Snapshot#
Google's day. Gemini 4 Carbon went from a cyber-defender preview to general availability at an introductory price that doubles on 12-01, and two independent indexes put it near Claude Fable 5.2 on score at a fraction of the cost per task. Two Chinese-lab-adjacent open-weight datapoints moved: Nex-N2.6 weights landed, and Baidu put ERNIE 5.5 Flash Preview on a free OpenRouter tier with Apache 2.0 weights promised.
- Gemini 4 Carbon is generally available (Google, 2026-11-10): $2 / $10 per 1M tokens through 11-30, then $4 / $20. Paid tier only, no free tier, and implicit caching only.
- Artificial Analysis scores Carbon (High) 60 at $7.76 per index task, priced at the $4 / $20 list price. Fable 5.2 (Max) scores 65 at $26.54.
- CursorBench 4.0: Carbon High 60.4% at $9.22 per task, Fable 5.2 Max 60.0% at $32.79. This confirms the Cursor number that was dropped yesterday for lack of a primary.
- MKA benchmark (Google DeepMind + AA, new): Fable 5.2 leads at 57% for $31.97 per task. Carbon scores 54% for $6.29.
- Antigravity /swarm (preview) and Cursor multiplayer Projects (beta) both shipped multi-agent coordination layers on the same day.
- Nex-N2.6 weights are on Hugging Face under Apache-2.0: FP8, about 321B parameters, built on GLM-5.3-Flash. Yesterday it was API-only.
- ERNIE 5.5 Flash Preview (Baidu): 280B total and 14B active parameters, 1M context, free on OpenRouter, Apache 2.0 weights "soon".
- Modern Web Guidance 1.0 (Chrome, Edge, Firefox, MDN): an agent skill plus CLI that steers coding agents toward native HTML/CSS. Show HN reached 397 points.
Do this today: run your hardest golden task on gemini-4-carbon at medium thinking, next to Fable 5.2 High. Compare cost per solved task before the price doubles on 12-01.
Frontier and open-weight models#
One frontier closed model went GA, two open-weight releases moved, and one Chinese frontier preview was announced with no specs.
- News Gemini 4 Carbon closed frontier (Google, 2026-11-10; HN 287 points). It was announced 10-28 for cyber defenders only.
- What: Google's strongest model, aimed at long-horizon coding, cybersecurity, knowledge work and science. Google says Gemini 4 Argon stays the default for most coding and knowledge work. Carbon "uses more tokens than Argon to explore thoroughly".
- Spec: text, image, audio, video and file input. Model ID
gemini-4-carbon. Per the pricing page, it supports long decode up to 1M tokens. OpenRouter lists a 1,048,576-token context and up to 943,718 output tokens. - Cost (Gemini API pricing): $2 input / $10 output per 1M (thinking included) and $0.10 cached through 11-30. From 12-01 it is $4 / $20 / $0.20. Batch is half that. There is no free tier. Caching is implicit only, with no explicit cache control. OpenRouter lists the same introductory prices and a batch ID at $1 / $5.
- Availability: Gemini API (paid tier), AI Studio, Android Studio, Gemini for business. Ultra subscribers get it in the Gemini app, AI Mode and Antigravity. It is also in GitHub Copilot for Pro+, Max, Business and Enterprise at provider list pricing, and in Cursor (@cursor_ai). Docs: model card, safety report.
- AA numbers (model page): Intelligence Index 60 (High), $7.76 per task at $4 / $20, 205M output tokens across the index, 26.1 minutes per task against a 12.6-minute median, and 78 tok/s output. AA calls it "notably slow per task and very verbose".
- Vendor numbers: Google says Carbon blocked 99% of attack trajectories in its simulated red-team setup and passed 92% of a 375-task agentic honesty benchmark (Argon: 70%). These are Google-run tests.
- Builder take: the introductory price is the story. At $2 / $10 it costs the same per token as Argon. After 12-01 it costs the same as Opus 5.5 per token, and AA already prices it at that rate.
- Update Nex-N2.6 open weights, Apache-2.0 (Nex AGI; HF repo created 2026-11-09 23:34 CST, after yesterday's brief went out).
