Snapshot#
A real model day. Google shipped Gemini 4 Hydrogen, its fast tier, and wired it into a planner/executor mode in Antigravity. Thinking Machines released the Inkling 2 Small weights under Apache 2.0, and Edge0 put out a 145B open-weight model that it says runs on one 24 GB GPU plus system RAM. On the harness side, Claude Code added an opt-in mode where the model writes JavaScript to call its tools, OpenCode made subagents run in the background by default, and llama.cpp tagged v1.0.0.
- Gemini 4 Hydrogen (Google, 2026-11-12): $1.50 / $7.50 per 1M tokens flat across a 1M context, four thinking levels including a new xhigh, on every Gemini API tier including free. On the AA Intelligence Index it scores 53.9 at xhigh for $3.16 per task, just under Sonnet 5.5 (max, 54.5) and GPT-6.1 Sol (max, 54.6).
- Antigravity Orchestrator: Carbon plans, Hydrogen edits and runs tests, and either can hand back to the other inside one session. Google says it reaches near-Carbon success with 24% less quota and 19% faster completion.
- Inkling 2 Small weights (Thinking Machines, Apache-2.0): 594B total / 17B active MoE, text and image in, up to 1M context. Day-zero vLLM and SGLang support. Agent Arena puts it #2 among open models.
- Edge0-145B / 120B (HN 288): open weights on a Qwen3.8-Flash-Next base, about 10B active per token, trained natively at 4 bits. The model card measures 23.25 GiB VRAM plus 38.27 GiB host RAM on an RTX 5090 D. Custom license; benchmark claims are Edge0's own.
- Claude Code 2.1.323: experimental Code Mode, where Claude writes one JavaScript script that calls its tools. It also changes the assumed context for unknown model IDs from 200K to 1M, which can break sessions on smaller-window third-party models.
- Self-hosted Claude Managed Agents (private beta): you run the server, PostgreSQL, workers and sandboxes; model calls go to Anthropic, Bedrock, Google, Foundry or your own gateway.
- llama.cpp v1.0.0 (2026-11-13 04:20 CST), with ggml also bumped to 1.0.0. The release notes say more information comes at 3 PM CET.
- Perplexity Decider v1.3 (Apache-2.0, Qwen3.8-27B base): open decision models that add
locate, which returns verbatim spans with character offsets.
Do this today: run your agent eval set against google/gemini-4-hydrogen at high and at xhigh with a hard thinking budget, and compare cost per solved task against what you run now.
Frontier and open-weight models#
- News Gemini 4 Hydrogen closed, fast tier (
gemini-4-hydrogen, Google DeepMind, 2026-11-12 ~12:00 CST; model card; HN 546).- What: a speed-first member of the Gemini 4 family. The model card says it is adapted from Gemini 4 Carbon. Inputs are text, images, audio and video in a 1M combined window, with up to 64K synchronous output tokens or up to 1M through background long decode.
- Thinking levels: low, medium, high (default) and a new xhigh.
- Cost: $1.50 / $7.50 per 1M input/output, flat across the full context; cached input $0.075 and batch $0.75 / $3.75 on OpenRouter.
- Availability: all Gemini API and AI Studio tiers including free, Gemini for business, the Gemini app, Search AI Mode, and paid AI Pro and Ultra in Antigravity. Day-zero in Cursor and Devin Desktop and CLI.
- Vendor numbers: Google says it beats Argon on the AA index at its default level (51 vs 50) and reaches 54 at xhigh; HLE-Diamond 58.8%, #1 on Vals Tax Agent Bench (52.3%), MKA 53.4%, Agents' Last Exam v1.1 71.6%.
- Independent numbers: see §3. AA summarizes it as just below Sonnet 5.5 and GPT-6.1 Sol at a third of Sonnet's cost and about 2.5x Sol's. Its standout is a low hallucination rate: 8.5% on AA-Omniscience at xhigh, against 43% for Sonnet 5.5 and 52% for Sol on the same page data.
- Pricing context: Simon Willison points out Hydrogen is only 25% cheaper per token than Carbon and Argon until they double on 12-01, and 15x the price of Haiku 5.5 and GPT-6 Luna ($0.10 / $0.50). He ships it in llm-gemini 0.37.
