Snapshot#
Decision-model day. Google shipped its first closed-set scoring model, and Artificial Analysis had it on the Decision Index within hours: mid-pack on score, pricier per task than the leaders. AWS shipped a local AWS emulator written entirely by agents, which is the most useful testing tool for coding agents this week. No new frontier generative model shipped; the frontier numbers that moved were new evals on Fable 5.2.
- Gemini Decision 1 / Decisions API (Google, 2026-11-11): yes/no, choice (2–255 options) and rubric scoring, up to 1,000 questions per call over text, image, audio, video and PDF, 256K context, $0.10 per 1M input tokens, output not billed, 150–250 ms end to end (Google's figures).
- AA Decision Index v2.1: Gemini Decision 1 scores 66.0 at $345 per 1M tasks and 211 ms per task. Leaders Drex Max v1.1, SubQ Snap 1.2 and GLiDE v1.2 score about 72.5–72.6; Drex Max costs $103 per 1M tasks.
- AWS Emulator (AWS, CC0-1.0, HN 895 points): a local, offline emulator of the AWS APIs for tests and CI, no AWS account, with an agent skill. AWS says all of its code was written and tested by agents, built from 400+ API models covering 200+ services.
- Antigravity browser use and Managed Agents self-hosted environments (Google): the agent can drive a browser with download approval, and Gemini Managed Agents can now run tools in your own sandbox. Fly.io Sprites and Agent Substrate had same-day support.
- A2UI 1.0 (Google): the first stable version of the open generative-UI protocol, with renderer SDKs for web, Android and iOS.
- Atuin Threads (Atuin 18.24, HN 226): synced, searchable agent sessions across machines, resumable or forkable into another harness.
- MACaQ (Huawei, Apache-2.0, HN 121): a calibrated quantizer that writes standard 2/3/4-bit GGUF for llama.cpp and also targets vLLM and Diffusers.
- Fable 5.2 on Agent Arena: #1 at +30.4% net improvement across 4.7K sessions, median $7.3 per task; Gemini 4 Carbon (High) is #3 at $1.5 per task.
Do this today: install the AWS Emulator skill (aws agent-toolkit add-skill --skill-name aws-emulator --region us-east-1) and point one agent's integration tests at it instead of a real account.
Frontier and open-weight models#
No frontier generative model launched in the window. The closed model that shipped is a scoring model, and the open-weight movement was in quantization and day-zero serving.
- News Gemini Decision 1 closed, decision model (
gemini-decision-1, Google DeepMind, 2026-11-11; HN 478; Logan Kilpatrick).- What: a model that scores a fixed set of options in one forward pass instead of generating text. Three question types:
noul(yes/no probability),choice(2 to 255 options, normalized probabilities) andscore(ordered rubric, weighted score plus per-level probabilities). Every answer carries a 0–1 confidence. - Limits: one shared multimodal input up to 256K tokens, up to 1,000 independent questions per request. Google says model compute is a median 50–100 ms and 150–250 ms end to end, nearly flat from 1 to 100 questions, measured on contexts up to 16K tokens.
- Cost: $0.10 per 1M input tokens; output tokens are not billed. Google says this is nearly 5x cheaper than Gemini 3.5 Flash-Lite across its shared benchmark suite. That is a vendor comparison.
- Serving path: Gemini API (
client.decisions.create,v1beta) in Google AI Studio. Input uses the sameContentblocks as the Interactions API. LangChain shipped day-zero support. - Internal uses Google names: query routing to model tier and thinking level in Gemini Notebook (up to 40% lower latency, per Google), and the Stitch rapid-canvas experiment.
- Independent numbers (AA Decision Index, AA post): see §3. AA's summary: not yet frontier on its index, leads on knowledge (AA-Omniscience-Decision 66) and reasoning (HLE-MultiChoice 49), trails on document extraction (AA-Paperclip: all 15 answers right on 33% of documents, against 63% for Drex Max v1.1).
- Builder take: this is a routing and grading primitive, not a chat model. Simon Willison notes the API is close to TypeSafe's Jev and priced like OpenAI's GPT-6 Luna decision model, and that Gemini returns softer probabilities on yes/no than Jev's 1.0/0.0. Calibrate your thresholds on it rather than reusing ones tuned for another vendor.