- What: a multimodal agentic model built on GLM-5.3-Flash. FP8 weights, about 321B parameters. Code is at sii-nex/Nex-N2.6.
- Vendor numbers: OSWorld 83.0 (base 78.0), Terminal-Bench 2.1 84.1 (base 84.3), AutomationBench 45.7 (base 48.8). The reasoning-repetition rate fell from 16.0% to 1.6%, and tool-list adherence is 100%, both on internal replay evals.
- Serving: the card points to the SGLang GLM-5.3-Flash cookbook with an FP8 configuration.
reasoning_effortis low, high or max (default max). The API is $0.12 / $0.40 on OpenRouter. - Builder take: a computer-use-tuned open model you can self-host, but it trails its base on AutomationBench. There is no AA entry yet.
- News ERNIE 5.5 Flash Preview CN lab, weights promised (Baidu, dated 2026-11-11; @Baidu_ERNIE 01:56 CST).
- What: the first post-trained model on ERNIE-5.5-Flash-Base (weights out since 10-21 under a custom ERNIE license). It is a MoE with 280B total and 14B active parameters, 1M context, and training aimed at coding, game engines (Unreal, Unity, Godot), Blender/CAD and knowledge work.
- Cost / serving: the API is on Baidu Qianfan, with OpenRouter
baidu/ernie-5.5-flash-preview:freeat $0 (listed 11-10 11:58 CST, up to 921,600 output tokens). Baidu's page says Apache 2.0 weights are coming "soon", and its X post says by 11-25. Baidu also claims over 200 tok/s per user on Qianfan. - Builder take: free to try today. There is no AA or independent score yet, so treat it as a candidate for a worker-model slot.
- News Qwen4-Max preview CN lab, thin (Alibaba Qwen, 2026-11-10 20:01 CST; HN 45). The X post says it is in preview on Alibaba's Token Plan and in QwenWork, Qoder, the Qwen app and QwenStudio. No spec sheet, price, benchmark or weights are published. Check back for a model card.
- Context Kimi K3.1 Pro Highspeed on Amazon Bedrock availability. It is a faster variant of the open-weight Kimi K3.1 Pro. Moonshot reports 2–3x the speed of K3.1 Pro. Moving from Kimi K3 is a model-ID swap (
global.moonshotai.kimi-k3.1, orus./in.for residency). AWS says it runs with zero data retention, and the post includes an OpenCode setup. - Context ERNIE-Light-DT-1.0 other OpenRouter IDs at $1.13 / $3.94 is built on open-source GLM models with Baidu Sparse Attention. I found no weights.
- Index view (OpenRouter benchmarks panel, AA Intelligence Index): Fable 5.2 (Max) 64.6, Gemini 4 Carbon (High) 60.0, Opus 5.5 (Max) 59.1, GPT-6 Astra (Max) 56.9, GPT-6.1 Sol (Max) 54.6, Sonnet 5.5 (Max) 54.5, Qwen3.8 Max 53.4, GPT-6.1 Terra (Max) 51.3, Gemini 4 Argon (High) 50.4, Fable 5.1 (Xhigh) 50.0. Carbon re-entered the top ten after yesterday's absence.
- Not printed: the Apple local-model Mac app (9to5Mac report, no Apple primary), PAN 2 and GR00T 2 (world and robot models, outside this brief's scope), and MiniMax H3-Image (open image model, trending since 11-03).
Pareto frontier (quality vs $/token vs $/task vs session economics)#
Three indexes priced the same frontier differently today. Each table names its index.
Coding: CursorBench 4.0 (Cursor's own harness and tasks, cost from each model's published per-token pricing; Cursor doesn't say whether Carbon is priced at the introductory or the 12-01 rate):
| Model / effort | Score | Avg $ per task |
|---|---|---|
| Gemini 4 Carbon High | 60.4% | $9.22 |
| Gemini 4 Carbon Medium | 60.1% | $8.49 |
| Gemini 4 Carbon Low | 58.6% | $11.89 |
| Claude Fable 5.2 Max | 60.0% | $32.79 |
| Claude Fable 5.2 High | 56.9% | $9.70 |
| Claude Opus 5.5 Max | 57.8% | $13.43 |
| Claude Opus 5.5 High | 56.0% | $3.97 |
| Claude Sonnet 5.5 Extra High | 53.1% | $2.81 |
| Gemini 4 Argon High | 52.9% | $4.57 |
| Claude Haiku 5.5 Max | 48.4% | $1.12 |
- News on Cursor's index, Carbon Medium matches Fable 5.2 Max for about a quarter of the cost per task. Carbon Low costs more than Medium, so lower effort is not automatically cheaper here. Opus 5.5 High remains the cheap point near 56%.