- Operator caveats: one tester saw xhigh spend 131K thinking tokens and about $1.08 on a single image question without answering, and budget exhaustion on 1.8% of a text benchmark (skalskip92). Set a thinking budget before you run it unattended.
- Builder take: the case for Hydrogen is accuracy you can trust at mid-tier prices, not raw cheapness. If your agent's failures come from confident wrong answers, test it; if you only need cheap tokens, Sol and Terra are cheaper per task on AA.
- News Inkling 2 Small open weights, US lab (announcement, 2026-11-12 14:50 CST). This closes yesterday's Still open item.
- What: 594B total / 17B active MoE, text and image input, up to 1M context, Apache-2.0. Experts are NVFP4 and MXFP8 and it ships with a native DSpark speculator, per vLLM.
- Serving path:
vllm serve thinkingmachines/Inkling-2-Smallon one GB200/GB300 node (vLLM says up to 630 tok/s per user at TP4), or SGLang (576 tok/s per user at batch 1 on Vera Rubin, SGLang). This is datacenter hardware, not a workstation model. - Hosted:
thinkingmachines/inkling-2-smallon OpenRouter at $0.30 / $1.00 per 1M, plus a:freevariant. - Index view: AA scores it 30.6 at $1.18 per task. Agent Arena: #1 US open model, #2 open overall, +2.9% net improvement over 11.5K sessions. Code Arena WebDev: 1647, #1 US open model.
- News Edge0-145B and Edge0-120B open weights, local, Qwen-derived (HF org, inference engine, 2026-11-12 ~12:00 CST).
- What: a "latent MoE" where each group of 64 experts shares two large matrices, about 10B of 145B parameters active per step, trained at 4 bits from the start. Base model is Qwen3.8-Flash-Next, per the model card.
- Hardware, per the card: builds target NVIDIA GPUs with CUDA 13 on Linux or Windows. With residency on, the NVFP4 145B used 23.25 GiB VRAM plus 38.27 GiB host RAM on an RTX 5090 D 24 GB. The blog's laptop numbers (29 GB peak, 80+ tok/s) and its MacBook Pro example are Edge0's; the card says its Apple M2 Ultra runs "do not establish macOS release availability."
- License: Edge0 Community Model License 1.0 for Edge0's parts, and the Qwen Community License for the Qwen materials. Not Apache, even though the engine repo is.
- Benchmarks (vendor): 86.5% on Terminal-Bench 2.1 against Claude Opus 4.6's 78.2%, from Edge0's own evaluations.
- Builder take: the most interesting thing a single 24 GB GPU owner can download today. Run your own coding tasks before you believe the Terminal-Bench number.
- News Perplexity Decider v1.3 and Decider Flash v1.3 open weights, decision models (announcement). Apache-2.0, multimodal, fine-tuned from Qwen3.8-27B (the Flash is about 4.5B parameters). New
locatequestion type returns verbatim spans with offsets and confidence, for NER, PII detection and extractive QA. The full model needs a CUDA GPU with room for about 49 GiB of BF16 weights. Perplexity says it is free on OpenRouter for a week; it was not in OpenRouter's public model list when checked at 07:45 CST. - Update SOV-Transcribe-1 open weights, speech (Salesforce) is the new #1 open-weights speech-to-text model on AA at 2.3% AA-WER, 4.6B active.
- Context Checkpoint forensics CN labs, provenance (HN 318). Will Depue reports that initialization patterns in released weights identify the GPU type used, and that six Chinese open models were initialized on Blackwell systems. Critics argue about the scale and what it means (teortaxes). No change to how you serve these models today.
- Still open ERNIE 5.5 Flash CN lab, weights promised. Full weights are still targeted for 11-25. llama.cpp PR #31925 is still a draft. The preview stays free on OpenRouter.
- Still open Qwen4-Max preview thin. Still no model card, price or weights.
- Not printed: OpenAI's unreleased cyber model (a livestream slide, no artifact), Alpamayo 2 Edge and LucidML's world model (outside core scope), MiniMax H3-Image (image model, covered before).
