GUI 36,730 / +3,100 Still open · Watch Covered 10-30; biggest daily gain on trending
lightpanda-io/browser Headless browser for agents (AGPL-3.0) 39,677 / +827 Still open · Runnable Covered 10-30
milind-soni/OpenMausBot Open agent with a VM 10,166 / +919 Still open · Watch Covered 10-30
Donchitos/Claude-Code-Game-Studios, patchy631/ai-engineering-hub Prompt/skill pack; tutorials collection +970 / +573 Skip Prompt packs and tutorials, not tools

Stars are GitHub API totals, and "+24h" values come from GitHub trending daily, read 2026-11-01 morning CST. "n/a" means the repo was not on today's trending page. One trending repo that collects model jailbreaks was skipped and is not linked.

Agent builder stack#

  • News: Temporal AI framework integrations GA (10-30; repo, MIT).
    • Durable-execution plugins are now GA for:
      • Google ADK (Python/Go/TypeScript)
      • OpenAI Agents SDK (Python/TypeScript)
      • Strands
      • LangGraph
      • LangSmith tracing
      • Vercel AI SDK
      • Google GenAI
    • Workflow Streams is in public preview.
    • Take: if your agent loop dies on deploys or long tool waits, this is the least-effort way to get retries and resume without rewriting the agent.
  • Update: Gemini API Computer Use is GA (10-30) on both the Interactions and GenerateContent APIs (docs).
    • It entered public preview in June with browser, mobile and desktop environments, configurable safety policies, and prompt-injection detection.
    • GA means production SLAs apply. It does not tell you anything about reliability, so measure on your own flows.
  • News: Slack Code API GA. "Code channels" let a coding agent work in a Slack channel tied to a repo.
    • Slack says Copilot opens a code channel in 30% of its sessions, and 150+ teams requested access.
    • Useful if your team already reviews agent work in Slack.
  • News: Ling 3.1 Flash 1019 free on Vercel AI Gateway until Nov 3.
    • Model ID inclusionai/ling-3.1-flash-1019, 1M context.
    • Set up with npx vercel ai-gateway setup.
    • A free window to benchmark it as a cheap sub-agent model.
  • Update: Deepgram is acquiring Daily (10-29; Daily's post). Yesterday this was dropped for lack of a primary; both company posts now confirm it.
    • Pipecat stays BSD-2 under an independent Pipecat Foundation.
    • Deepgram also proposes a "Real-Time Context Standard" open protocol.
    • If you build voice agents on Pipecat, nothing changes today. Watch the foundation's governance.
  • Context: security reports, not detailed here. Simon Willison linked a fixed ChatGPT account-takeover report (bounty paid). Cisco Talos published on an infostealer that uses LLMs (Talos post). No action is needed beyond keeping sessions and extensions current.

Infrastructure dash#

  • Update: DeepSeek-V4.1-Flash-1017-NVFP4 (section 2).
  • News (research): GAAC lossless KV-cache compression. It encodes BF16 KV exponents losslessly, with a fused attention kernel. Authors' numbers:
    • 1.405–1.407× lossless BF16 KV compression.
    • On DGX Spark: 1.19–1.27× decode-attention speedup at 64K context.
    • About 26% less resident KV memory and 13–29% more serving throughput under the same KV budget.
    • Also tested on Apple M4.
    • There is no public code link on the abstract page, so this is Watch only.
  • Context: FreeToken (section 5) is the consumer-GPU MoE engine to watch. Its semantic KV checkpoints target agent context edits, such as tool calls and thinking blocks, so the context doesn't have to be recomputed.
  • Engine blogs (vLLM, SGLang, TRT-LLM, InferenceX): not nominated by HN. The HN inference slice surfaced only arXiv KV-cache papers (GAAC, Gandalf SSD offload, UlaCache), each with 2 points or fewer.

Builder actions#

  1. Benchmark Ling 3.1 Flash for free before Nov 3. Run npx vercel ai-gateway setup, then point a sub-agent or extraction eval at inclusionai/ling-3.1-flash-1019 (changelog). Compare it with DeepSeek V4.1 Flash on the same prompts.
  2. Price your sessions with real data. Use OpenRouter's session-cost table.
    • If your default is Opus, try Sonnet 5.5 for a day and send background and sub-agent work to Haiku 5.5.
    • Set a credit limit on the API key.
    • To try other models in Claude Code: ori claude --model anthropic/claude-sonnet-5.5. Read the Ori install script before you run it.
  3. Prepare for pooled ChatGPT + Codex limits. Open Settings → Usage now to see your combined burn, and decide before December whether long Codex runs should move to an API key (OpenAI help).
  4. Try a five-role small-model team. Follow the memdoor team-solve "Run it" steps.
    • memdoor connect openrouter (or memdoor connect local with a 32K-context Ollama/llama.cpp/vLLM).
    • Copy examples/workflows/team-solve to ~/.memdoor/workflows/.
    • Run memdoor workflow run team-solve --wait on a problem that has its own tests/.
  5. Add a sandboxed reviewer to a repo. Copy the OpenRouter code-reviewer script (Python 3.11+, Git). Run it on a non-sensitive repo with Haiku 5.5 first; OpenRouter quotes about $0.015 per review.
  6. Study the Super-agent pattern. Read the dsh-superagent-prototype README. Run it against a throwaway repo with a narrowly scoped token. Copy the event-triggered router/worker split into your own harness.
  7. Make one agent loop durable. If you use OpenAI Agents SDK, ADK, LangGraph or the AI SDK, wrap one workflow with temporalio/ai-integrations, then kill the worker mid-run and confirm it resumes.
  8. Talk to your agent. Install Hex (macOS stable; Windows/Linux beta) for on-device dictation into your terminal or IDE.
  9. Blackwell box? Serve DeepSeek-V4.1-Flash-1017-NVFP4 with vLLM that includes PR #61530, and compare tok/s and accuracy on your own evals against the BF16/FP8 source.
  10. Still open move off Opus 4.8 fast mode before Nov 30 (Claude Code v2.1.316 notes).
  11. Still open export from Jules before Nov 11 (it becomes read-only after that).
  12. Still open if you stored AI provider keys on any platform named in Wiz's Black October report, finish rotating them and check provider billing.

