Snapshot#
A model-launch day, and a price-per-task day at both ends of the market. Anthropic shipped a new top model that is cheaper per token and more expensive per task, and Meta shipped a small model that sits on Artificial Analysis's cost-per-task line. A third source, Amazon's Strands team, put numbers on swapping a frontier model for a decision model.
- Claude Fable 5.2 and Mythos 5.2 (Anthropic, 2026-11-09): $8 / $40 per 1M tokens, down from $10 / $50. Artificial Analysis puts Fable 5.2 (Max) at 65, #1 on its Intelligence Index, at $26.54 per index task versus $10.04 for Claude Opus 5.5 (59).
- Fable's effort ladder (AA): Low 54 at $3.10, High 62 at $9.08, Xhigh 62 at $14.83. High already beats Opus 5.5 Max on score at slightly lower cost per task.
- Meta Muse Ember: $0.10 / $0.50 per 1M. AA's write-up scores it 36 at $0.14 per task, against $0.41 for Claude Haiku 5.5 at the same score. Closed weights.
- Strands Decider replaces Sonnet 4.6 for ticket labeling inside Amazon: about $500 versus $22,200 per million tickets, and 4.7x faster per call (authors' numbers). It follows yesterday's Decision Index v2.
- Claude Fable 5.2 API changes can break code that edits earlier conversation turns or still calls the older
computer_20251124tool. Claude Code, Managed Agents and the Agent SDK are unaffected by the history change. - Claude Code v2.1.321 makes Fable 5.2 the default Fable and fixes several permission and sandbox cases, plus MCP OAuth requests sending secrets to another host after a redirect.
- Fireworks AI incident: Fireworks now reports impact to secrets, datasets and incomplete models for a limited set of accounts. Notion says under 1% of workspaces may be affected.
Do this today: run your hardest golden task on Fable 5.2 at High next to Opus 5.5 at Max and compare cost per solved task, not the token price. Then price a routing slice (extraction, labeling, tool calls) on Muse Ember.
Frontier and open-weight models#
One frontier closed model, one small closed model, and a few open-weight datapoints.
- News Claude Fable 5.2 and Claude Mythos 5.2 closed frontier (Anthropic, 2026-11-09; HN 700 points, 863 comments).
- What: one underlying model, sold two ways. Fable is open to everyone. Mythos is for organizations verified through Anthropic's Cyber and Life Sciences programs. IDs are
claude-fable-5-2andclaude-mythos-5-2. - Spec: 1M context, 128k max output, adaptive thinking always on, five effort levels (low to max). Anthropic says output is up to 30% faster than 5.1. An optional fast mode runs up to 2.5x faster at $16 / $80.
- Cost: $8 / $40 per 1M tokens, cache reads $0.20 (Claude Code notes). OpenRouter lists the same prices, plus a batch variant at $4 / $20.
- Vendor numbers: FrontierCode 2.0 49.1% (Xhigh) versus 45.5% for Opus 5.5; FrontierSWE v2 69.9% versus 62.3%; OSWorld 2.1 offline subset 89.3%; AA-Briefcase v1.1 1785 Elo versus 1800 for Opus 5.5. Anthropic says Fable at High matched Opus 5.5 at Max on FrontierCode 2.0 at roughly half the cost per task.
- Operator caveats: Fable 5.2 requires data retention by default so Anthropic's safety classifiers can run; zero data retention is only for eligible enterprises (GitHub changelog). Anthropic's own advice is to start most coding and knowledge work on Opus 5.5, and move up only when evals say so.
- Availability: Claude API, Claude Code and Amazon Bedrock, GitHub Copilot (Pro+, Max, Business, Enterprise), and OpenRouter.
- Builder take: the token price fell but AA's per-task cost at Max rose to about 2.6x Opus 5.5. Evaluate it at High, not Max.
- What: one underlying model, sold two ways. Fable is open to everyone. Mythos is for organizations verified through Anthropic's Cyber and Life Sciences programs. IDs are
- News Meta Muse Ember closed small model (Meta, 2026-11-09; HN 316 points).
- What: Meta's smallest model, aimed at repeated agent work: tool calls, coding subtasks, routing, extraction, long-document parsing. 1M context on OpenRouter.
- Cost: $0.10 input / $0.50 output / $0.004 cached input per 1M (AA, matching the OpenRouter listing).