Pareto frontier (quality vs $/token vs $/task vs session economics)#
AA Intelligence Index v5 (read from page data at 07:45 CST; cost is AA's weighted cost per index task):
| Model | Index | $ per task | Time per task | Terminal-Bench 4.0 |
|---|---|---|---|---|
| Claude Fable 5.2 (Max) | 64.6 | $26.54 | 46.8 min | 59.6% |
| Gemini 4 Carbon (High) | 60.0 | $5.48 | 28.4 min | 65.2% |
| Claude Opus 5.5 (Max) | 59.1 | $10.04 | 25.2 min | 59.6% |
| GPT-6 Astra (Max) | 56.9 | $6.24 | 24.2 min | 59.1% |
| GPT-6.1 Sol (Max) | 54.6 | $1.27 | 23.3 min | 56.1% |
| Claude Sonnet 5.5 (Max) | 54.5 | $10.02 | 34.1 min | 63.6% |
| Gemini 4 Hydrogen (xhigh) | 53.9 | $3.16 | 13.7 min | 53.5% |
| GPT-6.1 Terra (Max) | 51.3 | $0.96 | 14.6 min | 55.6% |
| Kimi K3.1 Pro (open) | 46.5 | $3.68 | 38.4 min | 67.2% |
| Inkling 2 Small (open) | 30.6 | $1.18 | 29.6 min | 13.1% |
- News Hydrogen does not move the cost line; Sol and Terra sit below it on cost per task. It does move the speed line: 13.7 minutes per task is the fastest in the top eight. Hydrogen used about 154K output tokens per task at xhigh, according to AA's page data.
- Index disagreements on Hydrogen:
- Terminal-Bench: AA measures 53.5% on Terminal-Bench 4.0 at xhigh. Google's main blog cites Orchestrator results on Terminal-Bench 4.0, while the Antigravity post cites 57% for Hydrogen and the Orchestrator results on Terminal-Bench 4.1. Google's two pages don't agree on which version they mean.
- Agent Arena: #5 at +14.8% net improvement over 6.5K sessions, the fastest in the top 10 at 5.3 minutes median, $1.6 median per task.
- Vals Index: #6 at 68.5%, 2.0 points behind Carbon at 38% of Carbon's list price.
- Coding: CursorBench 54.3% (Gemini 3.8 Flash was 39.6%); Cognition FrontierCode 2.0 38.2%, just behind Sol; SWE-sweep 13.5% against Astra's 9.5%.
- Open-weight points: Kimi K3.1 Pro remains the top open model on AA (46.5) and has the highest Terminal-Bench 4.0 score in this table. Inkling 2 Small costs $0.26 median per task on Agent Arena, but on AA's index it trails Kimi by 16 points with a very low Terminal-Bench 4.0 score (13.1%). The two indexes measure different things; Arena counts real sessions, AA counts a fixed eval suite.
- Coding-agent number (Context, 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. Hydrogen and Orchestrator are not on it yet.
- Session economics (OpenRouter top apps, usage through 11-12): Hermes Agent is #1 at 1.85T tokens today. The cost-per-session panel is still client-rendered and unreadable from a fetch.
- Multi-model routing as a cost lever: Google's Orchestrator claim (near-Carbon success, 24% less quota, 19% faster) is the vendor version of a planner/executor split you can build yourself; the Antigravity SDK exposes the roles directly.
- Pricing calendar: Carbon and Argon go from $2 / $10 to $4 / $20 on 12-01 (Simon Willison's table). After that date Hydrogen is a 62.5% per-token discount on them rather than 25%.
Coding CLIs/IDEs/harnesses#
Times converted from UTC to CST.
- News: Claude Code 2.1.323 (2026-11-12 16:12 CST; changelog).
- Code Mode (experimental, opt-in): Claude writes JavaScript that calls its tools, including MCP tools, instead of one tool call per turn. Organizations can turn it off. The linked docs page returned 404 when checked this morning.
- Breaking for third-party models: the assumed context for an unknown model ID moves from 200K to 1M. Anthropic's own note says sessions on smaller-window models can fail every turn once full; set
CLAUDE_CODE_DISABLE_1M_CONTEXT=1if you point Claude Code at other vendors. - Session-to-session messaging: the number of messages your local sessions can send each other without a prompt goes from about 20 to 100,000. If you run agent groups, that safety stop is effectively gone.
