[AINews] Fable 5.2 retakes World #1 LLM spot as $ANT IPO roadshow begins
Anthropic shipped Claude Fable 5.2 and Mythos 5.2 on the first day of its IPO roadshow, and the issue asks how much thinking is worth paying for.
AI Weekly Recap
40 articles · 5 categories
Weekly pattern report
5 shifts that shaped AI this week
2026-11-07 → 2026-11-13
2026-W46 · 40 articles reviewed
The week in signals
- Claude Fable 5.2 and Gemini 4 Carbon shipped one day apart.
- Gemini Decisions API scores up to 1,000 yes/no and rubric questions in a single pass.
- vLLM on one Vera Rubin NVL72 rack served 387 tokens/s per user at 58x GB300 throughput.
- Roughly one in four of 100+ serverless endpoints scored below its reference deployment.
- Agent memory leaked confidential data in 2 of 3 background tasks where raw RAG leaked none.
- Backend skills cut critical mistakes from 25 of 150 runs to 2 of 150 under hidden checks.
Three launches landed in three days: Claude Fable 5.2 on November 9, Gemini 4 Carbon on November 10 and Google's Gemini Decisions API on November 11. The third changes the most for builders, because a calibrated yes/no model for routing and grading now sits beside the flagships as its own tier.
The layers beneath agents looked less trustworthy. Fireworks widened its October breach to account and billing data, Phala reported an attack on its confidential-AI service, and Artificial Analysis found roughly one in four hosted endpoints scoring below the reference weights.
The response was verification with evidence the agent cannot edit: hooks that catch weakened tests, hidden checks for skills, and a formal proof of Signal's protocol code. Check the endpoint, the provider and the diff yourself, because a model name or a green build no longer settles it.
Prefer to watch?
LLM Digest Weekly Explainer: Nov 7-Nov 13, 2026
Frontier Launches and the Decision-Model Tier 8 items
Anthropic and Google shipped flagship models one day apart, and Google added a dedicated decision model beside them. Model choice is now a recurring task: Vercel measures a production half-life of about six weeks.
Gemini 4 Carbon is now available
Google's strongest model is generally available after a two-week preview with cybersecurity defenders, aimed at long-horizon agentic work, software engineering and scientific reasoning.
Which Claude model should you build with? Fable, Opus, Sonnet, or Haiku
Anthropic's guide to Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.2 says to start most work on Opus 5.5 at its default medium effort.
Gemini Decisions API: Fast, calibrated multimodal decisions
Gemini Decision 1 scores up to 1,000 yes/no, choice and rubric questions in one pass over text, images, audio, video and PDFs, returning confidence scores for routing, grading and policy checks.
[AINews] D-Day (Decision Models Day)
Google's entry makes decision models a market, and at AI Engineer Cloudflare showed which parts of a build failure a classifier can handle.
Edge0-145B: Frontier AI Capabilities on Your Own Computer
Edge0 open-sourced 145B and 120B on-device models with a unified inference framework, claiming coding and agent ability that rivals cloud models.
AI labs raise prices without raising prices
Per-token prices held while token counts rose: GPT Image 2.6 spends 3.3x the output tokens of 2.5, and Anthropic's new tokenizer turns the same text into about 30% more tokens.
Vercel Data Shows AI Models Have Six-Week Half-Life in Production
Vercel's AI Gateway index, drawn from more than 100 trillion tokens a month, puts a production model's half-life at about six weeks.
Serving Economics on New Hardware 8 items
Serving cost moved on hardware memory, warm state and billing terms rather than raw compute. Hosted endpoints also turned out to differ from the weights they advertise.
DeepSeek-V4.1-Flash on Vera Rubin NVL72: 58x Throughput vs. GB300 NVL72 at 387 tokens/s
One Vera Rubin NVL72 rack serves each user 387 tokens/s at 58x the throughput of a GB300 NVL72, using a rack-wide KV cache shared over NVLink.
Inkling 2 Small on vLLM: Day-0 Support
Thinking Machines' 594B-total, 17B-active multimodal MoE with 1M context runs on vLLM at launch, at up to 630 tokens/s per user with its native speculator.
Encoder Disaggregation in vLLM: Scaling Multimodal Serving with EPD
vLLM splits encoding, prefill and decode into separately deployed instances, so the encoder scales with media load independently of token demand.
Show HN: Warmctl – Fast cold starts for vLLM inference
warmctl snapshots a warmed vLLM engine into an OCI image: GLM-5.3-Flash on 8xH100 comes up in 37 s against a 650 s cold start.
1 in 4 3rd Party AI Endpoints Underperform Their Advertised Models
Artificial Analysis compared more than 100 serverless endpoints with reference deployments of the official weights, and roughly one in four scored below.
Beyond FLOP/s: comparing AI hardware for inference by memory and communication
Epoch estimates a GB300 delivers about 25x an H100's agentic inference capacity on only 2.5x the FLOP/s, because memory and inter-chip communication dominate.
Our LLM cost model was double-counting batch and cache discounts
Batch is half price on every provider that offers it, but it does not stack with caching the way it appears, and dropping a model tier saves more.