- Serving: Meta's first-party API and Muse Code, plus OpenRouter. No weights.
- AA numbers: Intelligence Index 36 (max) and 34 (xhigh). Coding Agent Index 46 in Muse Code, versus 37 for Haiku 5.5 in Claude Code and 31 for GPT-6 Luna in Codex. Weak spot: AA-Omniscience accuracy of 21.1%, because it declines 64% of questions (hallucination rate 17.4%).
- Builder take: a cheap subagent or router target. Check it on your own tool-call traces before trusting the coding-agent number, which comes from Meta's own harness.
- News Nex-N2.6 open-weight derivative, weights unconfirmed (Nex AGI, OpenRouter listing 2026-11-09 21:46 CST).
- What: an agentic model built on GLM-5.3-Flash, about 321B parameters, aimed at computer use, coding and long tool chains.
- Cost: $0.12 / $0.40 per 1M, cache reads $0.024, 262k context.
- Gap: I found no
Nex-N2.6repo under nex-agi on HF (only older N2 models), no license, and no AA entry. Treat it as API-only until a card appears.
- Update Inkling 2 Small open-weights promise now has a free OpenRouter variant (added 2026-11-09 16:04 CST). It is a multimodal MoE with 17B active of 594B total parameters, 1M context. The announcement still says weights start closed on Tinker, with an intent to release under Apache 2.0 later. thinkingmachines on HF still lists only the first Inkling and Inkling-Small families. AA scored it 31 on 11-05.
- Context the OpenRouter benchmarks panel (AA Intelligence Index): Claude Fable 5.2 (Max) 64.6, Claude Opus 5.5 (Max) 59.1, GPT-6 Astra (Max) 56.9, GPT-6.1 Sol (Max) 54.6, Claude Sonnet 5.5 (Max) 54.5, Qwen3.8 Max 53.4, GPT-6.1 Terra (Max) 51.3, Gemini 4 Argon (High) 50.4, Claude Fable 5.1 (Xhigh) 50.0, Kimi K3.1 Pro 46.5. Gemini 4 Carbon, #1 since 10-28 per AA, is not in the panel's top ten.
- No new Chinese-lab weights were confirmed in the window. The HF trending list shows Kimi K3.1 Pro, Ripple-65B Preview and Arete-27B among the top entries, all covered earlier.
- Not printed: Qwen4 scores said to beat Opus 5.5 (Chinese-language recap only), and Cursor's "CursorBench 60%" claim for Fable 5.2 (AlphaSignal headline, no Cursor primary found).
Pareto frontier (quality vs $/token vs $/task vs session economics)#
Index scores and per-task costs are from Artificial Analysis model pages (Intelligence Index v5.0.1, list prices), read around 07:35 CST. Token prices come from the OpenRouter models API.
Fable 5.2 on the AA Intelligence Index, by effort:
| Model / effort | Index | $ per index task | Notes |
|---|---|---|---|
| Fable 5.2 Low | 54 | $3.10 | |
| Fable 5.2 Medium | 56 | $3.80 | Default in Claude apps and Claude Code |
| Fable 5.2 High | 62 | $9.08 | |
| Fable 5.2 Xhigh | 62 | $14.83 | Same score as High, 63% more cost |
| Fable 5.2 Max | 65 | $26.54 | 275M output tokens, 44.6 min per task |
| Opus 5.5 Max | 59 | $10.04 | 218M output tokens |
- News Fable 5.2 High (62 at $9.08) beats Opus 5.5 Max (59 at $10.04) on both score and cost per task. Xhigh adds cost with no AA score gain over High. Max buys 3 more points for roughly 3x the cost of High. AA's own summary calls Max "particularly expensive" and "very verbose".
- Vendor cross-check: Anthropic says the same thing on FrontierCode 2.0 (High about level with Opus 5.5 Max at half the cost per task). Two sources agree on direction here, but one is the vendor.
- News (small-model Pareto, AA article):
| Model | Index | $ per index task |
|---|---|---|
| Meta Muse Ember (max) | 36 | $0.14 |
| Meta Muse Ember (xhigh) | 34 | $0.12 |
| Claude Haiku 5.5 (max) | 36 | $0.41 |
| DeepSeek V4.1 Flash 1017 (max) | 33 | $0.60 |
| GLM-5.3-Flash | 28 | $0.58 |
| Meta Muse Wonder (max) | 44 | $6.95 |
- Why it moves the line: Ember and Haiku 5.5 share the same list price per 1M input and output. Ember's cached input is 60% cheaper and it uses about a third of the output tokens, and AA says cached input is most of the cost in agentic evals. Ember also beats the two open-weight Chinese models above on both axes, though neither is a like-for-like open-weight comparison with a closed API.