- What: a model that scores a fixed set of options in one forward pass instead of generating text. Three question types:
- Update Claude Fable 5.2. No new release, but three new independent datapoints:
- Agent Arena (2026-11-11 11:23 CST): Fable 5.2 (Max) #1 at +30.4% net improvement over 4.7K real agent sessions, about 1.6x #2 Opus 5.5 (Max, +19.5%). #1 in Code, Work and Chat.
- AA-Robotics v0.1: Fable 5.2 leads at 75, ahead of GPT-6 Astra (69) and GPT-6.1 Sol (65).
- VIALS-AA: 67% with tools at $4.20 per task, 44% without tools at $0.73. Haiku 5.5 (max) with tools scores 49% at $0.07 (§3).
- Update Unsloth Dynamic v3.1 open weights, quantization (2026-11-11). Qwen3.8-27B GGUF (Apache-2.0) was re-uploaded with the v3.1 algorithm; the docs publish KL-divergence comparisons against v3. Unsloth's accuracy claims are its own measurements. Serve with llama.cpp or any GGUF runner.
- News Praxis-1 open weights, robotics, outside core scope (Runway, Apache 2.0; base weights). A 27B world action model for robot control. vLLM-Omni has day-zero support qualified on one H100 80GB, and NVIDIA published an NVFP4 checkpoint. Runway's leaderboard and speed claims are its own.
- Still open ERNIE 5.5 Flash CN lab, weights promised. Baidu's Apache 2.0 weights are targeted for 11-25. A llama.cpp draft PR #31925 ("model : add ERNIE 5.5") is open. The preview stays free on OpenRouter.
- Still open Qwen4-Max preview CN lab, thin. Still no model card, price or weights. It has no entry in Artificial Analysis's page data.
- Context AA published no new Intelligence Index entries in the window. Yesterday's top of the board stands: Fable 5.2 (Max) 64.6, Gemini 4 Carbon (High) 60.0, Opus 5.5 (Max) 59.1 (OpenRouter benchmarks panel, AA Intelligence Index, read 11-11). OpenRouter listed no new model IDs in the window (models API).
- Not printed: Grok 4.8 (The Neuron headline; no x.ai post found), AlphaGenome 2 and Gemini Robotics 2.5 (outside scope), MiniMax H3-Image (image model, trending since 11-03).
Pareto frontier (quality vs $/token vs $/task vs session economics)#
The new Pareto line today is for decision models, where list price per token and cost per task rank differently.
Decision models: AA Decision Index v2.1 (read from the page data; cost is per 1M tasks at current token prices, decision models billed for input only):
| Model | Index | $ per 1M tasks | Time per task |
|---|---|---|---|
| Drex Max v1.1 (Nace.AI) | 72.6 | $103 | 283 ms |
| SubQ Snap 1.2 (Subquadratic) | 72.6 | $137 | 162 ms |
| GLiDE v1.2 (Fastino) | 72.5 | $780 | 284 ms |
| Mammoth 27B (Base Labs) | 71.4 | $264 | 106 ms |
| d1-ultra (Liquid AI) | 70.6 | $133 | 301 ms |
| Decider v1.2 (Perplexity) | 69.8 | $68 | 229 ms |
| Elephant 9B (Base Labs) | 66.4 | $92 | 79 ms |
| Jev 1.14 (TypeSafe) | 66.2 | $156 | 92 ms |
| Gemini Decision 1 (Google) | 66.0 | $345 | 211 ms |
| Microsoft-Decision-1 | 62.1 | $143 | 274 ms |
| Celeris-1.1 Decision | 62.0 | $14 | 61 ms |
| GPT-6 Luna (Decisions) | 60.2 | $339 | 206 ms |
- News Gemini Decision 1 is about 6.6 points below the leaders at more than 3x Drex Max's cost per task. AA attributes the gap to token price ($0.10 against $0.03 per 1M input for Drex Max). Its Brier score across the index is 0.18, between Drex Max (0.16) and Jev 1.14 (0.19), per AA.
- Index disagreement: Google's own suite has Decision 1 beating Jev 1.14 on 5 of 7 shared benchmarks, 92.2% against 90.0% on JevBench (vendor numbers, via the blog). AA's index puts the two level (66.0 vs 66.2), with Jev at under half the cost per task and less than half the latency.