Knowledge work: AA Intelligence Index (v5.0.1, list prices):
| Model | Index | $ per index task |
|---|---|---|
| Fable 5.2 Max | 65 | $26.54 |
| Fable 5.2 High | 62 | $9.08 |
| Gemini 4 Carbon High | 60 | $7.76 (at $4 / $20) |
| Opus 5.5 Max | 59 | $10.04 |
- AA prices Carbon at the post-12-01 list price. At the introductory $2 / $10 the token bill would be about half, but that is my arithmetic, not an AA figure.
Visual knowledge work: MKA on AA (article, 2026-11-10):
| Model | Climb WSR | $ per task |
|---|---|---|
| Claude Fable 5.2 (Max) | 57% | $31.97 |
| Gemini 4 Carbon (High) | 54% | $6.29 |
| Claude Opus 5.5 (Max) | 53% | n/a in the article |
| GPT-6.1 Sol (Max) | n/a in the article | $1.11 |
| GPT-6 Luna (Max) | n/a in the article | $0.18 (cheapest) |
- AA's read: cost per task spans nearly 30x within 10 points of the top score. Below Carbon, all four OpenAI models tested sit on the Pareto line, along with Claude Haiku 5.5 and Xiaomi MiMo-V2.6-Pro. GPT-6 Astra (Max) uses the fewest output tokens (8.39M for the run). The benchmark was built by Google DeepMind, which also built Carbon. AA runs it on its own Stirrup harness with a GPT-6 Luna judge.
- Index disagreement: Cursor puts Carbon level with or above Fable 5.2 Max on coding. AA's Intelligence Index puts Fable Max 5 points ahead, and MKA puts it 3 points ahead. All three agree Carbon is far cheaper per task than Fable Max.
Coding-agent number (AA Coding Agent Index v2.0, harness and model bundled, read from the page data): Claude Code + Opus 5.5 (Max) 61.7, Devin Fusion CLI (Opus 5.5 Max + GPT-6 Luna High) 61.5, GitHub Copilot Fusion (max) 59.9, Antigravity CLI + Gemini 4 Carbon (High) 59.4, Claude Code + Sonnet 5.5 (Max) 59.3, Claude Code + Fable 5.1 (Max) 57.6, Antigravity CLI + Argon (High) 53.2. Fable 5.2 is not on this index yet.
Open-weight and CN points (OpenRouter models API): Nex-N2.6 $0.12 / $0.40 (cache $0.024), ERNIE 5.5 Flash Preview $0 (free tier), Muse Ember $0.10 / $0.50 (closed), Carbon $2 / $10. No AA cost-per-task data exists yet for Nex-N2.6 or ERNIE 5.5 Flash Preview.
Session economics:
- OpenRouter top apps ("Today", usage through 11-10): Claude Code 1.91T tokens, Hermes Agent 1.69T, Kilo Code 908B, pi 636B, OpenClaw 622B, Codex 555B, Strix 401B, omp 349B, DeepSeek Harness 332B. Claude Code retook #1 from Hermes.
- The OpenRouter "Cost per session" panel is client-rendered and could not be read.
- Token-efficiency caveat: Anthropic's Fable 5.2 prompting guide says the "this codebase is yours" line used about 1.5x the output tokens in its tests (§6).
Coding CLIs/IDEs/harnesses#
Carbon landed in every major surface on day one. Two vendors shipped team-scale multi-agent coordination. Times converted from UTC to CST.
- News: Cursor Projects is now multiplayer (beta, all users, 2026-11-10).
- What: teammates share one Project with a task board, shared context (requirements, specs, mockups), and agents that see each other's plans. Coordinators flag overlapping work, such as a schema change against a feature that reads it, and agree on an order before anyone hits a merge conflict. On-call handoffs carry in-flight state.
