| ← All issues | |
| The Diff No. 042 Saturday, 14 November 2026 · AI dev news, only what is verified |
 |
Anthropic’s roughest reliability day in months - two resolved incidents on Friday, a third still recovering as this issue ships - lands next to Perplexity publishing the training recipe for Decider, the model that answers questions about state instead of generating text. Between 09:58 and 15:38 UTC on Friday, Anthropic ran two separate incidents across claude.ai, the API, Claude Code, connectors and Managed Agents - and its own status text says scheduled routines ran more than two hours late, with some runs skipped outright. A third incident, hitting purchases and plan changes, opened at 02:54 UTC Saturday and was still in Monitoring when we sent. Perplexity, meanwhile, put out Decider v1.3 and a distilled Flash sibling: a classification model on a Qwen3.8-27B base that reads text, JSON or images and returns typed, probabilistic answers - with the blog post, the model card, and a new Locate feature that extracts verbatim spans with character offsets. Reliability is the story twice over: the agents you scheduled may not have run, and the routers inside your stack are becoming products with published recipes - and free tiers. |
IN THIS ISSUE | 01 🚨 Anthropic went down twice in a day - and some scheduled runs never happened Two resolved Friday incidents hit claude.ai, the API, Claude Code, connectors and Managed Agents; a third, affecting purchases, was still recovering Saturday morning. All times and quotes from Anthropic’s own status page. |  | 02 🔀 Perplexity publishes how it trains Decider v1.3 - a model that decides instead of generates Decider v1.3 and a distilled Flash variant read text, JSON or images and return typed, probabilistic answers; a new Locate feature extracts verbatim spans with character offsets and confidence scores. Built on Qwen3.8-27B. |
|
🔥 VIRAL The loud ones on Hacker News. Numbers are a morning snapshot. | 01 status.claude.com, Nov 13-14 · HN 460 points, 372 comments at the time 🚨 Anthropic went down twice in a day - and some scheduled runs never happenedTwo resolved Friday incidents hit claude.ai, the API, Claude Code, connectors and Managed Agents; a third, affecting purchases, was still recovering Saturday morning. All times and quotes from Anthropic’s own status page. SCENE Picture the developer who wired a claude.ai routine to triage the night’s bug reports before standup. Friday morning, standup comes, and the triage simply does not exist - no error, no retry, no output. Anthropic’s own words, in the 15:33 UTC update: "Scheduled tasks and routines can run more than two hours late and some runs have been skipped, and Claude Security scans, reviews and project setup remain delayed." THE THING The verified timeline, all from status.claude.com. Incident one, 09:58-11:58 UTC: users could not authenticate to claude.ai and Claude Code; a fix deployed at 11:05, resolved 11:58 - claude.ai and Claude Cowork affected. Incident two, 13:09-15:38 UTC, broader: claude.ai plus mobile apps, the Claude API, Claude Code, connectors including Docs and Sheets, Claude Managed Agents, Claude in Meetings and Claude Security; new connectors were temporarily unavailable during recovery. Incident three opened Saturday 02:54 UTC: elevated errors on purchases, plan upgrades and cancellations; at 03:09 UTC Anthropic moved to Monitoring, warning that some recent payers "may need to wait up to an hour for their plan to update. There is no need to pay again." Separately, the Microsoft 365 connector outage that began Thursday was a Microsoft service incident, per Anthropic’s status text; Anthropic marked its side resolved at 16:40 UTC Friday while noting Microsoft is still completing recovery. Anthropic has stated no cause for the Friday incidents. ✂️ STEAL IT Steal the operational habit: a cloud agent that runs on a schedule is only as reliable as the provider’s worst day, and yesterday was one. After any provider incident, check the run history yourself - Anthropic said runs were skipped and nowhere promised they would be retried. And write routines to be idempotent, so re-running one by hand after an outage is boring instead of dangerous. Watch VERDICT Nothing to patch in your code today. If you depend on claude.ai scheduled tasks, routines, or Security scans, assume Friday’s runs need verifying by hand. Watch for a post-mortem: three incident banners in roughly 19 hours, the day after Code Mode shipped, is timing Anthropic will be asked about - but timing is all it is until Anthropic says more. Claude status: the incident history · Microsoft 365 status (the connector incident) |
 02 Perplexity blog + Hugging Face model card, Nov 12 · HN 107 points at 06:27 IST 🔀 Perplexity publishes how it trains Decider v1.3 - a model that decides instead of generatesDecider v1.3 and a distilled Flash variant read text, JSON or images and return typed, probabilistic answers; a new Locate feature extracts verbatim spans with character offsets and confidence scores. Built on Qwen3.8-27B. SCENE The model card’s own example is a support queue: a customer writes "My integration keeps failing with a 401 since we rotated our API key." One call returns three typed answers at once - team: support, urgent: true, frustration: mildly annoyed - no generation, no parsing, no prompt. That is the whole pitch: routing and triage are classification problems, so Perplexity built a classifier and now publishes how it trains and evaluates it. THE THING The facts, all from Perplexity’s own materials. Decider v1.3 is a multimodal classification model on a Qwen3.8-27B base, succeeding v1, v1.1, v1.2 and Flash v1.2; Flash v1.3 is distilled from it. It takes a state (text, JSON, or images) plus questions of three types - choice, noul, and score - and returns typed answers. New in v1.3: Locate, which extracts verbatim text spans from the state with character offsets and confidence scores, which is what an audit trail for automated