Claude Code Daily Briefing - 2026-11-11

Release Summary

v2.1.322 arrived just one day after v2.1.321 (11/9), but the only change is that fast mode now works in Vertex AI’s global, US, and EU regions, falling back automatically to standard speed whenever Vertex declines the request. Since v2.1.321’s main features and security fixes were already covered in the 11/10 briefing, today’s briefing focuses on two big announcements Anthropic made directly — the launch of Claude Fable 5.2 and Mythos 5.2, and the unveiling of Claude Whiteboards, Sheets, and Data — along with GeekNews discussion and service status.

Full release notes

New Features & Practical Usage

Claude Fable 5.2 and Mythos 5.2 are now generally available, marking the debut of Mythos as a dedicated scientific research model (11/11)

The 11/10 briefing only mentioned that v2.1.321 added Fable 5.2 as the default Fable model and that it scored 65 on the Artificial Analysis Intelligence Index — but Anthropic’s own announcement goes well beyond that and introduces Mythos 5.2 for the first time. Both models come from the same underlying model but differ in who can access them. Fable 5.2 ships with cybersecurity and biosecurity safeguards and is available to everyone, while Mythos 5.2 is restricted to vetted institutions doing security or life-sciences research. Anthropic recommends starting most coding and knowledge work with Opus 5.5, and reaching for Fable or Mythos only on the hardest tasks.

# Pricing: $8 per million input tokens, $40 per million output tokens (lower than 5.1)
# Fast mode: 2x the price ($16 in / $80 out) for up to 2.5x faster token generation
/model claude-fable-5-2

Anthropic says a Fable 5.2 agent formalized Fermat’s Last Theorem in Lean in about 30 hours — a sharp drop from the 11 days a prior internal attempt took with a dedicated harness and human guidance. The announcement also describes coordinating more than 1,000 subagents over 8 days to build an operating system prototype, and three pieces of electronics built from single-paragraph requests that worked correctly without any circuit fixes. Mythos 5.2, meanwhile, developed a method for predicting adeno-associated virus (AAV) capsid structures, accurately modeling 28 of 34 cases.

If you’re already using Claude for work that demands deep reasoning — complex math proofs, scientific research, hardware design — it’s worth following Anthropic’s own recommendation rather than routing everything through Fable: start with Opus 5.5, and reserve Fable and Mythos for the genuinely hard cases. Anthropic’s announcement · GeekNews

Claude Whiteboards and Sheets launch in beta, and Claude Data goes GA (11/11)

Claude Whiteboards, a team collaboration board, and Claude Sheets, a spreadsheet tool, are now available in beta, while Claude Dashboards — which started in beta back in October — has graduated to general availability as Claude Data. Whiteboards is a real-time board Claude and your teammates can edit together: turn a rough plan into a timeline with owners and dates, or pull action items out of meeting notes, then export the finished board as a doc or presentation, or send it to Lucid, Miro, or Vani by Zoho. Sheets lets you build spreadsheets from a prompt — tabs, formulas, formatting, and dropdowns included — highlights the cells Claude changes for easy change-tracking, and exports to .xlsx or CSV when you need full spreadsheet functionality. Data lets you query sources like Redshift, BigQuery, and ClickHouse, and adds certified queries (beta) so teams can approve a shared definition of a metric, plus live sharing (beta), which surfaces only pre-approved query results.

Team plan         → all three features on by default
Enterprise plan   → Whiteboards and Sheets off by default; Data turns on by default starting 11/19
Where to change it → Organization settings > Artifacts

If your organization has been bouncing between separate SaaS tools for whiteboarding, spreadsheets, and data analysis, this launch is a good excuse to handle the lightweight versions of that work directly inside Claude, and only hand off to specialized tools like Lucid, Miro, or Excel once you need serious editing power. Claude’s announcement · GeekNews

Fast mode now works in Vertex AI’s global, US, and EU regions (v2.1.322)

Fast mode used to be limited to a handful of regions; as of this release it also works in Vertex AI’s global, US, and EU regions, automatically falling back to standard speed if Vertex rejects the request.