- Numbers: a Cursor engineer reports 62 concurrent agents across 11 people and 972 merged PRs in 15 days (X). That is an internal anecdote, not an eval.
- Builder take: the useful idea is the shared context layer built once per project rather than per person. Start one from the left nav and invite one teammate first.
- News: Google Antigravity /swarm (preview).
- How it works: a coordinator spawns parallel workers in isolated git worktrees, plus adversarial verifiers that run tests and challenge diffs. Agents talk on a shared, append-only message board and "park at zero token cost when idle", per Google.
- Models: each agent can be pinned to a different model. Google suggests Carbon for coordinator, architecture and audit roles and lighter models for edits. In the CLI,
/swarm boardor Alt+S opens a live board.
- News: Gemini 4 Carbon in GitHub Copilot (GA, gradual rollout, provider list pricing). Admins control it through the model policy; new models are on by default unless disabled.
- News: Copilot CLI v1.0.109 (2026-11-10 11:02 CST). Adds IPv4/IPv6 CIDR rules for sandbox host allow/deny and Fable 5.2 in the model picker. In ACP mode, a
preToolUsehook that answers "ask" now forwards the decision to the ACP client instead of denying. - News: llm-gemini 0.36 adds Carbon to Simon Willison's
llmCLI (notes). He tested low, medium and high thinking; medium is the default. - Context: Codex shipped rust-v0.167.1 (11-10 15:32 CST) with no release notes beyond the compare link, plus 0.169.0 alphas through alpha.11. Gemini CLI v0.68.0 (11-10 15:17 CST) is fixes only, with no Carbon mention in the notes. Claude Code is still on v2.1.322.
- News Anthropic's ClaudeDevs says Pro, Max, Ultra and Team users get one free full reset across all models and limits, usable before 12-10 at 4:00 PM CST (2:00 PM PT). It cites fixed bugs that used up more of users' limits than intended.
- Still open: Claude Code issue #107757 (email in User-Agent) is open at 117 comments, last updated 2026-11-09 12:26 CST.
- Still open Conductor deadline Cloud goes read-only 11-13 and shuts down 11-20.
Usable tools and open projects#
Star counts are from the GitHub API, read around 07:50 CST. "+24h" is the stars-today figure on GitHub trending (daily) at about 07:40 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 |
|---|---|---|---|---|
| GoogleChrome/modern-web-guidance | Agent skill + CLI with curated guides for modern web APIs, plus legacy-scan |
3,929 / n/a (src repo +133) | Runnable (News; v1.0.0 11-09 21:24 CST; Apache-2.0) | npx modern-web-guidance@latest install. A skill with a CLI and an eval dashboard (linked from the README), not just a prompt file |
| shadcn/improve | Skill that audits a repo with an expensive model and writes plans for cheaper models to execute | 10,479 / +333 | Runnable (Context; MIT; last push 09-12) | /improve execute sends a plan to a cheaper model in a separate worktree and reviews the diff. This is the planner/executor split that today's Carbon and Fable pricing rewards |
| openai/codex-security | Codex Security CLI (cs) and TypeScript SDK: scan, validate, patch, export SARIF |
11,437 / +20 | Runnable (Apache-2.0; npm i -g @openai/codex-security) |
Pushed today. Needs Node 22.13+/24/26, Python 3.10+, and OpenAI auth |
| ArtificialAnalysis/Stirrup | AA's open agent harness, the one MKA runs in | 708 / n/a | Runnable (MIT) | Pair it with the 10 public MKA tasks (Apache-2.0) to grade your own agent on visual deliverables |
| agent-substrate/substrate | Actor-based agent runtime on Kubernetes (CNCF Sandbox) | 6,433 / n/a | Watch (Update; Apache-2.0) | NVIDIA wired it to Dynamo (§7). Needs a cluster |
| vercel-labs/agent-browser | Browser-automation CLI for agents | 49,625 / +712 | Runnable | Still climbing |
| homeassistant-ai/ha-mcp | MCP server for Home Assistant | 7,009 / +640 | Runnable (MIT) | Rides today's #1 HN story on local-first smart homes |
| meta/astryx, addyosmani/agent-skills | "Agent-ready" design system; skill pack | 24,277 / +2,842; 115,607 / +788 | Skip (momentum only) | No new runtime |
Keepers:
- Modern Web Guidance is a context-engineering artifact with an owner: four browser vendors, version 1.0, and evals that drop content models already know. Run
npx modern-web-guidance@latest legacy-scan .on a front-end repo before your next agent session. Its own notes say to check the guidance against the browsers you support. - improve is the cheapest way to test the expensive-planner, cheap-executor split. Run
/improve quickwith Fable or Carbon, then/improve execute 001with a Haiku- or Ember-class model.