- Sandbox fixes: several, including cases where the Bash sandbox could switch off mid-session or
sandbox.excludedCommandslet an unnamed command run outside it. Worth the upgrade on its own. - Plugins:
engines["claude-code"]inplugin.jsonpins version ranges;claude plugin validatetype-checks TypeScript hooks. - Regression report: #109446 (opened 11-13, 6 comments) objects to the removal of
subagentStatusLine, which per-agent cost and context displays relied on (HN).
- News: OpenCode v2.3.0 (2026-11-13 02:55 CST). Subagents now always run in the background and report back when idle, so you keep working in the parent session. Tool calls collapse into one activity row by default (
session.verbosityrestores detail). Experimental MCP events behindmcp_events. - News: Codex 0.168.0 (11-12 21:54 CST) and 0.169.0 (11-13 02:37 CST; changelog). Side-chat panes and tabs in an opt-in terminal workspace, custom model providers added from the terminal (with device-code sign-in for gateways over SSH), audio recordings for audio-capable models. Security-relevant fix:
codex execkeeps project config and MCP servers disabled in untrusted directories even with--sandbox workspace-write; use--trustto save trust explicitly. - News: Copilot CLI 1.0.111 (11-13 07:40 CST): regex support in enterprise-managed Shell, Bash and PowerShell permission rules, and
--host-allow-session-creationfor one-launch remote session creation. - News: Gemini 4 Hydrogen in Antigravity, plus Orchestrator (HN 445). Orchestrator (Carbon) and Orchestrator (Argon) are rolling out to AI Ultra; Pro gets Orchestrator (Argon). Carbon starts read-only to plan with explicit re-planning triggers, Hydrogen edits and runs tests, and hands back when it hits friction.
- News: Devin Teams update: Devin Desktop for the whole team without extra seats, and two concurrent sessions per user in a new Free mode on SWE-2.
- Update Conductor is back reversal. The team says it will keep developing Conductor and resume updates next week. Yesterday's shutdown dates (read-only 11-13, shutdown 11-20) no longer apply; check the announcement page before you migrate anything.
- News Claude incident outage, 2026-11-13 03:58–05:58 CST: authentication errors on claude.ai, Claude Code, Console and Cowork; a fix was deployed and was being monitored at last update (HN 306).
- Context Gemini CLI is still on v0.69.0-preview.0. Claude Code #107757 remains open at 118 comments.
Usable tools and open projects#
Star counts from the GitHub API around 07:50 CST. "+24h" is the stars-today figure on GitHub trending (daily) at about 07:43 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 |
|---|---|---|---|---|
| Edge0-AI/Edge0 | Inference engine and app for the Edge0 models | 10,817 / n/a | Runnable (News; engine Apache-2.0, weights under a custom license; CUDA 13) | A 145B MoE on a 24 GB card with experts paged to host RAM (§2) |
| ggml-org/llama.cpp v1.0.0 | Local inference | n/a | Runnable (News) | Version 1.0; the release also carries server fixes for SSRF in the CORS proxy and media URLs, and opt-in tools on non-loopback addresses |
| hyperionaidev/smolmaxxing | llama.cpp patches for tiny devices | 527 / n/a | Runnable (MIT; HN 407) | Measured runs of 16B MoE models on about 1 GB of RAM by paging experts from flash; needs its patched llama.cpp |
| Parcha-ai/agentrun | Self-hosted control plane, durable execution and self-improvement on top of your harness | 766 / n/a | Watch (Apache-2.0; Show HN 120) | An open counterpart to managed-agent platforms; check which harnesses it supports |
| Expo Simulators | Cloud iOS simulators and Android emulators for agents | n/a | Runnable (paid Expo plans; $0.10 per simulator minute, 60 minutes a month included until 2027) | Agents return a recording of the app running before you review |
| alibaba/open-code-review | Code review tool: deterministic pipelines plus an LLM agent | 50,233 / +206 | Runnable (Apache-2.0; OpenAI- and Anthropic-compatible) | Line-level comments with a built-in ruleset; a reviewer for agent-written PRs |