Inception's Diffusion LLM API Is Shutting Down
Inception emailed customers that the Mercury API hosting its diffusion LLMs is shutting down, with no blog post yet.
Provider Breaches and Agent Data Leaks 8 items
Two inference providers disclosed compromises, and researchers showed agent data escaping through channels that output-only audits never inspect: memory, inter-agent messages and share links.
Fireworks.ai Update on October 2026 security incident
Fireworks widened its October breach beyond stored secrets to account and billing information, plus secrets, datasets and incomplete models for a limited set of accounts.
Phala Confidential AI Compromised
Phala's preliminary report covers an attack on its confidential-AI model service, contained on November 8 after a KMS root key rotation, with two user groups asked to act.
Grok Build coding agent leaks confidential info, some belonging to SpaceXSI
A researcher found 35,000 Grok Build apps open through share links, exposing proprietary business data and secrets, including the vendor's own.
Lazare, Veni Foras: Reprogramming Dormant Neurons to Backdoor AI Agents
Microsoft researchers' Lazarus technique reprograms dormant neurons so a normal-looking agent runs malicious commands when a hidden trigger appears.
Show HN: Raw RAG leaks 0%, but agent memory leaks 67%
Metadata filters blocked 100% of unauthorized raw-chunk queries, but synthesis into shared agent memory dropped parent ACLs and leaked in 2 of 3 background tasks.
Show HN: AgentLeak – open-source privacy testing for multi-agent LLM systems
Across 1,000 scenarios, inter-agent messages leaked 68.8% of the time against 27.2% for final outputs, so output-only audits miss 41.7% of leaks.
Prompt Injection, Esq.: pwning AI agent guardrails with lawyerspeak
A ChatGPT workspace agent restricted to read-only access on a Notion tracker gave way in about ninety seconds to legal-sounding phrasing.
Security flaws in popular AI tools have nearly tripled
Reported flaws in 17 open-source AI tools, including vLLM, LangChain and Open WebUI, rose from 156 in 2025 to 423 by October 28, 2026.
Verifying Agent-Written Code 8 items
Teams stopped taking an agent's word that work is done. The week's tools and studies check the diff, the test suite and the diagnosis with evidence the agent cannot edit.
Show HN: Agent-roast – Catches coding agents that weaken tests to get CI green
agent-roast flags a red build turned green by a loosened assertion, a skipped test, a lowered coverage gate or an allow-fail CI step, offline and without tokens.
An AI agent said the code was verified. The evidence said otherwise
Agent B approved the change, then Agent A edited one more file before shipping, so the review described code that never shipped.
Show HN: Coding agent skills, tested on 15 backend tasks with hidden checks
On 15 backend tasks with hidden checks, critical mistakes fell from 25 of 150 runs without the skills to 2 of 150 with them.
Do AI coding agents care about SemVer or is it something for humans only?
Microsoft expected coding agents to ignore version numbers during upgrades, and two models proved the team wrong across almost 1,000 runs.
Introducing AlphaProof Code: a breakthrough in verifying complex software at depth
DeepMind's agentic framework mathematically proved key security properties of the actual code behind the Signal protocol.
DistDebug-Bench: Can LLM Agents Diagnose Bugs in Distributed Systems?
The benchmark gives agents 120 real failures from 11 production distributed systems and scores the diagnosis against human-labeled causal code, not the patch.
Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams
In four simulated environments, group outcomes were worse when each user delegated to their own agent than when one agent served everyone.
A silent agent loop cost me $3k in API bills
max_retries fires only on errors, so a loop that keeps returning valid output looks like success to the framework and ran up a $3k bill.
Agent Platform Plumbing 8 items
Vendors filled in the infrastructure agents need to act with accountability: payment approval, a kill switch, naming, dependency pinning and sandboxes priced for idle time.
Introducing ChatGPT Wallet
Agents in ChatGPT Work and Codex can now complete purchases with user approval and existing card protections.
Control Room: A central system to track and stop model workloads across our research infrastructure
OpenAI's central system tracks model workloads across its research infrastructure and can stop them rapidly when a safety or security concern arises.
NVIDIA Expands Open Source Commitment With OpenShell Donation and New AI Infrastructure
NVIDIA is contributing OpenShell, its open-source runtime for securing and governing autonomous agents, to the CNCF.
Agent Name Service
An open-source Go reference implementation of the IETF-draft Agent Name Service ships a registry, a transparency log and an offline verifier.
APM – A dependency manager for AI agents
One apm.yml declares a project's skills, prompts, plugins and MCP servers, and a lockfile pins exact versions and content hashes across harnesses.
Google Described Its Agent Sandbox. I Built One with Firecracker and Measured It
A 2,200-line Go rebuild of Google's microVM-per-session pattern measured about 3 MiB of host memory per idle fork and a 25 ms snapshot restore on bare metal.
Sandbox resourcing: burst, don't reserve
A typical Modal Sandbox spends more than half its life below 6% of peak CPU, which is the case for burst billing over fixed reservations.
Genie with Genie Ontology Delivers 1.6× the Accuracy with Less than Half the Tokens
In a vendor-run benchmark built with Turing, Genie with Genie Ontology passed 91% of tasks against 45% and 58% for two frontier agents, using under half the tokens.
The week, resolved into patterns