Agent builder stack (protocols, runtimes, evals, context engineering, sandboxes, computer use)#
- News MCP Apps in Claude protocol (public preview, ClaudeDevs 2026-11-10).
- New: Claude now supports more of the MCP Apps extension: side panel and full screen, entry points from the Artifacts page, app tools (your UI registers tools Claude can call to change what's on screen), messages from the app to Claude, embedded websites in an iframe, and downloads.
- Portability: because the app ships with your MCP server, it runs in any client that supports the extension.
- Caveat: Anthropic says implementation details may change while the MCP Apps community standardizes them.
- News MKA benchmark evals (Google DeepMind with Artificial Analysis).
- Design: 200 private tasks across 10 verticals and 83 occupations, each 3–8 hours of expert work. About 70 yes/no rubric checks per task, weighted 75% accuracy and 25% visual design. Models averaged 30–130 turns per task.
- Grading: AA runs it in Stirrup with a 250-turn limit and grades with a GPT-6 Luna agentic judge. Google says experts agreed with its autorater 96% of the time. There is a tech report and a public 10-task set.
- Builder take: the first widely run eval where file output (decks, sheets, marked-up images) is the thing graded. Word documents and text files are the hardest outputs; no model passes 52% on either (AA).
- News Antigravity /swarm orchestration patterns (worktree isolation, verifier agents, append-only board) and Cursor Projects (shared context across a team, agent-to-agent conflict checks). See §4 for details. Both move coordination out of a single coordinator's context window.
- News Prompting Claude Fable 5.2 context engineering (Anthropic docs; Simon Willison).
- The finding: the line "This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it." made Fable merge repeated code in about half of small coding tasks, against under a tenth without it.
- The cost: about 1.5x the output tokens, plus some changes nobody asked for.
- Also in the guide: sections on subagent counts, piled-up caution and long silent turns. The guide says to make these system-prompt changes only in new conversations, because of the Fable 5.2 prefix check.
- News AgentCore Web Search API tool runtime (AWS, GA).
- What: web results with title, URL, publication date and an extract, from an Amazon-operated index, signed with existing AWS credentials, with domain and date filters.
- How: the LangChain AWS tool lets code fix the date window, a source blocklist and a search budget so the model can't relax them. AWS's sample dependency watcher took 36 s in parallel against 99 s sequential, from one example run.
- Context AgentCore Evaluations playbook evals (built-in, DeepEval and AutoEval evaluators). The useful part is the method: start from the business outcome, then pick session-level and tool-use checks.
- News TIS-100 Bench evals, community (HN 82). 16 LLMs ran 47 game puzzles plus 8 community puzzles, one scored attempt each, 10–45 minutes per puzzle, with the public leaderboard withheld. It scores a multi-objective frontier (cycles, nodes, instructions) rather than pass/fail.
- Context VulnCheck's Glasswing tracker security evals (HN 428). It counts 442 CVEs attributed to Anthropic and 37,966 reported findings, of which 19.4% are public and 217 marked fixed. curl rejected 19 of 34 published reports. If you run an AI scanner, budget maintainer triage time and track your own true-positive rate.
- Not printed: "pinned MCP servers resolving fresh dependencies" (Agentic Digest). The only linked source was a security-news recap, and I found no original report.
Infrastructure dash#
- News NVIDIA Agent Substrate with Dynamo agent runtime + inference (2026-11-10).
- Routing: Dynamo's router uses session identity, KV-cache location and tool-call boundaries to keep related calls on workers that already hold the context.