| ChromeDevTools/chrome-devtools-mcp | DevTools for coding agents over MCP | 55,248 / +60 | Runnable (Context; Apache-2.0) | Still the standard browser-debugging MCP server |
| backnotprop/plannotator | Plan, annotate and review agent work | 10,286 / +100 | Watch (Apache-2.0) | A review surface between you and the agent's plan |
| Vercel UI SDK | 50+ interface behaviors for agents to build UIs with | n/a | Watch (React, Vue, Svelte, Solid, Web Components per @ctatedev) | Behaviors rather than finished components, designed for agent-generated UI |
| openai/codex-security, GoogleChrome/modern-web-guidance-src | Security CLI; web best-practice skill | 12,843 / +754; 4,263 / +1,875 | Runnable (Context) | Still climbing; covered earlier this week |
| ZhuLinsen/daily_stock_analysis, tradermonty/claude-trading-skills | Stock analysis app; Claude Code trading skill pack | 71,290 / +272; 3,914 / +60 | Skip | Momentum, not builder infrastructure |
Keepers:
- Edge0 is the one to try if you have a 24 GB NVIDIA card and 40+ GB of RAM. Read the license and the residency flags (
--resident-budget-gib,--expert-profile) on the model card first. - llama.cpp v1.0.0: if you expose
llama-serveron a network, upgrade for the SSRF and tool-exposure fixes regardless of what 1.0 brings.
Agent builder stack (protocols, runtimes, evals, context engineering, sandboxes, computer use)#
- News Self-hosted Claude Managed Agents runtime, sandboxes (private beta, 2026-11-12; ClaudeDevs; Anthropic article).
- What you run:
cma-server(API, agent loop, session storage), PostgreSQL 16+, and workers (ant beta:workeror an SDK worker) that run shell and file tools in your sandboxes. The server never runs tools itself. - Model routing: Anthropic, Amazon Bedrock, Claude on Google Cloud, Microsoft Foundry, or your own gateway. The model still runs outside your network.
- Access: through your Anthropic account team only. Some hosted features are missing; check the docs' differences section before planning a migration.
- Builder take: the harness/sandbox split it documents (agent loop and credentials outside the sandbox, a proxy attaches credentials to MCP calls) is a good reference design even if you never use it.
- What you run:
- News Claude Code Code Mode (§4). Letting the model batch tool calls in a script cuts per-call round trips and the context spent on tool-call plumbing, at the cost of reviewing a script instead of individual calls. Try it on read-heavy tasks first.
- News Brave Search API over x402 payments for agent tools (HN 166). Endpoints answer unpaid requests with HTTP 402 and a price; the agent signs a USDC payment on Base and retries. $0.005 per request for LLM Context and web search, no account or API key needed, with gas covered for a limited time. Brave points to Stripe's Purl CLI wallet.
- News Expo Simulators verification infrastructure (GA). Each session records the agent's actions, video, screenshots, logs, network traffic and CPU and memory use, and there is an
eas-simulatorskill. Same pattern as yesterday's AWS Emulator: give the agent a place to check its own work. - News instabox sandboxes, multi-OS cloud sandboxes (Linux, Android, Mac, Windows) with GPUs; the company says users created 1M+ boxes in under a month of beta.
- News ParseBench-2 evals (LlamaIndex), a document-parsing benchmark with severity-weighted error types and open data and eval code, and OWE-Bench (Mercor), 26 long-horizon engineering projects where Fable 5.2 scores 54.6%, GPT-6 Astra 44.2% and Opus 5.5 41.6%.
- News Decider v1.3
locateopen decision models (§2) gives you a local, Apache-licensed span extractor with offsets, useful for PII redaction before context injection. - News OpenAI Daybreak Defense Factory and Daybreak Red Team security products, report only. Defensive vulnerability handling and automated red teaming products. No public artifact to run.
- Not printed: Simular box's OSWorld claim and Handoff² (vendor posts only, primaries not opened), SoMem and TRUST papers (no code confirmed), HF Daily Papers "Actor, Judge, or Teacher?" on decision models in agents (no code confirmed).