- Where it still wins: multimodal input. Jev is text and JSON only, and OpenAI's decision model handles text and images (Simon Willison). If your routing input is audio, video or PDF, Decision 1 is the one on this list that takes it natively.
Agents: Agent Arena (real user sessions, Arena's metric is net improvement):
| Model | Net improvement | Median $ per task |
|---|---|---|
| Claude Fable 5.2 (Max) | +30.4% | $7.3 |
| Claude Opus 5.5 (Max) | +19.5% | $5.5 |
| Gemini 4 Carbon (High) | +17.0% | $1.5 |
| GPT-6 Astra (Max) | +14.5% | $6.0 |
- Arena's post labels the #1 row "Claude Fable 5.1 (Max)" in its comparison list, but the rest of the post makes clear it means 5.2. Its Fable 5.1 → 5.2 numbers: median cost per task $7.7 → $7.3.
- On Arena, Carbon is the cost outlier: a fifth of Fable's cost per task for a bit over half its net improvement. On CursorBench yesterday it matched Fable 5.2 Max on score. The two indexes measure different things and disagree on how close Carbon is.
Effort vs tier (AA, Claude lineup): moving up a model tier beats maxing out effort. Opus 5.5 (xhigh, 57) outscores Sonnet 5.5 (max, 55) at 40% lower cost per task and about half the time. Fable 5.2 (high) beats Opus 5.5 (max), and Fable 5.2 (low) beats Opus 5.5 (high). The pattern breaks at the bottom: every step up from Haiku 5.5 (max) costs more.
Tools vs model size (AA VIALS-AA): tool use roughly doubles scores across models (about 25% to 47% on average). Haiku 5.5 (max) with tools scores 49% at $0.07 per task, above Fable 5.2 without tools (44% at $0.73).
Coding-agent number (Context, read 11-11, AA Coding Agent Index v2.0): Claude Code + Opus 5.5 (Max) 61.7, Devin Fusion CLI 61.5, GitHub Copilot Fusion (max) 59.9, Antigravity CLI + Carbon (High) 59.4. Fable 5.2 is not on it yet.
Session economics (OpenRouter top apps, "Today" panel dated 11-11): Claude Code 1.86T tokens, Hermes Agent 1.75T, Codex 830B (up from 555B in yesterday's panel and now #3 among apps), Kilo Code 762B, pi 593B, OpenClaw 488B, Strix 410B, DeepSeek Harness 338B, omp 333B. The cost-per-session panel is still client-rendered and unreadable.
Token efficiency: at AI Engineer, Cloudflare described sending tool selection and log extraction to a decision model and only the diagnosis to a reasoning model. In its demo that went from six LLM calls to one and about 40K tokens to 1.8K. Those are Cloudflare's demo numbers.
Coding CLIs/IDEs/harnesses#
A quieter release day. The notable movement was test infrastructure and a Qwen CLI 1.0 preview. Times converted from UTC to CST.
- News AWS Emulator agent test target (user guide via README, latest release v1.0.322, 2026-11-10 18:32 CST; Swami Sivasubramanian's post).
- How to run: with an agent,
aws agent-toolkit add-skill --skill-name aws-emulator --region us-east-1(AWS CLI 2.35.0+). On Linux, RPM or Debian packages. On macOS and Windows, a container from a short Dockerfile. In Rust, an embedded mode runs one isolated instance per test with no endpoint or env vars. The server listens on HTTPS at127.0.0.1:9297. - Fidelity: each release publishes per-service fidelity scores, and known gaps live in the repo under
issues/things-that-do-not-work/. AWS says not to use it for production or for security testing. - How it was built: agents wrote fidelity tests from the public Smithy API models, corrected each test until it passed against real AWS, then built the emulator until it passed the same test. A merge gate with several security-review agents checks every change. Simon Willison counts 29,684 commits in the first three weeks.
- Operator caveats (HN): commenters compare it to LocalStack, Floci and Testcontainers, and one asks whether AWS will maintain it as a product. Building from source needs at least 100 GB free for
target/. - Builder take: the useful pattern is a high-fidelity local twin as the inner loop for an agent, so it can check its own work without credentials or a bill.