If you’re running Claude Code through Google Cloud Vertex AI, it’s worth trying /fast right away in those regions — no extra configuration needed. Full release notes

Developer Workflow Tips

Google Chrome’s modern-web-guidance skill lifted Claude Code’s eval pass rate from 57% to 88% (11/11)

This skill nudges coding agents away from the stale JavaScript patterns baked into their training data by searching a CLI-indexed guide of roughly 250 web platform features and 203 usage examples, then feeding the relevant results into context. It searches offline using embeddings and BM25, with no network calls or API keys required, and uses Baseline data to suggest fallback implementations that match current browser support. In a Playwright-graded before/after comparison, Claude Code’s pass rate rose from 57% to 88%, Codex CLI’s from 54% to 82%, and Antigravity CLI’s from 58% to 94%.

# Install the modern-web-guidance skill in your project
npx modern-web-guidance

If Claude Code has been writing outdated polyfills or needlessly heavy JavaScript, it’s worth installing this skill to steer it toward current CSS, DOM, and accessibility features, then checking whether your own pass rate actually improves. GeekNews

When adopting a community skill into your project, use keep / bind / rewrite / drop / review to record your reasoning (11/10)

Claude Code and Codex skills pulled from GitHub are written against the original author’s project, so they often assume tools your project doesn’t have or reference test commands that don’t match yours — skilladopt addresses this by walking through the skill paragraph by paragraph and recording a keep (preserve as-is), bind (wire it in), rewrite, drop, or review (needs a closer look) decision along with the reasoning behind it. When your project changes, it flags only the lines affected by that change; it never executes files from inside the skill, reading only a pinned commit instead; and the agent making the calls runs inside a read-only sandbox.

npx skilladopt setup
npx skilladopt anthropics/skills frontend-design
npx skilladopt impact   # see only the lines worth re-checking after a project change

If you’ve ever hand-edited someone else’s skill and then lost track of which lines needed another look every time a teammate changed a test command or rule, this tool is worth trying — it makes the reasoning behind each decision, and the lines it affects, traceable. GeekNews

Security & Limitations

Claude service status — a roughly 4-minute error spike for Fable 5.2 and Mythos 5.2 on 11/10, all systems normal as of 11/11

Checking status.claude.com directly shows a roughly 4-minute incident on 11/10 where error rates rose for Claude Fable 5.2 and Mythos 5.2 requests (since resolved). 11/9 showed “No incidents reported,” and as of today (11/11), claude.ai, the Claude API, Claude Code, Claude Console, Claude Cowork, and Claude for Government are all showing “All Systems Operational.” Earlier, on 11/8, there was a roughly 14-minute period of degraded performance on Fable 5.1.

Brief error spikes right after switching to a new model aren’t unusual, so a short-lived incident like the one on 11/10 was likely resolved just by retrying — if you run into trouble today, it’s more sensible to check your local setup or network first. Claude Status

Ecosystem & Plugins

Show GN: Herdr CKD — an open-source AI coding orchestrator bundling 13 MCP tools for Claude Code (11/11)

This orchestrator tackles the problem of coding agents declaring a task “done” while build errors or missing requirements remain, not through prompt instructions but through physical permission separation and git worktree isolation. A fixed coordinator (Core C) that can’t edit files or run commands assigns work and issues the final PASS/FAIL verdict; implementation threads (A1 through An) work in their own worktrees; and audit threads (B1 through Bn) perform read-only reviews based solely on git status and git diff --stat output — never the implementer’s own explanation. The key idea is that auditors never see the implementer’s self-justification, which cuts down on tasks passing purely because the agent talked its way through. It also supports hybrid routing, mixing a cloud model like Claude for the coordinator with local models from LM Studio or Ollama for the implementation and audit threads.

Core C (coordinator, no read/write) → A1-An (per-worktree implementation) → B1-Bn (read-only, git-diff-based audit) → PASS/FAIL

If you’ve been running Claude Code with accept-edits or auto mode across multiple concurrent agents and found it hard to trust an agent’s own “it’s done” report, this structure is worth a look for physically separating implementation from audit. GeekNews

Val Town is shutting down in phases after losing ground in the cloud computing market (11/11)

Val Town stops billing on 11/10, switches to read-only mode on 12/10 (except for sqlite and blob storage), and shuts down completely on 1/10/2027. Its source code is slated to go open source on February 1, 2027. The team also published migration guides to Cloudflare, Vercel, Netlify, and Render, saying they expect users to lean on current LLMs to help with the move, on the premise that today’s models handle this kind of migration work well.