- Coding-agent number (AA Coding Agent Index): Muse Ember in Muse Code 46, Claude Haiku 5.5 in Claude Code 37, GPT-6 Luna in Codex 31. Harness and model are bundled, so this is not a pure model comparison.
- Update Strands Decider E4B vs Sonnet 4.6 decision-model economics, author-reported on one of Amazon's internal ticket-labeling flows (20 labels).
| Claude Sonnet 4.6 (Bedrock) | Strands Decider E4B (SageMaker, ml.g5.xlarge) |
|
|---|---|---|
| Average time per call | 4,215 ms | 897 ms |
| Total tokens per ticket | 6,724 | 2,010 |
| Cost per ticket | $0.0222 | about $0.0005 |
| Cost per 1M tickets | about $22,200 | about $500 |
- How they ran it: a three-way AppConfig switch (
BEDROCK,DUAL,DECIDER) ran both models on live traffic and judged every disagreement. The authors say Decider matched Sonnet; the page shows no accuracy table. - Caveats: list prices only, Lambda cost excluded, written by Amazon engineers about an AWS-built model. The post also says JSON key order in the prompt moved Decider's output probabilities by up to 0.19.
- Link to yesterday: this is a production version of the AA Decision Index v2 argument that billing shape, not the token price, drives $/task.
- Context OpenRouter Auto Code routers as a cost lever (OpenRouter, 2026-11-06; nominated by TLDR AI 11-09). OpenRouter says Auto Code scores in line with Opus 5.5 across five benchmarks at 74% less cost per task, and that its routing step is 3x faster at p90. Those are OpenRouter's numbers; set
modeltoopenrouter/auto-codeto test it. It returned to higher-priced model tiers with this release, and acost_tieroflowrestores the old cheap bias. - Context Fable 5.2 $8 / $40, Muse Ember $0.10 / $0.50, Nex-N2.6 $0.12 / $0.40, Ripple-65B Preview $0.20 / $0.60, Claude Opus 5.5 $4 / $20, GPT-6 Astra $10 / $50 per 1M.
- OpenRouter top apps (rankings, usage through 11-09): Hermes Agent 1.95T tokens, Claude Code 1.38T, Kilo Code 851B, pi 558B, OpenClaw 492B, Strix 405B, omp 351B, DeepSeek Harness 347B, Codex 337B.
- Not available: the OpenRouter "Cost per session" panel and weekly model-usage chart are client-rendered and could not be read. AA has no new cost-per-task data for Kimi, Qwen or DeepSeek today.
Coding CLIs/IDEs/harnesses#
Fable 5.2 landed across the coding surfaces on day one. Dates below are converted from the GitHub UTC stamps to CST.
- News: Claude Code v2.1.321 (2026-11-09 13:22 CST) and v2.1.322 (20:29 CST).
- Added:
claude-fable-5-2as the default Fable (1M context, $8 / $40, $0.20 cache reads),--settings <file>forclaude remote-control, and fast mode on Vertex (global, US, EU). - Fixed: MCP OAuth requests sending refresh tokens to a different host after a server redirect. A read-only Bash command running outside the sandbox without a prompt. File-write permission checks judging the wrong folder. Deny rules not applying to
cpandmvinto a directory. - Builder take: these are permission-boundary fixes. Update, then rerun any allow-list you rely on.
- Added:
- News: GitHub Copilot adds Fable 5.2 for Pro+, Max, Business and Enterprise at provider list pricing under usage-based billing. GitHub says it solved more tasks than Fable 5.1 with fewer steps and tool calls, which is a vendor claim. Copilot CLI v1.0.108 (11-09) adds
/indexcommands with managed tgrep indexing, and a sandbox fix for local MCP servers after/cd. - News: OpenCode v2.2.7 (2026-11-10 04:37 CST). New "GitHub Copilot Auto" model option that lets Copilot's router pick per session. Breaking:
ImageResponse.imageis removed from@opencode/ai, so readresponse.images[0]. - News: pi v1.5.1 (11-09 17:47 CST) and v1.5.2 (11-10 06:25 CST).