Infrastructure dash#
- News vLLM on Vera Rubin NVL72 with DeepSeek-V4.1-Flash engine, nominated by Latent Space (2026-11-12 16:31 CST). 1.7M tok/s per GPU at 387 tok/s per user on SemiAnalysis AgentX, which vLLM puts at 58x GB300 NVL72 at that operating point and 2.7x at peak (2.52M tok/s per GPU). A rack-wide Mooncake KV pool across DRAM and spare HBM held the growing prefixes. These are vLLM's numbers.
- News TileRT megakernels on GB300 NVL72 engine: 1,725 tok/s per user on DeepSeek v4.1 Flash at TP24, which SemiAnalysis says is 2x general-purpose engines for decode.
- News SGLang-Neuron serving is public in an AWS Deep Learning Container: an out-of-tree plugin that runs SGLang serving and RL rollouts on Trainium with native PyTorch.
- News Inkling 2 Small on vLLM and SGLang day-zero serving (§2).
- News llama.cpp v1.0.0 local and the Edge0 engine (§5). A separate Arm NEON prompt-processing study claims 6.5x faster prefill on Arm; it has 7 HN points and has not been checked independently.
- Context NVIDIA's CEO resigned on 11-12 (NVIDIA announcement); CFO Colette Kress is interim CEO. Nothing changes in CUDA or serving stacks today.
- Context a Microsoft 365 service degradation dominated HN (1,028 points); not AI-specific.
- Still open Val Town platform shutdown goes read-only on 12-10.
- Engine blogs: vLLM (nominated by Latent Space, confirmed on vLLM's posts), SGLang (nominated by Price Per Token and Latent Space, confirmed on the AWS blog and SGLang's post). TRT-LLM and InferenceX: not nominated. The HN inference slice returned nothing for sglang, trt-llm, tensorrt-llm, mooncake or "kv cache"; vllm and llama.cpp surfaced smolmaxxing, llama.cpp v1.0.0 and the NEON study.
Builder actions#
- Price Hydrogen against your current model: run your eval set through
google/gemini-4-hydrogenathighandxhighand compare cost per solved task, not per token. Cap thinking tokens first (budget-exhaustion report). - Build a planner/executor split: wire Carbon as planner and Hydrogen as executor with the Antigravity SDK, or the same pattern with any two models in your own harness, and measure success rate and cost against the single big model.
- Try Edge0 on a 24 GB GPU: follow the model card build steps (CUDA 13) and run five of your own coding tasks. Read the license before any commercial use.
- Hit Inkling 2 Small for free: call
thinkingmachines/inkling-2-small:freeon OpenRouter on an agentic task before deciding whether to self-host it withvllm serve thinkingmachines/Inkling-2-Small. - Upgrade Claude Code and set one variable: update to 2.1.323 for the sandbox fixes, and set
CLAUDE_CODE_DISABLE_1M_CONTEXT=1if you use non-Anthropic models. - Turn on Code Mode for one read-heavy session: compare turn count and tokens with and without it (release notes).
- Re-check agent group limits: if your local Claude Code sessions message each other, add your own cap now that the default prompt-free limit is 100,000 (release notes).
- Try background subagents: update to OpenCode v2.3.0 and run two subagents while you keep working in the parent session.
- Check
codex execin CI: after updating to 0.169.0, add--trustwhere CI jobs relied on project config or MCP servers in a fresh checkout. - Upgrade
llama-serveron any networked box:git pull && cmake --build buildto v1.0.0 for the SSRF and tool-exposure fixes. - Add a local PII span extractor: run pplx-decider-flash-v1.3 with a
locatequestion on 50 real tickets and check offsets before you trust it for redaction. - Prototype pay-per-call search: try one Brave Search API x402 call from a test wallet with a small balance and see how your harness handles a 402.
- Give a mobile agent a simulator: run one PR through Expo Simulators inside the 60 included minutes.
- Still open — price Carbon and Argon before 12-01: they double to $4 / $20 then (pricing table).
- Still open — ERNIE 5.5 Flash weights: due by 11-25; watch llama.cpp #31925.
- 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.
- Closed today: Inkling 2 Small weights are out (item 4), and the Conductor Cloud deadline no longer applies (Conductor).