If you’re running an app on Val Town, you’ll need to finish migrating before January 10, 2027 — and this is also a good opportunity to consider having Claude Code automate the move to another hosting environment. GeekNews

Community News

• The Fable 5.2 launch sparked suspicions of an Opus 5.5 regression and benchmark skepticism (11/11): Some users claimed Opus 5.5 quietly regressed before the Fable 5.2 launch, but another developer who ran their own benchmarks said they found no clear degradation. Commenters also pointed out that comparisons with Carbon (an unreleased model) mixed medium and xhigh settings, that Carbon was missing from the cybersecurity benchmark, and that some Artificial Analysis results were left out entirely. GeekNews
• The Claude Whiteboards, Sheets, and Data launch raised concerns about Anthropic encroaching on SaaS territory (11/11): Some commenters worried about the trend of Anthropic pulling more SaaS tools into Claude and said they’d prefer the market stay competitive, while others predicted the features are too basic to replace dedicated tools like Figma and will likely fizzle out within six months. Others noted that handing enterprise data entirely over to Anthropic would require solid access controls, compliance, and data sovereignty guarantees. GeekNews

Minor Changes

All of the following are v2.1.321 (11/9) items not covered in the previous briefing.

• Fixed a bug where the PreToolUse mcp_tool hook for stdio and HTTP MCP servers would be skipped if the response exceeded 16MiB — now the tool call itself is rejected instead.
• Fixed a bug where a read-only command written with a variable, like grep -r foo "$PWD/src", triggered an unnecessary permission prompt that the same command with a literal path wouldn’t.
• Fixed a bug where flags like --permission-mode, --dangerously-skip-permissions, --model, --effort, and --agent were ignored on the claude agents command whenever another flag such as --verbose was also present.
• Claude now refuses to send files whose names hide directional or control characters that make them display differently in the Claude app.
• Fixed high memory usage in long-running sessions that repeatedly read large files line by line.
• Fixed a bug where the git status passed to Claude at session start was misread as “clean” if git status itself failed — it’s now reported as “unknown” instead.
• Notifications for background commands over 1,000 characters, and TaskStop results, now show Claude only the start and end instead of the full text, saving context.

Recommended Reads

• Using AI makes codebases fragile (11/11): The author, reflecting on their last few months at Meta where roughly 90% of their code was AI-generated, says productivity rose only about 2x rather than the promised 10x, and argues codebases become fragile in four ways: shallow tests, excessive complexity, eroding understanding, and reviews that can’t keep pace. A counterargument attached to the piece says clear specs and explicit “don’t do this” guidance dramatically improved agent output, suggesting the real problem is a lack of standards rather than AI itself. GeekNews
• The web is more than enough (11/9): The author argues that as the encryption, 3D, and audio capabilities already built into browsers get easier to use, a single person can now build what used to take a large team — and backs it up with their own office tool, which drew hundreds of thousands of visitors at zero hosting cost. A comment noting that Claude actually wrote the code for the author’s browser-based 3D game, Battle Drones, adds an interesting wrinkle to what “built by hand” really means. GeekNews
• Flide — speeding up Jev-style decision model inference by up to 9.5x (11/10): An open-source runtime that processes shared input once, then forks the model state per question to keep answers identical while cutting inference time — on a single H200, it ran 8 questions against 10,000 emails in 28 seconds, 4.1x faster than the fastest comparison engine. A comment notes the speedup is largest when asking many questions against shared input, but smaller for the more common pattern of asking the same question across many different inputs — a reminder, much like the Hightouch Jev hybrid-routing case from the 11/7 briefing, that teams running cheap classifier models should check whether their own query pattern actually fits before adopting it. GeekNews

Interesting Projects & Tools

• Show GN: flickbook — replays AI-made changes step by step using Claude Code hooks (11/11): When an agent finishes a task, the diff usually shows up as one large blob that’s hard to trace back to individual decisions. flickbook fixes this by marking a starting point with flickbook start, recording steps with flickbook step (done automatically if you use Claude Code hooks), and then letting you walk through each step’s request, actual diff, and related concepts in a local web viewer via flickbook replay. It only reads files and never sends any data out, and unlike Claude Code’s checkpoints, which exist for reverting changes, it’s built purely for revisiting changes at any point.

npx flickbook --help

GeekNews