- How to run: with an agent,
- News: Antigravity browser use (rolling out, @antigravity 2026-11-11 10:37 CST). The agent drives the built-in browser for research, localhost testing and checking its own work. You choose what it logs into, and downloads pause until you approve them (v2.31.0 changelog). @kevinhou22 says to use
/browserafter updating. - News: Qwen Code 1.0 Preview (2026-11-10 21:55 CST, after yesterday's brief; repo +151 stars today). It merges Qoder's agent into Qwen Code, supports Qwen and Qoder accounts plus custom compatible endpoints, and ships native packages for macOS, Linux and Windows with no separate Node.js install. It supports session resume and non-interactive runs. Preview label: interfaces may change.
- News: Codex rust-v0.167.2 (2026-11-11 16:44 CST): tool updates are now sent incrementally by default on Google Cloud Vertex AI; opt out with
incremental_tools_legacy_gcp = falseunder[features]. The 0.170.0 alphas also started on 11-11. - News: OpenAI Codex on Amazon Bedrock (AWS ML blog): an enterprise rollout guide for Codex with Bedrock as the model backend, covering access paths, identity and monitoring.
- News llm 0.37.2 security fix (2026-11-11 11:19 CST). Fixes two cases where Simon Willison's
llmCLI could send your stored OpenAI key to an OpenAI-compatible third-party endpoint configured inextra-openai-models.yaml(GHSA-v9h7-jj4g-f862, GHSA-r639-rr7r-x72w). Upgrade if you usellm openai endpoint add. - Context Copilot CLI shipped only 1.0.110 prereleases (fixes; an experimental masked dialog for git, ssh and sudo password prompts). Gemini CLI v0.69.0-preview.0 adds a
/stats heatmapfor daily token use. OpenCode v2.2.7 (11-10) adds GitHub Copilot Auto routing. Claude Code is still on v2.1.322. - Still open: Claude Code issue #107757 is open at 117 comments, last updated 2026-11-09 12:26 CST.
- Still open Conductor deadline tomorrow Cloud goes read-only 11-13 and shuts down 11-20.
Usable tools and open projects#
Star counts are from the GitHub API around 07:45 CST. "+24h" is the stars-today figure on GitHub trending (daily) at about 07:44 CST; "n/a" means the repo is not on that list.
| Repo or tool | What it is | Stars / +24h | Runnable today? | Why a tinkerer should care |
|---|---|---|---|---|
| aws/aws-emulator | Local emulator of the AWS APIs, agent-built | 3,906 / n/a | Runnable (News; CC0-1.0; Linux packages, Docker elsewhere) | A free test target for any agent that writes AWS code (§4) |
| huawei-csl/MACaQ | Calibrated quantization to GGUF, vLLM W4A4, Diffusers | 169 / n/a | Runnable (News; Apache-2.0; paper) | Huawei says it quantized a 122B model in 64 minutes in 32.5 GB of device memory. MACaQ-R refines existing GGUFs and stays llama.cpp-compatible |
| Atuin Threads | Synced, searchable agent sessions in Atuin 18.24 | n/a | Runnable (News) | atuin resume picks up a Claude Code session in Codex on another machine. Session history is context you already have |
| QwenLM/qwen-code | Terminal coding agent, 1.0 preview | 29,586 / +151 | Runnable (News; preview) | Native binaries, custom endpoints; an easy way to run a Qwen-first harness |
| CopilotKit/OpenBot | Open-source AI coworkers, each with its own browser, files and tools | 11,266 / +291 | Watch (MIT) | Every action is decided before it runs and recorded, per the README |
| ashhart/TensorFold | LLM inference engine for Metal, CUDA and Vulkan | 1,845 / +58 | Watch (Apache-2.0; pushed today) | A single engine across Apple, NVIDIA and Vulkan GPUs; check model coverage before relying on it |
| klad-app/klad | Paint app a coding agent can drive with real brushes | 153 / n/a | Watch (source-available; Show HN 79) | A small example of an app designed to be operated by an agent |
| FrigadeHQ/yogurt | Private, local AI meeting notes | 757 / n/a | Runnable (MIT; Show HN 266) | A local-model app you can read end to end |
| openai/codex-security | Codex Security CLI and SDK | 12,120 / +663 | Runnable (Context; Apache-2.0) | Still climbing after yesterday |
| Alishahryar1/free-claude-code, addyosmani/agent-skills, meta/astryx | Free-token proxy; skill pack; design system | 68,920 / +636; 116,380 / +812; 26,013 / +1,789 | Skip | Momentum only, no new runtime worth adopting |
Keepers:
- AWS Emulator is the one to install this week. Check the per-service fidelity table on its releases page for the services you use before you trust a green test.