- v1.5.1:
pi auth login <provider>without starting a session, idle prompt-cache warming for up to 30 minutes between runs when it is expected to pay for itself, and Anthropic message threads that send only new messages. - v1.5.2: forking a session whose folder no longer exists.
- Builder take: the cache-warming setting matters most for Fable at $0.20 cache reads versus $8 uncached input. Check it before a long idle gap.
- v1.5.1:
- News: Hermes Agent v0.23.1 (11-09 19:42 CST). A patch for a plugin-catalog size cap: since the 11-09 catalog deploy,
hermes plugins installsilently fell back to the catalog shipped with older installs. v0.23.0 and earlier stay on the stale catalog until updated. Hermes is #1 on the OpenRouter apps list (§3). - Context: Codex stable is still 0.167.0 (11-05). Prereleases through 0.169.0-alpha.7 are rolling. Gemini CLI is at v0.67.0 stable, with v0.68.0-preview.0 from 11-04.
- News the Fable 5.2 docs say Claude Code, Managed Agents and the Agent SDK build requests safely, but custom clients may now get a 400 on edited history (§6). The HN thread on the launch runs to 863 comments, and a separate HN post (430 points) reports a one-shot game build. Neither is an eval.
- 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 07:40 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 |
|---|---|---|---|---|
| expo/exact | App framework designed for agents to write: one codebase for web, macOS and iOS, UI in a small declarative language called Contract | 1,697 / n/a | Runnable (News; HN 127; MIT) | The README is a prompt for a coding agent. Needs a Mac with Xcode, Rust, Bun and Chrome. Its example run took about 57 minutes, mostly first builds |
| ranxianglei/billion-context | Context-compression plugin for Claude Code, OpenCode and pi | 6,013 / +374 | Watch (Context; npm install -g billion-context) |
README claims 5x fewer tokens and a 95–97% cache-hit rate. Third-party benchmark folder is in the repo. License shows as NOASSERTION, so read it first |
| sshah03/recalc | A spreadsheet where each cell can run a Claude Code, Codex or pi session | 55 / n/a | Runnable (News; Apache-2.0) | Visible, side-by-side agent runs with checks in adjacent cells. The README warns it is early beta and may break between versions |
| githubnext/ace | GitHub Next experiment: team channels, one worktree per task, hosts joined over a Tailscale tailnet | 71 / n/a | Watch (Context; MIT; site) | Agents run as you with a shell on your Mac, per the page. Run it on a throwaway machine first |
| strands-agents/sdk-python | Python agent harness SDK, now with /box sandbox, /evals and Decider docs |
9,126 / n/a | Runnable (Context; Apache-2.0) | The Decider post above is built on it |
| garrytan/g-harness-evals | Model-fixed, harness-swapped eval (covered 11-09) | 41 / n/a | Runnable (Still open; MIT) | Updated today. Rerun the matrix with Fable 5.2 |
| homeassistant-ai/ha-mcp | MCP server for Home Assistant | 6,362 / +792 | Runnable | A narrow, testable MCP tool surface |
| vercel-labs/agent-browser | Browser-automation CLI for agents | 48,929 / +775 | Runnable | Still climbing |
| anthropics/skills, addyosmani/agent-skills, mattpocock/skills, meta/astryx | Skill packs; an "agent-ready" design system | +1,676; +465; +1,223; +3,111 | Skip (momentum only) | Prompt and skill packs with no new runtime |
Keepers:
- expo/exact is the one new runnable thing. The interesting part is not the framework but the setup: a repo whose README is the agent prompt, with tests the agent must pass on three targets.
- recalc is a cheap way to see several agents on the same task in one grid, which fits the harness-cost work from yesterday.
Agent builder stack (protocols, runtimes, evals, context engineering, sandboxes, computer use)#
- News Claude Fable 5.2 migration changes API / runtime, breaking. Four changes can break code:
- Edited history now errors: a request that changes anything before a returned thinking block (system prompt, tools, or an earlier message) gets a 400 ("The block is bound to a different conversation") on every account. A stopgap is the
thinking-binding-controls-2026-08-01beta header withprefix_mismatch_behaviorset to"drop_block", which can raise output tokens. Details are in the preserved-thinking guide. - Thinking is model-bound: only Fable 5.2 and Mythos 5.2 read Fable 5.2's thinking blocks. Switching to another model mid-conversation loses that reasoning.
