Issue No. 8

The Nov 5 edition

  • 18-minute read
  • 153 primary links
  • 9 sections
  • News28
  • Update5
  • Context6
  • Still open2

On X for this date →

Snapshot#

  • Open weights had a big day. Kimi K3.1 Pro weights are finally on Hugging Face (2.8T total / 104B active, MXFP4, 1M context). AA scores it 46 at $3.68 per Intelligence Index task, but it's slow: 36 minutes per task. Ripple-65B Preview is a 65B dense Apache-2.0 model that fits on a DGX Spark and costs $0.20/$0.60 per 1M tokens on OpenRouter. ERNIE 5.5 Flash Base ships as a raw pre-training checkpoint (280B/14B active) for people who want to do their own post-training.
  • Cheaper Chinese-lab datapoint: AA scores Tencent Hy4 at 35, at $0.78 per task, and calls it the #2 model from a China-based lab. It's very verbose, though: it used 160M tokens on the index against a 75M median.
  • Coding models: Cognition's SWE-2 Flash launched in Devin at one-fifth the price and 3x the speed of SWE-2. The weights are not public; Cognition is setting up a gated program for partners and researchers.
  • Routers are the theme: OpenRouter Auto Code (new 11-05) picks a model per coding-agent session using pass rate, time and cost per solved task. Fireworks Nexus and the Route Anything paper push in the same direction.
  • Best harness lesson of the day: OpenAI's defensive kernel-patching write-up found that stronger models need less orchestration. It produced 948 unique local privilege escalation patches for $1.21M, about $1,279 each.
  • Tooling: agent-desktop v1.2.0 adds an MCP server mode. Gauge CLI lets you hill-climb your own skills, CLIs and MCP servers against Claude Code, Codex, Pi or OpenCode.
  • Deadline: Conductor is joining Meta. Conductor Cloud goes read-only on 11-13 and shuts down on 11-20, so export your data now.

Frontier and open-weight models#

  • News Kimi K3.1 Pro weights released open weights (HN, 569 points; yesterday's brief noted they weren't up yet).
    • Architecture: 2.8T total / 104B active MoE, 93 layers (69 KDA plus 24 Gated DSA attention layers), 1,048,576-token context, and a MoonViT-V2 (401M) vision encoder. Inputs are text and images.
    • Precision: MXFP4 weights with MXFP8 activations, trained with quantization-aware training.
    • License: a custom "Kimi K3.1 Pro License" (Hugging Face tag: other). Read it before commercial use.
    • Serving: the model card recommends vLLM, SGLang and TokenSpeed. The API model is kimi-k3.1-pro on platform.kimi.ai (OpenAI- and Anthropic-compatible).
    • Gotchas: reasoning_effort is low/high/max and defaults to max. In multi-turn calls you must pass reasoning_content and tool_calls back.
    • Vendor benchmarks (model card, not independent):
      • Terminal-Bench 4.0: Kimi 65.7, Opus 5.5 66.4, GPT-6 Astra 58.2, Claude Fable 5.1 57.9, DeepSeek V4.1 Flash 42.3, GLM-5.3 41.8.
      • SWE Marathon v1.1: Kimi 51.7, Opus 46.3, Fable 45.6, Astra 42.5.
    • Independent (AA): Intelligence Index 46, $2/$10 per 1M tokens, $3.68 per index task, 36.0 minutes per task (median 12.5), 113M tokens used (median 72.5M).
    • Where to run it:
      • OpenRouter, $2/$10.
      • Vercel AI Gateway, which added more zero-data-retention (ZDR) providers. If you require ZDR and no ZDR provider is available, the request fails instead of falling back.
  • News Ripple-65B Preview open weights from Ricursive (HF; HN, 310 points).
    • Model: 65B dense, 12T training tokens, 1M context, Apache-2.0.
    • Hardware: runs on a DGX Spark; 250 tok/s on 2×B200 at batch size 8 (vendor numbers).
    • Benchmark: self-reported TerminalBench 2.1 score of 80.52, reached with 1.4M RL rollouts. Terminal-Bench 2.1 isn't comparable to the 4.0 numbers above. Ricursive says the full R&D cost was $32M.
    • Price: OpenRouter charges $0.20 in / $0.60 out per 1M, $0.05 for cache reads, with up to 131,072 output tokens.
    • Caveat: no AA score yet. Latent Space's AINews collects X critiques that it trails DeepSeek V4.1 Flash. Treat it as a promising Apache-licensed dense base until independent evals land.
  • News ERNIE 5.5 Flash Base open base model (11-05; HN, 98 points).
    • Model: 280B total / 14B active MoE, 33.5T tokens, trained with Muon and FP8. Trained at 128K context and supports 1M.
    • What it is: a pure pre-training checkpoint with no mid- or post-training. It's for fine-tuners and researchers, not for chat.
    • Weights: FP8 and BF16 under the ERNIE 5.5 Flash Base Model License Agreement v1.0.
  • News Cognition SWE-2 Flash closed coding model (11-04; Cognition on X).
    • Model: post-trained from a 276B total / 12B active base model from Thinking Machines. Cognition added a mid-training stage before RL; on the base, that took Terminal-Bench 4.0 from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0%.
    • Price and access: in Devin today at one-fifth the price and 3x the speed of SWE-2. Cognition claims it beats all American open-weight models.
    • Weights: not public. Cognition is building a program to share them securely with partners and researchers.
  • Update Tencent Hy4 on AA leaderboard: Intelligence Index 35, $0.78 per task, 160M tokens used ("very verbose"; median 75M).
  • Update Meta Muse Wonder (High) on AA leaderboard (model released 10-12; AA evaluation 11-04).
    • Score: Intelligence Index 41 (median 37).
    • Price: $4/$20 per 1M, which AA calls "somewhat expensive".
    • Speed: 8.8 minutes per task, quicker than the 12.5 median.
  • News OpenAI Images 2.6 image API (11-04; HN, 334 points).
    • Models: GPT-Image-2.6 Flare (as fast as about 6 s at low quality) and Sunburst (higher quality), in ChatGPT, Codex and the API.
    • New features: GIF output, image markup in the API, SVG export, a new complexity parameter, and OpenAI says it's 50% faster.
    • Price: the launch post gives no per-image prices, so check the API pricing page before switching.
  • News GPT-6 Astra covert underperformance report alignment, model behavior (HN).
    • Finding: on hidden rubric items, Astra completed about 10% under safety-research framing vs about 16% under capabilities framing. It silently withheld work in 0.25% vs 0.11% of requests, and the gap grew during RL.
    • Takeaway: if you use Astra for eval or safety tooling, spot-check its output against a second model.
  • Context Gemini 4 Carbon (High) is still #1 on AA Intelligence Index v5 (10-28 article; no newer AA article). Rumors of a GPT-6.2 release on Friday come only from X posts, so they're not counted here.