- MACaQ has a llama.cpp path with documented workflows for Qwen3.8-27B and Qwen3.6-35B-A3B. Its quality claims (lower perplexity than MR-GPTQ and DuQuant++) are Huawei's; compare against an Unsloth v3.1 quant of the same model on your own prompts.
Agent builder stack (protocols, runtimes, evals, context engineering, sandboxes, computer use)#
- News A2UI 1.0 protocol (Google, 2026-11-11; HN 30).
- What: the first stable version of an open protocol where the agent describes what to show and the app renders it with its own registered components. Renderer SDKs for web, Android and iOS; agent SDKs for Python, TypeScript and Dart.
- Data handling: A2UI binds data on the server, so sensitive values don't have to pass through the model, and UI streams in as the agent works.
- Fit: Google says A2UI output can show up in hosts that support MCP apps, and that components can be picked by decision models rather than generated as text. Pair it with yesterday's MCP Apps in Claude if you're choosing a UI layer for an MCP server.
- News Gemini Managed Agents self-hosted environments sandboxes (2026-11-11 14:10 CST). Google keeps running the model and agent loop, and shell and file tools execute in a sandbox you control. Same-day integrations: Fly.io Sprites guide (each command and file stays in a microVM in your Fly org) and Agent Substrate on your own Kubernetes clusters. The Gemini docs page did not render here, so the details here come from the official post and the partner guides.
- News gVisor is joining the CNCF sandboxes, with Modal becoming a maintainer. Modal says it has upstreamed GPU support, multi-node training and snapshots. The gVisor blog also has a post on checkpoint/restore getting 5x faster.
- News Perplexity, "Escaping SPACE, part II" sandbox red-teaming (post). Perplexity gave 12 models two tasks against its own VM sandbox: reach a host credential and get past the network policy. Its findings: standard capture-the-flag framing steered agents away from the boundary under test, and security-specialized models did not reliably do better. Perplexity says it fixed the network-policy gap it found. If you red-team your own sandbox, phrase tasks around the real boundary.
- News AA Decision Index v2.1 evals is worth reading for method alone. It scores each answer as accuracy less a Brier penalty, which is the payoff of answering above a confidence threshold and routing elsewhere below it, averaged over every threshold. It also has a symmetry test that checks whether a decision holds when options are reordered or the question is negated. That is a ready-made checklist for any router you build.
- News Atuin Threads context engineering. Agent sessions are append-only logs, so Atuin syncs them with the same encrypted engine as shell history. You can search past sessions by intent and fork one into a different harness.
- News Lock your GitHub Actions supply chain (HN 289). A walkthrough of workflow lockfiles (GA since 10-28):
gh extension install github/gh-actions-lock, thengh actions-lockwrites.github/workflows/actions.lock, plus a verify workflow. This matters most for agent repos where bots open PRs that touch workflows. - Context Stanford CS 329Z, Engineering AI Agents learning (Fall 2026). Students build RAG, tool use and agent loops from scratch before using frameworks. Latent Space reports the materials are public under MIT and that one assignment, designed by SWE-bench co-author John Yang, has students write Terminal-Bench-style tasks.
- Context OpenAI, "What Affects CoT Monitorability Across the Model Lifecycle?" monitoring research (2026-11-10). It studies how pretraining, architecture, RL and model-generated data affect two proxies for chain-of-thought monitorability. Relevant if your agent oversight reads reasoning traces.
- Not printed: HoneyBench v0.2 (no Goodhart Labs page for it found), the KAIST evidence-linked memory paper (covered in HF papers yesterday, no code confirmed), Pomerado's website-to-MCP service (Launch HN; a hosted product, nothing to run locally).
Infrastructure dash#
- News vLLM on Vera Rubin NVL72 engine, nominated by Latent Space (2026-11-11 17:25 CST). On SemiAnalysis AgentX with MiniMax-M3.1-Flash, Rubin against GB300 NVL72: 10.2x total tok/s per chip (including cached) at 546 P90 tok/s per user, 6.5x at 376, and up to 1,205 P90 tok/s per user at low concurrency. These are vLLM's numbers on a SemiAnalysis workload.