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 | Changelog shows four Hydrogen evaluations on 11-12 and JT-4.1 Flash on 11-11. No new article since 11-10. Index page data read for the §3 table. AA's X post on Hydrogen and SOV-Transcribe-1 used as AA's own statements |
| OpenRouter rankings + models API | Hit | New IDs: google/gemini-4-hydrogen and :batch (11-12 12:01 CST). Benchmarks panel not yet showing Hydrogen. Decider v1.3 not listed at check time. Cost-per-session panel client-rendered |
| Hacker News front_page API | Hit | Gemini 4 Hydrogen (546), Hydrogen and Orchestrator (445), smolmaxxing (407), Claude errors (306), Edge0 (288), Apple Foundation Models (231), Brave x402 (166), AgentRun (120). NVIDIA CEO, Anthropic IPO and Microsoft 365 outage stories not builder items |
HN last-24h (Algolia, created_at_i within 24h, points>15) |
Hit | 102 stories. Added checkpoint forensics (318), Claude Code #109446 (25), Expo Simulators (24), Conductor is back (32), Libreagi (149, not printed), Inception diffusion API shutdown (13, no URL, dropped) |
| HN inference slice (per keyword) | Hit | inference 16, vllm 26, llama.cpp 2; sglang, trt-llm, tensorrt-llm, mooncake and "kv cache" returned 0. Promoted smolmaxxing and llama.cpp v1.0.0 |
| GitHub trending daily | Hit | open-code-review, modern-web-guidance-src, codex-security, daily_stock_analysis, sure, plannotator, uptime-kuma, buzz, claude-trading-skills, tailcat, chrome-devtools-mcp |
| HF papers 2026-11-13 + 2026-11-12 | Hit | SO-Arena, G-STAR, Actor/Judge/Teacher decision models in agents, Agent-R², SteerDistill. None promoted; no code or serving path confirmed in the window |
| HF models trending | Hit | MiniMax H3-Image, Arete-27B, LTX-2.5, Qwen3.8-27B, Ripple-65B Preview, Cactus Mirage, Kimi K3.1 Pro, Inkling-2-Small. Edge0 found through HN and its HF org |
| pricepertoken.com/news (+ agents, coding, open-source, model-releases) | Hit | LS AINews, SGLang-Neuron, Natera AgentCore, Amplitude/OpenAI, LiteLLM gateway on Bedrock, Hydrogen (three items in model-releases last 24h) |
| news.learnprompt.pro | Partial | SPA shell only |
| LearnPrompt data JSON | Hit | latest-24h.json generated 07:06 CST, 326 items. Hydrogen (many), Edge0, Codex CLI 0.168/0.169, Expo Simulators, Burnstate, Foxpit, Offstage Cloud, Claude status |
| radarai.top/en | Hit | Hydrogen, Vercel UI SDK (shadcn), Decider v1.3, Inkling 2 Small on Arena, Netflix observability graph. No CN-lab release in the window |
| The Agentic Digest | Hit | 11-13 issue: user-scoped knowledge-base search for Bedrock agents (AWS how-to, not printed) |
| simonwillison.net 2026/Nov/12 | Hit | llm-gemini 0.37 and the Gemini 4 pricing table |
Dailies in rank order
- Latent Space AINews: Hit. 11-13 issue "Jensen Huang resigns after 33 years leading NVIDIA" loaded. Promoted Hydrogen third-party numbers, Inkling 2 Small serving, Edge0, Decider v1.3, self-hosted Managed Agents, Expo Simulators, instabox, Devin Teams, ParseBench-2, OWE-Bench, vLLM Rubin and Mooncake, TileRT, checkpoint forensics. Simular, Handoff², SoMem, TRUST, KQuant and Alpamayo 2 Edge not printed.
- AlphaSignal: Partial (several items PRO-gated). Claude Meetings, self-hosted Managed Agents, Hydrogen, Cactus Mirage (PRO), Claude trading toolkit (PRO, skip), World Labs Atlas, Cursor Hydrogen. Nothing new beyond the above promoted.
- Code4U digest: Partial. Latest entries dated 11-12 (Google Tabletop, Fractile and Agile Robots funding, NY AI center, AlphaProof Code, Decisions API). Nothing new promoted.