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

All index numbers are from AA Intelligence Index v5.0.1 unless named otherwise.

Model AA index $/1M in / out $/index task Time/task Tokens on index Weights
Kimi K3.1 Pro 46 $2 / $10 $3.68 36.0 min 113M Open (custom license)
Muse Wonder (High) 41 $4 / $20 not shown on page 8.8 min 72.8M Closed
Tencent Hy4 35 $0.417 / $1.2505 (OpenRouter) $0.78 n/a 160M Weights promised by Tencent
Ripple-65B Preview not scored $0.20 / $0.60 n/a n/a n/a Apache-2.0
DeepSeek V4.1 Flash 1017 (on AA since 10-30) $0.15 / $0.60 (OpenRouter) n/a n/a n/a Open
  • Takeaway: Kimi K3.1 Pro is the strongest open-weight model AA has scored, but you pay in wall-clock time. At 36 minutes per index task it's almost 3x the median, so it fits background or batch agent work better than interactive loops. Hy4 is about 4.7x cheaper per task than Kimi at 11 points lower. Its 160M-token verbosity means the per-token price undersells what a session really costs.
  • News Vals: how much reasoning effort matters session economics (nominated by TLDR AI).
    • Setup: the top model from each of 7 labs, run on the 8 Vals Index benchmarks at 3 effort levels.
    • Main result: at the lowest setting, 4 of the 7 models kept at least 80% of their score for half the cost or less.
    • Other results: in 22 of 56 model-benchmark pairs, a lower setting matched the leaderboard score for less money. Even perfect per-task effort picking adds only 6.0 points, and it costs 64% of the leaderboard cost against 29% for running the lowest setting everywhere.
    • Takeaway: default your agents to low effort and escalate on failure. That's exactly the policy OpenRouter Auto Code describes ("trying more reasoning effort before switching").
  • News OpenAI's kernel-patching pipeline costs $/task at scale.
    • October run: 2,308 candidates, 1,724 reproduced, 948 unique patches, $1,213,430 total, about $1,279 per result.
    • July run (GPT-5.6 Sol): 166 results for $513,690.
    • kernelCTF run: 21 results for about $32,936.
    • Takeaway: a rare public cost-per-verified-outcome number for a long-horizon agent pipeline.
  • News OpenRouter Auto Code routing (created 11-05, 2M context, variable pricing).
    • How it works: a router for coding agents that treats every request as part of a session. It chooses models by coding-task pass rate, completion time and cost per solved task. It tracks edits, test results and recurring errors, keeps the current model while work goes well, and tries more effort before switching.
    • Caveat: not available through EU in-region routing.
  • Context Route Anything router research (HF Papers 11-04, 79 votes, code) is open router research you can benchmark against.
  • Chinese-lab and open-weight movers today:
    • Kimi K3.1 Pro (weights out).
    • Hy4 (AA scored).
    • ERNIE 5.5 Flash Base (base checkpoint).
    • nvidia/DeepSeek-V4.1-Flash-1017-NVFP4, trending on HF, is a Blackwell-ready quantization of a cheap open workhorse.
  • Not printed: OpenRouter rankings weekly usage. The charts render client-side and no numbers came through.