- Feedback loop: Dynamo exposes projected time-to-first-token and GPU-pool saturation, and Substrate uses those to decide whether to keep, suspend or resume an agent session.
- Upstream: request parking in the router, per-container limits for microVM sandboxes, and GPU-state pause/restore through Kubernetes DRA. See the deploy guide.
- News vLLM on the LibTorch Stable ABI engine, nominated by Latent Space (2026-11-06). As of v0.33.0, vLLM publishes ABI-stable CUDA wheels, so a compatible PyTorch upgrade no longer forces a vLLM rebuild. Python APIs, minimum CUDA and third-party dependencies still constrain compatibility.
- News TPU 8i Max and TPU 8tx chips (Google, 2026-11-09).
- 8i Max: two TPU 8i chips in one package, for double the compute, SRAM and HBM bandwidth, aimed at long contexts.
- 8tx: TPU 8t silicon with AMD EPYC hosts for confidential inference and fine-tuning, with up to 2x lower latency than 8t per Google.
- Gap: no pricing, regions or availability dates.
- News Carbon's pricing page says caching is implicit only. You can't pin a cached prefix, so keep prompt prefixes stable and measure the cache hit rate in usage metadata. Every token is billed, including thinking tokens and grounding tool instructions.
- Context Kimi K3.1 Pro Highspeed on Bedrock managed serving. The global profile costs less per token than the geographic ones. There is a Flex tier and explicit prompt caching (see §2).
- Context An Open Field Guide to Inference Engineering learning (Modelplane; HN 89). An introduction for platform engineers new to running models on GPUs.
- Still open Val Town is winding down platform shutdown (HN 402). Signups closed 11-10, read-only on 12-10 (excluding sqlite and blob storage), all other services end 01-10, open-sourcing planned for 02-01. There are migration guides for Cloudflare, Vercel, Netlify and Render.
- Engine blogs: vLLM only, nominated by Latent Space. SGLang, TRT-LLM and InferenceX were not nominated. The HN inference slice returned nothing above noise for sglang, trt-llm, tensorrt-llm, mooncake, "kv cache" or llama.cpp.
Builder actions#
- Price Carbon before 12-01: run your golden tasks on
gemini-4-carbonat medium thinking next to Fable 5.2 High. Record cost per solved task at both the $2 / $10 and $4 / $20 rates. - Check Carbon's cache behavior: caching is implicit only. Keep the system prompt and tools byte-stable and read cached-token counts from responses before trusting a long-session cost estimate.
- Try /swarm on one multi-file change: in Antigravity, pin Carbon to the coordinator and verifier and a Flash model to workers. Compare tokens against a single-agent run.
- Open a multiplayer Cursor Project: create one with a shared spec and invite one teammate. Check whether the agents flag the overlap on a deliberately conflicting pair of tasks.
- Pull Nex-N2.6 if you self-host computer-use agents:
huggingface-cli download nex-agi/Nex-N2.6and serve with the SGLang GLM-5.3-Flash guide (FP8). Or hit OpenRouter at $0.12 / $0.40 first. - Try ERNIE 5.5 Flash Preview for free: set
modeltobaidu/ernie-5.5-flash-preview:freeon OpenRouter and run one tool-calling trace. Watch the ERNIE release page for the Apache 2.0 weights. - Modernize a front-end repo:
npx modern-web-guidance@latest legacy-scan .(repo), then install the skill for your agent. - Split planning from execution: install shadcn/improve, run
/improve quickwith a frontier model, and/improve executewith a cheap one. - Add UI to an MCP server: follow Get started with MCP Apps and register one app tool.
- Grade your agent on visual deliverables: clone Stirrup and run the 10 public MKA tasks.
- Test the ownership line on Fable 5.2: add the sentence from the prompting guide in a new conversation and compare output tokens and diff size.
- Rebuild-free vLLM upgrades: if you pin PyTorch for vLLM, move to v0.33.0+ stable-ABI wheels and test one torch bump.
- Use the Claude reset deliberately: it is one free full reset before 12-10 4:00 PM CST (ClaudeDevs). Save it for a long eval run.
- Still open — Conductor Cloud: export before 11-13 (announcement).