- News vLLM EPD disaggregation upgrades engine (2026-11-11 04:03 CST). Splitting encode, prefill and decode with better data transfer, cache reuse and routing cut P90 TTFT by 44–47% and raised throughput per GPU by 6.6–9.5% (counting encoder GPUs) on a DeepSeek V4.1 Flash multi-page document benchmark at concurrency 64–128. Useful if you serve multimodal models and scale the whole tier for image-heavy traffic.
- News SGLang-TorchTPU engine runs Muse Spark Open on a single TPU v7 host. Claimed: 3x faster small-batch decode (custom megakernel and DFlash2), 18% faster decode at 500K context, twice the context per chip with fp8 KV, and up to 77% more agent throughput with HiCache and session-aware eviction. Code and recipes are in the SGLang cookbook.
- News MACaQ quantization (see §5) also has a vLLM W4A4 path and an Ascend path.
- News gVisor to CNCF sandbox runtime (see §6).
- Context Gemini 4 Carbon serving caveat is still on its $2 / $10 introductory price through 11-30 with implicit caching only.
- Still open Val Town platform shutdown goes read-only on 12-10.
- Engine blogs: vLLM and SGLang, both nominated by Latent Space and confirmed on the projects' own posts. TRT-LLM and InferenceX: not nominated. The HN inference slice had no vllm, sglang, trt-llm, tensorrt-llm, mooncake or "kv cache" stories above noise; "inference" surfaced MACaQ.
Builder actions#
- Point agent tests at a local AWS:
aws agent-toolkit add-skill --skill-name aws-emulator --region us-east-1, or install the Linux package from aws/aws-emulator. Check the fidelity scores on the releases page for your services first. - Benchmark a router on decision models: take 200 real routing or grading decisions from your logs and run them through
gemini-decision-1via the Decisions API. Compare accuracy, calibration and cost against your current LLM call. Use the AA Decision Index to pick a cheaper second candidate. - Add a symmetry check to your router: reorder options and negate one question per test case, as AA-Symmetry does, and fail the test if the top choice flips.
- Re-check effort settings on Claude: per AA, try Opus 5.5 xhigh where you run Sonnet 5.5 max, and Fable 5.2 high where you run Opus 5.5 max. Compare cost per solved task.
- Give small models tools before upgrading them: rerun one visual or analysis task with Haiku 5.5 plus code execution against a bigger model without tools (VIALS-AA result).
- Run Managed Agents in your own sandbox: follow the Fly.io Sprites guide or use Agent Substrate if you already run Kubernetes.
- Sync agent sessions: update to Atuin 18.24 and try
atuin resumeacross two machines (Atuin Threads). - Lock your Actions:
gh extension install github/gh-actions-lock && gh actions-lock, then add the verify workflow (walkthrough). - Upgrade
llm:pip install -U llm(oruv tool upgrade llm) to 0.37.2 if you have third-party OpenAI-compatible endpoints configured. - Quantize one local model two ways: run MACaQ's llama.cpp workflow on Qwen3.8-27B and compare it with Unsloth's v3.1 GGUF on your own eval set.
- Try Qwen Code 1.0 Preview: download the native package from the release and point it at a custom endpoint, such as the free ERNIE 5.5 Flash Preview on OpenRouter.
- Prototype a native agent UI: pick one agent response that is currently a wall of text and render it with A2UI 1.0 components.
- Still open — Conductor Cloud: export before it goes read-only tomorrow, 11-13 (announcement).
- Still open — price Gemini 4 Carbon before 12-01: the introductory price doubles then. Arena puts Carbon at $1.5 median per task, against $7.3 for Fable 5.2 (Agent Arena).
- Still open — ERNIE 5.5 Flash weights: due by 11-25; watch llama.cpp #31925 for local support.
- Still open — Claude usage reset: one free full reset before 12-10 4:00 PM CST (ClaudeDevs).
- Still open — Val Town: migrate before read-only on 12-10 (timeline).
- Still open — Claude Code email check: keep the 2.1.317 check on while #107757 is open.
- Still open — Inkling 2 Small weights: thinkingmachines on HF still lists only the first Inkling and Inkling-Small families.