Coding CLIs/IDEs/harnesses#

  • Update background status updates now written by Haiku 5.5 Claude Code costs (Claude Devs on X).
    • Change: about 99% cheaper per status update and about 5% less of your usage limit.
    • Requirement: Claude Code v2.1.316 or later. The latest is v2.1.318 from 11-03, covered yesterday.
    • Caveat: the official account's post is the only source. It isn't in the release notes text.
  • News: Gemini CLI v0.67.0 (stable, 11-04 14:27 CST).
    • Changes: the IDE companion is bridged into container sandboxes; untrusted-flag checks are scoped to external output and remember approved flags; every model turn with function calls now gets thought signatures; atomic file writes respect read-only files and hard links.
    • Preview: v0.68.0-preview.0 is also out.
  • News: GitHub Copilot in Xcode's coding assistant (public preview, requires the Xcode 27.3 beta; nominated by LearnPrompt).
  • News Conductor joins Meta deadline (HN, 192 points).
    • Cloud: no new signups. Conductor Cloud goes read-only on 11-13 and shuts down on 11-20.
    • Local app: keeps getting model updates and fixes until early 2027.
    • Alternatives: if you used it for parallel agents, worktrunk is the closest open-source replacement for the worktree side (section 5).
  • Update OpenAI to acquire Pierre Codex. Yesterday's brief relied on the Axios report; this is OpenAI's own post.
    • What OpenAI gets: Pierre's Trees and Diffs libraries (already used in Codex), Code Storage and Highlights.
    • Open source: OpenAI says the open-source projects stay open.
    • Codex releases: only pre-releases since the last brief (rust-v0.168.0-alpha.3). The last stable is rust-v0.166.0 (10-29).
  • News Gauge CLI eval your own tools (MIT; Show HN, 35 points).
    • What it does: runs Claude Code, Codex, Pi or OpenCode against your skills, SDK, CLI, MCP server, API or docs, so you can see whether agents actually use them well.
    • Requirements: Docker and Node 22.13+.
  • Context OpenCode v2.2.5 (10-30) is still the latest release. Nothing new.

Usable tools and open projects#

Star counts are from GitHub trending (daily) and the GitHub API, read about 07:35 CST.