- 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 | New 11-10: MKA article and eval page. Read Carbon model page and Coding Agent Index page data. No new Intelligence Index entries since 11-09 |
| OpenRouter rankings + models API | Hit | Usage through 11-10. New IDs: gemini-4-carbon (+batch), ernie-5.5-flash-preview:free, ernie-light-dt-1.0-intl. Cost-per-session panel is client-rendered |
| Hacker News front_page API | Hit | Carbon (287), Modern Web Guidance (397), VulnCheck/Glasswing (428), Val Town (402), Claude Whiteboards/Sheets/Data (248), Maya LiDAR agents (147). Other front-page items were not builder items |
HN last-24h (Algolia, created_at_i within 24h, points>15) |
Hit | 70 stories. AI items picked by hand. The AI OR LLM OR agent query returned 0 (Algolia ANDs the terms), so the all-stories list was used |
| HN inference slice (per keyword) | Hit | vllm 37 hits (none promoted), inference 15 (promoted the Modelplane guide, 89). sglang, trt-llm, tensorrt-llm, mooncake, "kv cache" and llama.cpp returned 0 |
| GitHub trending daily | Hit | meta/astryx, daily_stock_analysis, modern-web-guidance-src, shadcn/improve, agent-skills, agent-browser, sure, ha-mcp, codex-security |
| HF papers 2026-11-11 + 2026-11-10 | Hit | Evidence-linked agent memory, MMDM, DepthGraft, Infrathon, InfraRSI, e-means, RidgeFill. None promoted (no code or serving path confirmed in the window) |
| HF models trending | Hit | Kimi K3.1 Pro, MiniMax H3-Image, Ripple-65B Preview, Arete-27B, ERNIE-5.5-Flash-Base. Nex-N2.6 found through the nex-agi org listing |
| pricepertoken.com/news (+ agents, coding, open-source, model-releases) | Hit | Carbon (Google posts + OpenRouter), MKA, ERNIE 5.5 Flash Preview, ERNIE-Light-DT, Kimi on Bedrock, AgentCore web search and evals, Meta TLX, NaiveAI (LS) |
| news.learnprompt.pro | Partial | SPA shell only |
| LearnPrompt data JSON | Hit | latest-24h.json generated 06:43 CST, 346 items. Carbon in Copilot, DeepMind Carbon post, Qwen4 and ERNIE (AIbase), codex-security trending, Agent Substrate, Grok Build leak (Ars) |
| radarai.top/en | Hit | Carbon, AlphaProof, BotPsy, a Rust inference engine "Kinetic" (no repo found; dropped). No CN-lab release beyond what HF/OpenRouter showed |
| The Agentic Digest | Hit | 11-11 issue: AgentCore Web Search, Carbon, pinned MCP dependency drift, AgentCore Evaluations, Pinecone embeddings, Modern Web Guidance, Kimi on Bedrock, Adobe MCP |
| simonwillison.net 2026/Nov/10 | Hit | llm-gemini 0.36 (Carbon), Fable 5.2 prompting note, Muse Ember pricing tiers, Yonkers Rises |
Dailies in rank order
- Latent Space AINews: Hit. 11-11 issue ("Mercor $750m Series D") loaded. Promoted vLLM stable ABI and MKA; Carbon partner numbers re-read on Cursor, AA and Google pages. Vals, ARC Prize, Perplexity WANDR, SemiAnalysis and Roboflow figures were not opened, so they are not printed.
- AlphaSignal: Partial (several items PRO-gated). Claude Sheets, ChatGPT merchant agents, Qwen4-Max, Gemini Robotics 2.5, Carbon on CursorBench, Cursor Projects. Permalinks had no outbound primaries; confirmed on Cursor, Qwen's X post and Google instead.
- Code4U digest: Partial. 11-10 entries were Fable 5.2, Fireworks, IPO and policy items, and Skyfall "Steve". Nothing promoted.
- TLDR AI: Hit (11-10 issue). Promoted Google TPU 8i Max/8tx. NVIDIA AutoCompress and GR00T 2 not promoted.
- TLDR flagship: Not used (TLDR AI was not blocked).
- The Neuron: Hit. 11-11 issue "3 AI labs shipped for you": Carbon, Claude Meetings/Sheets, Businesses in ChatGPT. Nothing new promoted.