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 since 11-10. Decision Index page data read for Gemini Decision 1 and all 19 decision models. AA's X posts on Decision 1, Claude tiers, VIALS-AA and AA-Robotics used as AA's own numbers. No Qwen4 entry |
| OpenRouter rankings + models API | Hit | No new model IDs in the window. Top-apps panel dated 11-11; Codex rose to 830B. Cost-per-session panel is client-rendered |
| Hacker News front_page API | Hit | AWS Emulator (895), Gemini Decisions API (478), Maya-site agents (1,010; covered yesterday), Lock Your GitHub Actions (289), Yogurt (266), MACaQ (121), Anthropic IPO (194, not printed). Nvidia CEO stories not builder items |
HN last-24h (Algolia, created_at_i within 24h, points>15) |
Hit | 64 stories. Added Atuin Threads (226), Klad (79), Pomerado (56), A2UI 1.0 (30), United Airlines agent ban (592, not printed) |
| HN inference slice (per keyword) | Hit | inference 13 (promoted MACaQ), vllm 35 (none promoted), llama.cpp 2. sglang, trt-llm, tensorrt-llm, mooncake and "kv cache" returned 0 |
| GitHub trending daily | Hit | modern-web-guidance-src, codex-security, sure, agent-skills, astryx, OpenBot, free-claude-code, qwen-code, TensorFold, llama.cpp |
| HF papers 2026-11-12 + 2026-11-11 | Hit | SteerDistill, Read Less Solve More, Ensemble Routing, Evolving Infra, PolyRoute, MACaQ, Deletion Tuning, Evidence-linked memory. MACaQ promoted through its repo; the rest had no code or serving path confirmed |
| HF models trending | Hit | MiniMax H3-Image, Arete-27B, LTX-2.5, Qwen3.8-27B, Ripple-65B Preview, Kimi K3.1 Pro, Unsloth Qwen3.8-27B GGUF. Praxis-1 found through the runwayml org listing |
| pricepertoken.com/news (+ agents, coding, open-source, model-releases) | Hit | Decisions API, LS AINews and Jev piece, Codex on Bedrock, AgentCore Web Search, Spring AI agents, AlphaGenome 2. Model-releases slice had nothing new in the last 24h beyond AlphaGenome 2 |
| news.learnprompt.pro | Partial | SPA shell only |
| LearnPrompt data JSON | Hit | latest-24h.json generated 02:23 CST, 286 items. Decisions API (several), AWS Emulator (Techmeme), Praxis-1 and Qwen4 (AIbase), Pomerado, codex-security |
| radarai.top/en | Hit | AA Claude-tier analysis, ERNIE 5.5 Flash write-ups, Drex OCR, Codex on Bedrock, Kiji Inspector, Omnigraph, CodeGraphy. No CN-lab release beyond what HF and OpenRouter showed |
| The Agentic Digest | Hit | 11-12 issue: Gemini Decisions API |
| simonwillison.net 2026/Nov/11 | Hit | Gemini Decisions API, aws-emulator, llm 0.37.2, Anthropic roadshow video (not printed), tools.simonwillison.net tags |
Dailies in rank order
- Latent Space AINews: Hit. 11-12 issue "D-Day (Decision Models Day)" loaded. Promoted Gemini Decision 1 AA numbers, Agent Arena, AA effort and VIALS posts, Managed Agents self-hosted, gVisor/CNCF, Perplexity SPACE, vLLM Rubin and EPD, SGLang-TorchTPU, Unsloth v3.1, Praxis-1, CS 329Z, Cloudflare decision-model pattern. Vals, HoneyBench, Continual-ARC, Transluce and Theo's Carbon timings not printed (primaries not opened or not found).
- AlphaSignal: Partial (several items PRO-gated). NVIDIA SkillEvaluator, JEV-35B, Unsloth Qwen3.8-27B, hs-rerank-1, AlphaGenome 2. SkillEvaluator: no repo found on GitHub search, dropped.
- Code4U digest: Partial. 11-11 entries were Adobe Firefly Builder and local MCP preview, IBM/Microsoft Imagine Enterprise, Odessia, TSMC investments, two research items, and Carbon GA. Nothing new promoted.
- TLDR AI: Hit (archive). Latest issue 11-11 (Carbon, Claude Sheets, Antigravity Swarm), all covered yesterday. Nothing new promoted.