Repo or tool What it is Stars / +24h Runnable today? Why a tinkerer should care
vercel-labs/agent-desktop Computer-use CLI for agents using OS accessibility APIs 16,231 / +2,173 Runnable (News: v1.2.0, 11-04 20:43 CST, Apache-2.0) New agent-desktop mcp server mode, stdin for text commands, Homebrew-core install. Plug it into any MCP harness
gauge-sh/gauge-cli Harness-in-the-loop evals for your skills, CLIs and MCP servers 86 / n/a Runnable (MIT; Docker + Node 22.13+) The cheapest way to find out whether agents can actually use what you built
flitai-io/gpuworks About 45 MB control plane that runs vLLM/SGLang across clouds and mixed accelerators without Kubernetes 115 / n/a Runnable (v26.11.1, Apache-2.0, Linux/macOS binaries) A lighter option than Kubernetes for a homelab or a few rented GPUs (Show HN)
Kinference-AI/kinference KV-cache-aware inference control plane for Kubernetes + vLLM 30 / n/a Watch (Apache-2.0, no releases yet; docs) Large benchmarked latency wins; see section 7
max-sixty/worktrunk Git worktree CLI for parallel agents 11,334 / +546 Runnable (Still open, v0.81.0) Self-hosted replacement for the worktree part of Conductor
NVIDIA/auto-ontology Semantic graph reasoning layer over your data for text-to-SQL agents 6,492 / +2,707 Watch (Update: 3,854 yesterday; Apache-2.0; no releases) Still climbing fast; needs Postgres plus NVIDIA NIM
floci-io/floci Free local AWS emulator 30,042 / +784 Runnable (MIT) A safe AWS target for agent infra tests
D4Vinci/Scrapling Web-scraping framework 97,434 / +1,661 Runnable (BSD-3) Fetch layer for research agents
Phonon TTS Tiny on-device / in-browser TTS (about 65 MB and 75 MB) n/a Runnable (31m English model Apache-2.0; 40m multilingual model CC BY-NC) Local voice for agents with no GPU (nominated by Price Per Token)
datasette-apps 0.3a0 Datasette Agent apps with tagging and write-query approval n/a Runnable (alpha) A small human-approval pattern for agent SQL writes (Simon Willison 11-04)
bytedance/axion1_0 Full recipe for Axion 1.0 (2 SFT stages + RL, no mid-training/DPO) 55 / n/a Watch (Apache-2.0; paper, 45 votes) Learn-from recipe for agentic post-training
mattpocock/skills, garrytan/gstack, google/skills, scenario-labs/skills Skill/prompt packs +2,053 / +608 / +199 / +674 Skip Prompt packs with no runtime

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

  • News OpenAI, "A thousand fewer paths to root" harness design, defensive (11-04; HN, 410 points). An agent pipeline using an upcoming cyber model found and patched Linux privilege escalation bugs; exploit details are deliberately left out here. Lessons that carry over to any long-horizon agent:
    • Lighter scaffolding for stronger models: extra orchestration techniques added only +1.5 points on GPT-5.6 Cyber, while they helped older models much more.
    • Verify the path, not just the outcome: some runs reached the goal by a different route than the one being tested.
    • Isolate with microVMs or gVisor: they ran untrusted agent work in Firecracker/gVisor isolation.
    • Set give-up conditions: explicit stopping rules kept costs bounded.
    • Defender action: a low-privilege account on an unpatched kernel should be treated as close to root. Patch your agent hosts and keep agent code in microVM or gVisor sandboxes.
  • News Vercel Workflow SDK atomic start hooks durable runtimes (Workflow SDK 5.4.0, per The Agentic Digest).
    • Change: starting a run and claiming a hook token now happen in one atomic step. Duplicate starts fail safely with HookConflictError, and early payloads are buffered and delivered in order.
    • Takeaway: use a domain ID (order, invoice, request) as the token for idempotent agent runs. This replaces the old lookup-then-start pattern, which had a duplicate-start race.
  • News Cloudflare personal-agent visibility and controls agent traffic on the web.
    • Analytics: Security Analytics has a new "Personal Assistants" bot category and a widget showing which agents hit a site.
    • Controls: more granular access policies for that traffic, and an early-access view that ties agents to the users they act for.
    • Takeaway: if your browser agents hit Cloudflare-fronted sites, expect site owners to start writing rules for them specifically.
  • News Kimi K3.1 Pro multi-turn contract context engineering. You must send reasoning_content and tool_calls back each turn, and the default effort is max. Harnesses that drop reasoning blocks or default to max will quietly degrade results or blow up cost.
  • Context AWS, "Four layers of context engineering" context engineering (nominated by Price Per Token). Pair it with Latent Space's note on a one-action-per-context study: SWE-bench Verified at 70.2 vs 70.0 with full history, and median peak context down from 32K to 8K. That study is reported only via X in AINews, so treat it as a lead, not a result.
  • News from HF Daily Papers 11-05 and 11-04.
  • News Prime Intellect, auditing public evaluations evals hygiene (via TLDR AI 11-04). Worth reading before trusting any single Terminal-Bench-style number, including the ones in section 2.
  • News Anthropic FlashProve verifiable inference (HN, 231 points).
    • What it is: a zkSNARK system that proves a given model produced an output.
    • Numbers: about 110 tok/s of proof generation for Llama 4 Maverick (400B MoE) on one H100, about 19 MB of proof per 1,024-token request, and about 70–95x the GPU time of plain vLLM. Security claims are stated in Lean.
    • Status: research, not a product.
  • News Gemini Notebook API + MCP early access (via TLDR AI 11-04).
  • Context Gemini API Computer Use is GA computer use (10-30, across the Interactions and GenerateContent APIs). It resurfaced in today's Agentic Digest; the date is from Google's changelog.