- AI Daily Brief: Miss. Latest edition is 11-09.
Nominated → confirmed or dropped
| Event | Status |
|---|---|
| Gemini 4 Carbon GA | nominated (HN, PPT, LS, Agentic Digest, Neuron, LearnPrompt, RadarAI, AlphaSignal) → confirmed (Google blog, Gemini API pricing, AA, OpenRouter, GitHub changelog) |
| Carbon CursorBench 60.4% / Fable 5.2 on CursorBench | nominated (AlphaSignal, LS, X pulse) → confirmed (cursor.com/cursorbench) |
| MKA benchmark | nominated (PPT, LS) → confirmed (Google blog, AA article and eval page, HF dataset) |
| Antigravity /swarm | nominated (LS, Agentic Digest, Neuron) → confirmed (Antigravity blog) |
| Cursor multiplayer Projects | nominated (LS, AlphaSignal, X pulse) → confirmed (Cursor blog) |
| ERNIE 5.5 Flash Preview | nominated (PPT, LearnPrompt/AIbase, X pulse) → confirmed (Baidu ERNIE page, OpenRouter, HF base model) |
| Nex-N2.6 weights | found in HF org listing during confirm → confirmed (HF model card, Apache-2.0) |
| Qwen4-Max preview | nominated (HN, AlphaSignal, LearnPrompt/AIbase) → partial (official Qwen X post only; no spec, price or weights) |
| Modern Web Guidance 1.0 | nominated (HN, Agentic Digest, GitHub trending) → confirmed (GitHub repo and v1.0.0 release) |
| MCP Apps in Claude | nominated (LS) → confirmed (Claude docs, ClaudeDevs post) |
| Claude usage reset | nominated (LS) → confirmed (ClaudeDevs post) |
| AgentCore Web Search + Evaluations | nominated (Agentic Digest, PPT) → confirmed (AWS ML blog) |
| Kimi K3.1 Pro Highspeed on Bedrock | nominated (PPT, Agentic Digest) → confirmed (AWS ML blog) |
| Agent Substrate + Dynamo | nominated (LearnPrompt) → confirmed (NVIDIA blog, GitHub) |
| vLLM LibTorch Stable ABI | nominated (LS) → confirmed (vLLM blog) |
| TPU 8i Max / 8tx | nominated (TLDR AI) → confirmed (Google blog) |
| Fable 5.2 "codebase is yours" | nominated (Simon Willison, LS) → confirmed (Anthropic prompting docs) |
| shadcn/improve, codex-security | nominated (GitHub trending, LearnPrompt) → confirmed (GitHub repos) |
| VulnCheck Glasswing tracker, Val Town, Modelplane guide, TIS-100 Bench | nominated (HN) → confirmed exists (author pages; claims are the authors') |
| llm-gemini 0.36 | nominated (Simon Willison) → confirmed (GitHub release) |
| Apple local-AI Mac app / "Harbor" | nominated (HN, LearnPrompt) → no Apple primary; dropped |
| Pinned MCP servers dependency drift | nominated (Agentic Digest) → only a recap link, no original report found; dropped |
| Grok Build coding-agent data leak | nominated (LearnPrompt via Ars Technica) → no vendor primary opened; dropped |
| Anthropic AMD MI355X bring-up with Mythos | nominated (X pulse) → roadshow video quotes only, no written primary; dropped |
| Vals Index, ARC-AGI-3, WANDR, SemiAnalysis GDPx, Roboflow on Carbon/Fable | nominated (LS) → primaries not opened; not printed |
| Kinetic Rust inference engine | nominated (RadarAI) → no repo found; dropped |
| Claude Whiteboards/Sheets/Data, Claude Meetings, Businesses in ChatGPT | nominated (HN, Neuron, AlphaSignal) → confirmed exists (Claude page); consumer/workspace features, not printed beyond MCP Apps |
| PAN 2, GR00T 2, Gemini Robotics 2.5, MiniMax H3-Image | nominated (X pulse, TLDR, LS, AlphaSignal) → outside builder scope today; not printed |
Engine blogs: vLLM (nominated by Latent Space, confirmed). SGLang, TRT-LLM, InferenceX: not nominated.