- Computer use: the older
computer_20251124tool errors on every platform except Amazon Bedrock. - Advisor tool: a Fable 5.2 executor needs a Fable 5.2 or Mythos 5.2 advisor.
- Builder take: keep
systemandtoolsfixed for a conversation, append only, and send assistant turns back exactly as returned.
- Edited history now errors: a request that changes anything before a returned thinking block (system prompt, tools, or an earlier message) gets a 400 ("The block is bound to a different conversation") on every account. A stopgap is the
- News Anthropic's model-selection note says Fable 5.2 can act as an Advisor to smaller models, or as an Orchestrator that delegates simple tasks downward. Pair this with Ember or Haiku as the worker model.
- News Strands Decider rollout evals, production method. The three-way switch (old model only, both in parallel with the old result applied, new model only) with every disagreement judged is a reusable template for swapping any model in a live flow.
- News AWS Nitro Sandboxes sandboxes run an EKS pod, ECS task or lightweight VM in its own hardware-isolated VM, including AI agents and LLM-generated code. Attestation can gate secrets so only verified sandboxes receive them. Available on most Nitro-based EC2 instances in all commercial regions at no extra charge. See the EKS guide and the EC2 docs.
- News AlphaProof Code verification (Google DeepMind). A multi-model agent framework formally verified key security properties of Signal's actual implementation code. Humans reviewed every formal specification, agents wrote over six million lines of Lean, and the agents flagged about 1% of the codebase for a second human look. HN 147 points. A proof covers only the specified properties.
- Context The components of agent memory context engineering (Anthropic, 2026-11-06; nominated by TLDR AI 11-09). Argues for capturing experience in the moment and consolidating it into long-term memory during idle time. Read it next to the billion-context and cache-warming approaches in §4 and §5.
- Context OpenRouter Auto Router and Auto Code router runtime weigh eleven request factors with live pricing, provider health and cache status, and honor governance policies such as zero data retention and allowed-model lists (see §3).
- News Bedrock Managed Knowledge Base retrieval adds single-shot RAG and retrieval-only modes to
AgenticRetrieveStreamthroughmaxAgentIteration(1 for one retrieval plus answer, 0 for passages only). AWS says single-shot can cut latency and model usage. Test answer quality on your own queries.
Infrastructure dash#
- News Fireworks AI incident, third-party inference. Fireworks's update says account and billing information was affected, and for a limited set of accounts also secrets, datasets and incomplete models. It ties the activity to a financially motivated group that sells access to stolen LLM tokens. Notion's notice (HN 86) says under 1% of workspaces may be affected, only unattributable Insights chat transcripts, and that it stopped serving open-weight models through Fireworks. If you ran jobs or stored keys there, rotate them.
- News GitHub secret scanning for Actions logs availability is in public preview. It scans workflow logs when a run completes and files alerts that persist after the logs are deleted. It is limited to Enterprise Cloud with Secret Protection or Advanced Security.
- News AWS Nitro Sandboxes sandbox infra (see §6).
- News Claude Fable 5.2 on Bedrock and Claude Platform on AWS model on AWS, and NVIDIA Nemotron 3.5 Super VL on SageMaker JumpStart (NVFP4 and BF16; the listed JumpStart configurations use 520 to 15,600 tokens even though the model supports 1M).
- News 120B MoE on a 32GB M1 Max local inference, author-reported (HN 33). The author keeps experts on SSD and caps process memory at 16 GiB. On gpt-oss-120b they report 7.84 tok/s against 5.78 for TUFF 10.2.2 on the same Mac, about 36% faster. Engine code is not released.
- Context GPU rental price spread price data (HN 170). The repo's headline is an H100 at $2.69 or $12.29 an hour depending on where you rent it. It is a community dataset (18 stars), so verify the provider pages before buying.
- Engine blogs: not nominated. The HN inference slice returned no builder items for vllm, inference, sglang, tensorrt, trt-llm, mooncake, "kv cache" or llama.cpp above the noise floor.
Builder actions#
- Compare Fable effort levels on your own task: run your hardest golden task on Fable 5.2 at High, Xhigh and your current Opus 5.5 Max setting. Compare cost per solved task against the AA effort ladder in §3.