triage needs. Self-hosting wants a CUDA GPU with about 49 GiB for BF16 weights plus working memory - and the Hugging Face repository is private, gated behind an access token, which directly contradicts the "open-weight" framing in some third-party coverage. For everyone else, OpenRouter lists a free tier. This is the second decision-model release in a week, after Google’s Gemini Decision 1 - which scored 66 on Artificial Analysis’ Decision Index in yesterday’s issue; Decider is not on that index yet. ✂️ STEAL IT Steal the pattern for your own queues: anywhere you route, tag or score with a prompt today, a decision model gives you typed outputs with probabilities instead of prose to parse. The three question types cover most triage logic, and Locate’s character offsets mean every automated decision can cite its evidence span - build that citation habit now, before your queue is too big to audit by hand. Try today VERDICT Decider is on OpenRouter’s free tier, so pointing one real triage queue at it costs an afternoon, not a budget. It fails you if you need self-hosting this week: the weights sit behind HF access approval and want serious VRAM, and there is no independent benchmark of it yet - Perplexity’s evals are Perplexity’s. Perplexity: the Decider post · Hugging Face: the model card · OpenRouter: Decider V1.3 (free) |
📊 THE ATTENTION GAP HN points at 06:27 IST. Attention, not quality. | FAA ground stop, lifted after 4.5h | | 1406 HN | | Anthropic outage day | | 460 HN | | Perplexity Decider v1.3 | | 107 HN |
|
ALSO ON THE RADAR Claude Code’s docs now lay out five ways to run agents in parallel. The "Run agents in parallel" page at code.claude.com compares subagents, agent view (claude agents, research preview), agent teams (experimental, off by default), Projects (public beta on Pro, Max and Ultra, with cloud threads that keep running when your machine is off) and dynamic workflows - plus a /batch skill that splits one large change into 5 to 30 worktree-isolated subagents, and cross-session messaging between your sessions. The docs are undated; the HN discussion (66 points) is from overnight. Claude Code docs: run agents in parallel → Claude Code 2.1.325 shipped overnight with a security fix - the settings.local.json fix is still absent on day 8. The changelog’s own headline: "Fixed a security issue in how zstd-compressed responses are decoded." Also fixed: permission-rule startup warnings being marked as seen when scrolled out of view, and /voice now naming the settings source - managed, project or settings.local.json - that holds a value. No 2.1.324 is listed, and nothing addresses secret scanning or the settings.local.json exposure a week after the HN comments attributed to Claude Code’s head promised prototypes. Claude Code changelog → Brave Search API now takes pay-per-request agentic payments over x402. No account, no API key: an unpaid request gets an HTTP 402 with the price, the agent signs a USDC payment on Base, and retries with the payment attached. Brave runs its own facilitator; LLM Context and the web, news, image, video and local endpoints all price at $0.005 per request ($5 per 1,000), with Base gas fees covered for a limited time. The suggested path is Purl Wallet, a CLI wallet Stripe develops. All mechanics and pricing from Brave’s Nov 12 post. Brave: agentic payments via x402 → |
🧪 ONE THING TO TRY TONIGHT Five minutes: did your agents actually run yesterday? If you use claude.ai scheduled tasks or routines, open the run history and check Friday’s runs against the incident windows - 09:58-11:58 UTC and 13:09-15:38 UTC. Anthropic said some runs were skipped; it did not say they would be retried. If a skipped run was supposed to produce something someone is waiting on, send it late with a note rather than letting it vanish. And while you are in there, make each routine safe to run twice: same input, same result, no duplicate side effects - the next outage is already scheduled, you just don’t know the date. |
What this issue could not see. Every incident fact - times, affected surfaces, quotes - is from Anthropic’s own status page; Anthropic has stated no cause for the Friday incidents, and their timing a day after the 2.1.323 Code Mode release is timing, not evidence of cause. The Microsoft 365 connector incident was Microsoft’s service incident, per Anthropic’s status text, and Microsoft had not completed recovery at send time. Perplexity’s Decider claims - performance, training, evaluation - are Perplexity’s, from its blog and model card; the Hugging Face repository is private, so "released" here means documented and API-available, not downloadable weights, and third-party "open-weight" framing contradicts the model card. Brave’s pricing and x402 mechanics are from Brave’s own post. The parallel-agents capabilities are from Anthropic’s undated docs; the HN attention is from overnight. Two Nvidia items are deliberately not covered: the report that a director sold $946M in stock weeks before the resignation (Moneywise/Yahoo, 66 HN points) could not be verified against a primary Form 4 this morning, and the "engineered scarcity" piece (73 HN points) is an allegation. The FAA ground stop was lifted after 4.5 hours, per NBC; the FAA attributed it to a telecommunications outage, per the Washington Post - off this newsletter’s beat unless a verified systems cause emerges. VIDEO OF THE DAY is omitted for the ninth day running: no fresh, relevant and verifiable video on today’s stories could be confirmed. Reddit remains absent by policy; X blocks anonymous access and was not consulted. Deduped against every published issue, No. 001-041, via the kit shipped-index and the live home archive: the outage day, Decider v1.3, the parallel-agents docs, 2.1.325 and Brave x402 are new; Claude Code 2.1.323 Code Mode and the AA Decision Index result were No. 041 and are referenced, not re-reported. |
| Images: credited in each report; source screenshots and announcement artwork, not our tests. |
|