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 home Miss Changelog has nothing new since 10-30
AA articles Miss No new articles since 10-28; re-read the Kimi K3.1 Pro article for weights status
AA Coding Agent Index Hit Re-materialized 10-30 13:04 CDT; all 34 rows identical to yesterday
OpenRouter rankings Hit Apps through Oct 31 (omp, DeepSeek Harness, Strix new); weekly model tokens; benchmarks tab
OpenRouter models API Partial No new IDs; list prices changed on several models; Poolside laguna-s-2.1/xs-2.1 removed
OpenRouter blog Hit Code-reviewer tutorial (11-01), Claude Code session costs (10-30)
HN front page Hit Memdoor, OpenAI usage merge, DeepSeek super-agent, OpenIDEA, Nolan Lawson framework cost
HN Algolia last 24h (AI/LLM/agent keywords, one query per keyword) Hit Anthropic training-use block (X only), Talos report, Sol/Factorio anecdote, Brazilian court prompt injection
HN Algolia inference slice (vllm, sglang, trt-llm, tensorrt-llm, mooncake, "kv cache", inference) Partial Only arXiv KV-cache papers (≤2 points)
GitHub trending daily Hit hex, FreeToken new; totals checked against GitHub API
HF Daily Papers 2026-11-01 / 2026-10-31 Miss Weekend; API returned empty lists
HF models trending Partial No major release; NVIDIA NVFP4 DeepSeek, unsloth DeepSeek GGUF, LiquidAI d1-2B, SenseNova-Loop-9B-Preview (10-28)
Price Per Token news + /model-releases Partial Model releases last 24h: 0
LearnPrompt/ai-news-radar live site / data JSON (data/latest-24h.json) Miss / Hit The live SPA rendered empty; the JSON was generated 11-01 05:50 CST with 145 items
RadarAI Hit DeepSeek Super Agent, OpenRouter Responses API tutorial, VIDRAFT POCKET (no primary)
The Agentic Digest Hit Sunday issue (Mistral Large 4, carryover); 10-31 issue: Temporal GA, Slack Code API, Vercel Ling, Gemini Computer Use GA
Simon Willison 2026-10-31 Hit datasette-llm 0.1a9; ChatGPT account-takeover write-up link
X Not checked Quiet day; nothing nominated that needed X

Dailies in rank order

  1. Latent Space AINews: Partial. The latest is still the Oct 31 Google AI Mode / Search revenue issue. It is market narrative, and none of it was printed.
  2. AlphaSignal: Partial. Some items are Pro-gated.
    • Nominated NVIDIA DeepSeek NVFP4 (confirmed).
    • Also nominated Sharp-Spark-X2.5-4B-GGUF (a community repack that needs a llama.cpp PR), a Qwen3.5-4B uncensored fine-tune, LFM3 and Chronos-Bolt Mini. None of those were printed.
  3. Code4U AI Digest: Partial. The 10-31 items were Deepgram/Daily (confirmed), the robots.txt AI-training opt-out (policy), a NY chatbot law, funding news, and Mistral (carryover).
  4. TLDR AI archives: Miss. The latest issue is 2026-10-30, and there is no weekend issue.
  5. TLDR flagship: Not checked. It is only used if TLDR AI is blocked, and it wasn't.
  6. The Neuron: Partial. The latest is the Mistral 1T open model piece, a carryover.
  7. AI Daily Brief archive: Partial. The latest is 2026-10-30. Nothing new.

Confirm opens: nominated → confirmed

Dropped

  • "Anthropic blocks training-related use" (HN, 399 points): it rests on an X post only, with no Anthropic primary.
  • "Anthropic blocking game reverse-engineering" (HN via Mashable and Tom's Hardware): no Anthropic primary.
  • Anthropic watermark loophole (LearnPrompt via the-decoder) and Claude Duet (engadget): no primary, and consumer-facing.
  • OpenAI to train on Daybreak Blue data (LearnPrompt): policy, scored below 5.
  • "AI startup hacks Anthropic" (LearnPrompt via India Today): dated 10-26, stale, and no primary.
  • VIDRAFT POCKET LLMs (RadarAI): no primary.
  • agentmash (LearnPrompt via Product Hunt): not opened, scored below 5.
  • Nace AI Drex Max v1 (LearnPrompt via MarkTechPost): the HF repo was created 2026-10-13. A stale recap, scored below 5.
  • GPT-5.6 Sol beats Factorio (HN): a single anecdotal run.
  • Brazilian court prompt injection (HN): legal news, not a builder action.
  • bootgpt (Show HN), Kronos (trending; last push April), AiSOC (trending): not opened in depth, or scored below 5.
  • A trending jailbreak collection repo: skipped and not linked.
  • Sharp-Spark-X2.5-4B-GGUF and Qwen3.5-4B Heretic (AlphaSignal): a community repack and an uncensored fine-tune, scored below 5.
  • Latent Space "AI Mode hurts Search revenue": market narrative.

Engine blogs: vLLM, SGLang, TRT-LLM and InferenceX were not nominated. The only engine dependency today is the vLLM PR that the NVIDIA NVFP4 model card cites (section 2).