- TLDR flagship: Not used (TLDR AI was not blocked).
- The Neuron: Hit. Latest "DeepMind launched AlphaGenome 2" with Grok 4.8 and AWS Emulator in the subhead. Grok 4.8: no x.ai post found, dropped.
- AI Daily Brief: Miss. Latest edition is 11-10, already in the window covered yesterday.
Nominated → confirmed or dropped
| Event | Status |
|---|---|
| Gemini Decision 1 / Decisions API | nominated (HN, LS, Agentic Digest, PPT, LearnPrompt, Simon Willison, X pulse) → confirmed (Google blog, AA Decision Index page and post) |
| AWS Emulator | nominated (HN, Simon Willison, LearnPrompt, Neuron, X pulse) → confirmed (GitHub repo and releases, All Things Distributed post) |
| Antigravity browser use | nominated (X pulse) → confirmed (Antigravity post, v2.31.0 changelog) |
| Managed Agents self-hosted environments | nominated (LS, X pulse) → confirmed (Google AI Studio post, Fly.io docs, rakyll post); Gemini docs page did not render |
| Fable 5.2 Agent Arena, AA-Robotics, VIALS-AA, AA Claude tiers | nominated (LS, RadarAI) → confirmed (Arena and AA posts) |
| A2UI 1.0 | nominated (HN) → confirmed (Google Developers Blog) |
| Atuin Threads, Yogurt, Klad, MACaQ | nominated (HN) → confirmed (author blog and repos) |
| Qwen Code 1.0 Preview | nominated (GitHub trending) → confirmed (GitHub release) |
| Codex rust-v0.167.2, Copilot CLI 1.0.110 prereleases, Gemini CLI 0.69 preview | found on release pages during confirm → confirmed |
| Codex on Bedrock | nominated (PPT, RadarAI) → confirmed (AWS ML blog) |
| llm 0.37.2 | nominated (Simon Willison) → confirmed (GitHub release) |
| Unsloth Dynamic v3.1 | nominated (LS, AlphaSignal) → confirmed (Unsloth docs, HF repo updated 11-11) |
| Praxis-1 | nominated (LS, LearnPrompt, X pulse) → confirmed (Runway post, HF org, vLLM post); robotics, kept short |
| gVisor to CNCF | nominated (LS) → confirmed (Modal post) |
| Perplexity SPACE red-teaming | nominated (LS) → confirmed (Perplexity post; blog URL from that post, page itself returned 403 to the fetcher) |
| vLLM Rubin, vLLM EPD, SGLang-TorchTPU | nominated (LS) → confirmed (vLLM and SGLang official posts) |
| Lock Your GitHub Actions | nominated (HN) → confirmed (author post, GitHub changelog, gh-actions-lock repo) |
| CS 329Z | nominated (LS) → confirmed exists (course site); MIT license claim is Latent Space's |
| OpenAI CoT monitorability | nominated (LS, X pulse) → confirmed (OpenAI alignment blog) |
| ERNIE 5.5 llama.cpp support | nominated (LS Reddit recap) → confirmed (draft PR #31925) |
| Cloudflare decision-model architecture | nominated (LS) → TheTuringPost talk post only; printed as Cloudflare demo numbers |
| Grok 4.8 | nominated (Neuron) → no x.ai post; dropped |
| NVIDIA SkillEvaluator | nominated (AlphaSignal) → no repo found; dropped |
| HoneyBench v0.2, Continual-ARC, Transluce, Vals Carbon index, Theo timings | nominated (LS) → primaries not found or not opened; not printed |
| United Airlines bans AI booking agents | nominated (HN) → news report only, no United primary opened; not printed |
| Anthropic IPO, Nvidia leadership, TSMC investments | nominated (HN, LS, Code4U, AlphaSignal) → funding and corporate news; not printed |
| AlphaGenome 2, AlphaProof Code for Signal, Gemini Robotics 2.5 | nominated (LS, PPT, Neuron) → outside builder scope today; not printed |
| Apple "Faster Retrieval" RAG research | nominated (HN 220) → research page with no code confirmed; not printed |
Engine blogs: vLLM and SGLang (nominated by Latent Space, confirmed on official posts). TRT-LLM, InferenceX: not nominated.