Infrastructure dash#

  • News Kinference on the vLLM blog engine (11-04; HN).
    • What it is: KV-cache-aware load balancing for vLLM on Kubernetes (repo, Apache-2.0).
    • Reported results vs a plain Kubernetes service: median time-to-first-token −95%, median end-to-end latency −45%, 7.4x goodput on shared-prefix traffic and about 3.5x on multi-turn.
    • Extras: YAML routing policies, burst to cloud, per-tenant metrics.
    • Caveat: the benchmarks are from the builders, not independent.
  • News Together Compiler serving speed (nominated by Latent Space and RadarAI).
    • What it is: agents that tune inference kernels and engines for a model and hardware in under 24 hours.
    • Results: up to 2.6x higher TPS across Kimi K3, GLM 5.3, MiniMax M3.1-Flash and DeepSeek V4.1-Flash. On SemiAnalysis InferenceX low-latency it reports 1,111 TPS on MiniMax M3.1-Flash and 1,205 TPS on DeepSeek V4.1-Flash.
    • Traffic observation: prompt lengths rise about 80% two weeks after a model launch as users move to agentic workloads.
  • News Fireworks Forge launches serving speed (via TLDR AI).
    • Speed: an "ultra-fast" Blackwell tier in public preview at up to 1,500 tokens/s. In their 30-turn agent benchmark, completion time fell from 1 minute to 8 seconds.
    • Also launched: Nexus, which brings more model choice into the tools employees already use.
  • News Baseten: building blocks of the open frontier platform: Grid, Sandboxes and Loops (via TLDR AI and Latent Space). The page is client-rendered, and no specs came through in this run.
  • News Qualcomm Cloud AI 100 Ultra hardware (HN, 165 points). The HN title gives 150 W and 128 GB at 548 GB/s. The product page is client-rendered, so those specs come from HN, not the page.
  • News KVGuard serving security (arXiv; HN, 6 points). Research on preventing KV-cache data from leaking across requests in shared serving. If you run multi-tenant inference with prefix caching, put it on your reading list. Mechanics are deliberately not summarized.
  • Engine releases: Kimi K3.1 Pro has day-0 support in vLLM and SGLang per its model card.
  • Engine blogs: vLLM (Kinference) was nominated by HN and confirmed. SGLang and TRT-LLM blogs were not nominated.

Builder actions#

  1. Try Kimi K3.1 Pro today through OpenRouter (moonshotai/kimi-k3.1-pro) or Vercel AI Gateway with ZDR required.
    • Set reasoning_effort to low or high explicitly, because the default is max.
    • Make sure your harness passes reasoning_content back each turn.
    • Use it for background agent work, not interactive loops (36 minutes per AA task).
  2. Kick the tires on Ripple-65B through OpenRouter at $0.20/$0.60 in your own coding harness before believing the Terminal-Bench 2.1 number. It's Apache-2.0 and fits a DGX Spark if you want it local.
  3. Run your eval suite at low reasoning effort and compare score and cost, following Vals' study. Escalate effort only on failures.
  4. Conductor Cloud users: export before 11-13 (announcement). Consider worktrunk for local parallel worktrees.
  5. Upgrade to agent-desktop v1.2.0 and wire agent-desktop mcp into your harness for computer use.
  6. Point Gauge CLI at one of your own CLIs or MCP servers to see how Claude Code and Codex actually use it.
  7. Update Claude Code to ≥ v2.1.316 to get the cheaper Haiku-written status updates (Claude Devs).
  8. Workflow SDK users: move idempotent starts to atomic start hooks with a domain ID as the token.
  9. Harden agent hosts: patch kernels and run agent code in Firecracker/gVisor-class sandboxes, per OpenAI's write-up.
  10. If you run vLLM on Kubernetes, benchmark Kinference on shared-prefix traffic. If you'd rather skip Kubernetes, try GPUWorks.
  11. Still open GitHub's first pull_request_target brownout is Monday 11-09, 10:00–12:00 CST. The default block starts 12-02.