- Test your request builder before moving to Fable 5.2: append-only messages, fixed
systemandtools, assistant turns returned verbatim. Check the preserved-thinking guide and swap anycomputer_20251124use. - Update Claude Code: update Claude Code and confirm you are on v2.1.322. Rerun your permission allow and deny rules after the v2.1.321 fixes.
- Price a Muse Ember worker slice: point a routing or extraction step at
meta/muse-emberon OpenRouter and compare cost per task with your current mini model. Keep AA's 21.1% Omniscience accuracy in mind for anything that needs facts. - Try a decision model for labeling: read the Strands Decider post and copy its three-way switch. Run
DUALmode on a day of traffic before switching. - Test OpenRouter Auto Code in your harness: set
modeltoopenrouter/auto-code(announcement). Setcost_tiertolowif you want the old cheap bias. - Rotate anything stored with Fireworks: read the Fireworks update and rotate API keys and secrets you used there if you are on it. Check the Notion notice if your workspace used Insights.
- Update Hermes before installing plugins: upgrade to v0.23.1 so
hermes plugins installreads the live catalog. - Rerun the harness matrix with Fable 5.2: clone g-harness-evals and add Fable 5.2 to one harness pair.
- Try an agent-written app: with a Mac, Xcode, Rust, Bun and Chrome, give your coding agent the prompt in the Exact README. Time the first build.
- Still open — Conductor Cloud: export before 11-13 (announcement).
- Still open — Claude Code email check: keep the 2.1.317 check on while #107757 is open.
- Still open — Inkling 2 Small weights: watch thinkingmachines on HF for the promised Apache 2.0 release.
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: Fable 5.2 (five effort levels), Muse Ember article (11-09). Coding Agent Index page read for the Ember comparison |
| OpenRouter rankings + models API | Hit | Usage through 11-09. New IDs: nex-n2.6, inkling-2-small:free, claude-fable-5.2 (+batch), muse-ember. Cost-per-session panel and weekly usage chart are client-rendered |
| Hacker News front_page API | Hit | Fable 5.2 (980 and 700), Muse Ember (316), AlphaProof Code (147), Exact (127), Strands Decider (428), Nitro Sandboxes (27). Other front-page items were not builder items |
HN last-24h (Algolia, created_at_i within 24h, points>15) |
Hit | 79 stories. AI items picked by hand |
| HN inference slice (per keyword) | Hit | vllm 29 hits (none above noise), inference 15, sglang, tensorrt, trt-llm, mooncake, "kv cache" and llama.cpp 0. Promoted the MoE-on-M1 post (33) |
| GitHub trending daily | Hit | meta/astryx, anthropics/skills, sherlock, mattpocock/skills, ha-mcp, agent-browser, claude-quickstarts, addyosmani/agent-skills, mujoco, billion-context |
| HF papers 2026-11-10 | Hit | Tactile manipulation, WorldFoundry, Sage-Taiyi, Reframe, Longtail-SWE. No builder item promoted |
| HF models trending | Hit | speridlabs/idol, Kimi K3.1 Pro, Ripple-65B Preview, Arete-27B. No new open-weight model from the window |
| pricepertoken.com/news (+ agents, coding, open-source, model-releases) | Hit | Model releases: AlphaProof Code, Nex-N2.6, Inkling 2 Small (free), Fable 5.2, Muse Ember, Nemotron 3.5 Super VL. Promoted Bedrock retrieval modes |
| news.learnprompt.pro | Partial | SPA shell only |
| LearnPrompt data JSON | Hit | latest-24h.json generated 02:27 CST, 320 items. Fable 5.2 in Copilot, Claude Code releases, secret scanning for Actions logs, Control Room (OpenAI page returned 403) |
| radarai.top/en | Hit | GPT-6 Cyber Linux audit claim (no primary opened; dropped), AlphaProof Code (confirmed via Google). No CN-lab release |
| The Agentic Digest | Hit | 11-10 issue: Fable 5.2, AlphaProof Code, Bedrock retrieval modes, Muse Ember, Nemotron on SageMaker, recalc |
| simonwillison.net 2026/Nov/9 | Hit | details-summary post, llm-anthropic 0.31 (adds Fable 5.2), a mysterious-beep post. No further nomination |
Dailies in rank order
- Latent Space AINews: Hit. 11-09 issue loaded: Fable 5.2 retakes #1. Used for Fable context; all numbers re-read on Anthropic, AA and OpenRouter pages.