- TLDR AI: Hit (archive). Latest issue 11-12 (Grok Bot 1.0, AlphaGenome 2, Gemini Decisions), all outside the window or covered yesterday.
- TLDR flagship: Not used (TLDR AI was not blocked).
- The Neuron: Hit. Latest "Jensen Huang resigned as Nvidia CEO." Nothing promoted.
- AI Daily Brief: Miss. Latest edition is 11-11.
Nominated → confirmed or dropped
| Event | Status |
|---|---|
| Gemini 4 Hydrogen | nominated (HN, LS, PPT, LearnPrompt, RadarAI, AlphaSignal, Simon Willison) → confirmed (Google blog, DeepMind model card, AA page data, OpenRouter models API) |
| Antigravity Orchestrator | nominated (HN, LS) → confirmed (Antigravity blog) |
| Hydrogen third-party evals | nominated (LS) → confirmed (Arena, Vals, Cursor, Cognition, KLieret posts) |
| Inkling 2 Small weights | nominated (LS, RadarAI, HF trending) → confirmed (HF model page, Thinking Machines, vLLM, SGLang and Arena posts) |
| Edge0-145B/120B | nominated (HN, LS, LearnPrompt) → confirmed (Edge0 blog, HF model card, GitHub repo) |
| Decider v1.3 | nominated (LS, RadarAI) → confirmed (HF model cards); "free on OpenRouter" not visible in OpenRouter's list at check time |
| Self-hosted Claude Managed Agents | nominated (LS, AlphaSignal, X pulse) → confirmed (Claude platform docs, Anthropic article, ClaudeDevs post) |
| Claude Code 2.1.323 / Code Mode | found on the release page during confirm; #109446 nominated by HN → confirmed (GitHub release and issue) |
| OpenCode v2.3.0, Codex 0.168/0.169, Copilot CLI 1.0.111 | found on release pages during confirm (Codex also LearnPrompt) → confirmed |
| llama.cpp v1.0.0 | nominated (HN inference slice) → confirmed (GitHub release) |
| Expo Simulators | nominated (HN, LS, X pulse, LearnPrompt) → confirmed (Expo blog) |
| Brave Search x402 | nominated (HN) → confirmed (Brave blog) |
| AgentRun, smolmaxxing | nominated (HN) → confirmed (repos) |
| Conductor is back | nominated (HN) → confirmed (founder's post) |
| Claude incident | nominated (HN, LearnPrompt) → confirmed (status page) |
| SGLang-Neuron | nominated (PPT) → confirmed (AWS ML blog) |
| vLLM Rubin/Mooncake, TileRT | nominated (LS) → confirmed (vLLM and SemiAnalysis posts) |
| Vercel UI SDK | nominated (RadarAI, X pulse) → confirmed (Vercel post) |
| ParseBench-2, OWE-Bench, instabox, Devin Teams, Daybreak | nominated (LS) → confirmed (vendor posts) |
| Checkpoint forensics | nominated (HN, LS) → confirmed (report page); printed as context |
| Arm NEON prefill study | nominated (HN inference slice, LearnPrompt) → repo exists; printed with caveat |
| OpenAI unreleased cyber model | nominated (LS, X pulse) → livestream slide only; not printed |
| Inception diffusion LLM API shutdown | nominated (HN, no URL) → no primary; dropped |
| Simular box, Handoff², KQuant, SoMem, TRUST | nominated (LS) → primaries not opened or no code; not printed |
| Apple Foundation Models framework | nominated (HN 231) → page loaded, no dated change found; not printed |
| Libreagi | nominated (HN 149) → no builder artifact; not printed |
| NVIDIA CEO change, Anthropic IPO, Fractile and Agile Robots rounds | nominated (HN, LS, Neuron, Code4U) → corporate news; NVIDIA kept as one context line |
Engine blogs: vLLM (nominated by Latent Space, confirmed on vLLM's posts), SGLang (nominated by Price Per Token and Latent Space, confirmed on the AWS blog and SGLang's post). TRT-LLM, InferenceX: not nominated.
Banner: generated today (pale steel ice-blue with deeper copper-amber, long parallel lanes with brief crossings; seed Gemini 4 Hydrogen and Orchestrator).