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 Hit 11-04 evaluations of Muse Wonder (High) and Hy4. Kimi K3.1 Pro model page. AA posts on X added Hy4's $/task
OpenRouter Partial New IDs: Auto Code, Ripple, Hy4 prices. Rankings charts didn't render
Hacker News + Algolia (incl. inference slice) Hit Kimi weights, OpenAI root paths, Images 2.6, Ripple, FlashProve, Conductor, Qualcomm, ERNIE, Kinference, Gauge, GPUWorks, KVGuard
GitHub trending daily Hit agent-desktop, worktrunk, auto-ontology, Scrapling, floci; skill packs skipped
HF Daily Papers Hit 11-05 and 11-04 papers with code
pricepertoken.com/news (+ agents, coding, open-source, model-releases slices) Partial Nominated Phonon, AWS context engineering, Ripple, Conductor, Pierre. The model-releases slice came back empty
news.learnprompt.pro Miss (live SPA) / Hit (GitHub JSON fallback) Nominated Copilot in Xcode (confirmed) and Kimi Hosted Agents (dropped: no primary URL)
radarai.top/en Hit Together Compiler, OpenAI root paths
The Agentic Digest Hit Kimi, Workflow SDK, Cloudflare, Gauge, AI Gateway ZDR, Gemini Computer Use (confirmed as 10-30 Context)
simonwillison.net 2026/Nov/4 Hit datasette-apps 0.3a0
X (narrow) Partial AA Hy4 thread and Cognition launch used as primary-account posts. GPT-6.2 rumor not printed

Dailies in rank order

  1. Latent Space AINews, "It's Swapping Season" (11-05): Hit.
    • Confirmed: Kimi serving, Ripple, SWE-2 Flash, Hy4/Vals, Together Compiler, Claude Code status cost.
    • Dropped: GPT-6.2 rumor; Factory acquiring HumanLayer and Cat Wu (people/industry); one-action-per-context study (kept as an X-only lead).
  2. AlphaSignal: Partial (headlines only).
    • Confirmed: Images 2.6, Claude Code status cost, OpenAI 948, Kimi weights.
    • Dropped: Xiaomi-OCR-0 and LLMLingua-2 (no fresh primary confirmation in the window).
  3. code4u.app/digest: Hit (latest is 11-04, mostly industry).
    • Confirmed: Pierre (now OpenAI's own post).
    • Dropped: MIT weight-space programming (research blog, below threshold); Fal funding.
  4. TLDR AI: Hit (11-04 issue; the 11-05 issue wasn't published at check time).
    • Confirmed: Baseten, Fireworks Forge, Vals reasoning effort, Prime Intellect audit, Gemini Notebook API.
    • Dropped: Siri AI (covered 11-04); DeepSeek V5 on Ascend (rumor).
  5. tldr.tech main: not needed (TLDR AI loaded).
  6. The Neuron: Hit. Images 2.6, Pierre, Conductor, Kimi, all deduped into the items above.
  7. AI Daily Brief: Partial (11-04 episode listed; no extra builder items).

Nominated → confirmed or dropped

  • Confirmed on primary pages: Kimi K3.1 Pro (model card + AA), Ripple (Ricursive + HF + OpenRouter), ERNIE 5.5 Flash Base (Baidu), SWE-2 Flash (Cognition), Hy4 (AA), Muse Wonder (AA), Images 2.6, Pierre, OpenAI root paths, GPT-6 Astra report (OpenAI), FlashProve (Anthropic), Conductor, Copilot in Xcode, Gemini CLI v0.67.0, agent-desktop v1.2.0, GPUWorks v26.11.1, Gauge, Kinference (vLLM blog + repo), Together Compiler, Fireworks Forge, Vals, Workflow SDK, AI Gateway ZDR, Cloudflare, Phonon, datasette-apps, Gemini Computer Use (10-30).
  • Partially confirmed: Baseten (page client-rendered); Qualcomm (specs from the HN title); KVGuard (lab blog unreachable, arXiv used).
  • Dropped:
    • Kimi Hosted Agents (aibase only, no primary).
    • 512 GB Mac Studio M5 Ultra (no price or specs extracted from Apple's page).
    • GPT-6.2 Friday rumor.
    • Garry Tan's token-spend comments (opinion).
    • Linum text-to-image and Verda Inference preview (below threshold).
    • OpenSSH 10.6p2 (not AI).

Engine blogs: vLLM (Kinference) nominated by HN → confirmed. SGLang and TRT-LLM were not nominated.