- AlphaSignal: Partial (some items PRO-gated). Fable 5.2, Dots (no primary), Antigravity knowledge-work view (no primary), Muse Ember, Cursor CursorBench claim (no primary), Scantist GLM-5.3 shrink and UkisAI Swift 1.5 (PRO, dropped).
- Code4U digest: Partial. Page loaded, but the newest entries I found were policy and IPO items. Nothing promoted.
- TLDR AI: Hit (latest 11-09 issue). Promoted OpenRouter Auto Code, Anthropic agent-memory post, Ace. No 11-10 issue yet.
- TLDR flagship: Not used (TLDR AI was not blocked).
- The Neuron: Partial. 11-10 issue is "Anthropic released Claude Fable 5.2", with a Fireworks breach item that was promoted and confirmed.
- AI Daily Brief: Miss. Latest edition is 11-08.
Nominated → confirmed or dropped
| Event | Status |
|---|---|
| Claude Fable 5.2 / Mythos 5.2 | nominated (HN, Agentic Digest, LS, AlphaSignal, PPT) → confirmed (Anthropic post, API docs, AA, OpenRouter) |
| Muse Ember | nominated (HN, PPT, Agentic Digest, AlphaSignal) → confirmed (Meta post, AA article, OpenRouter) |
| Nex-N2.6 | nominated (PPT) → partial (OpenRouter listing only; no HF card or license found) |
| Inkling 2 Small free variant | nominated (PPT) → confirmed (OpenRouter, Thinking Machines post, HF listing check) |
| Strands Decider at Amazon | nominated (HN) → confirmed (Strands blog; authors' own numbers) |
| AWS Nitro Sandboxes | nominated (HN) → confirmed (AWS What's New) |
| Fireworks AI / Notion incident | nominated (HN, The Neuron) → confirmed (Fireworks bulletin, Notion notice) |
| AlphaProof Code | nominated (HN, PPT, Agentic Digest, RadarAI) → confirmed (Google DeepMind blog) |
| Claude Code v2.1.321/322, OpenCode, pi, Hermes, Copilot CLI | nominated (LearnPrompt, GitHub) → confirmed (GitHub releases) |
| Fable 5.2 in GitHub Copilot | nominated (LearnPrompt) → confirmed (GitHub changelog) |
| OpenRouter Auto Code | nominated (TLDR AI) → confirmed (OpenRouter blog; vendor numbers) |
| Exact | nominated (HN) → confirmed (GitHub repo and README) |
| recalc | nominated (Agentic Digest) → confirmed (GitHub repo and README) |
| Ace | nominated (TLDR AI) → confirmed (GitHub Next site, repo) |
| Anthropic agent-memory post | nominated (TLDR AI) → confirmed (Anthropic research page) |
| Bedrock single-shot retrieval | nominated (Agentic Digest, PPT) → confirmed (AWS ML blog) |
| GitHub secret scanning for Actions logs | nominated (LearnPrompt) → confirmed (GitHub changelog) |
| MoE on M1 Max, GPU price spread | nominated (HN) → confirmed exists (author pages; claims are the authors') |
| Cursor "CursorBench 60%" with Fable 5.2 | nominated (AlphaSignal) → no Cursor primary found; dropped |
| OpenAI Dots multi-computer control | nominated (AlphaSignal) → no OpenAI primary found; dropped |
| Google Antigravity knowledge-work view | nominated (AlphaSignal) → no Google primary found; dropped |
| OpenAI Control Room | nominated (LearnPrompt, AlphaSignal) → primary page returned 403; dropped |
| Gemini routing personal files to annotators | nominated (HN 664) → source is an individual's GitHub repo, no Google statement; dropped |
| GPT-6 Cyber Linux audit | nominated (RadarAI) → no OpenAI primary opened, security scope; dropped |
| Qwen4 beats Opus 5.5, Meituan LongCat-2.5, DeepSeek Harness DAU | nominated (LearnPrompt via AIbase, InfoQ China) → no lab primary found; dropped |
| Scantist GLM-5.3 shrink, UkisAI Swift 1.5, Xenova BGE-Small, NVIDIA Cosmos3-Nano | nominated (AlphaSignal PRO) → no primary opened; dropped |
| Opus 5.5 nerf reports (11-09 brief) | no new facts in the window; not repeated |
Engine blogs: not nominated