All postsROUNDUP3 min read
Best of the Week: Claude Fable 5.2's Lower Task Costs and 4 More AI Tools
This week's five topics: Claude Fable 5.2, Claude Code secrets, Phone Use, Claude Dreaming, and Brave's x402 agent search.
This week, I'm looking at agents getting cheaper and more capable, and the setup checks that come with giving them more access. These are the five Shorts I'd send an engineer comparing task costs, checking saved permissions, or connecting agents to phones, memory, and paid search.
Claude Fable 5.2: Lower Task Cost in Anthropic's Test?
Claude Fable 5.2 launched November 9 with API ID claude-fable-5-2 at $8 input and $40 output per million tokens, 20% cheaper than Fable 5.1 and double Opus 5.5's token prices. Anthropic reports 45.3% on FrontierCode 2.0 at High effort versus Opus 5.5's best 45.5%, for roughly half the cost per task, and says output is up to 30% faster than Fable 5.1. If you're choosing a coding model, I'd test it on your own workload: these are Anthropic's early results, and the useful comparison is what a completed task costs you.
Full breakdown with sources: 500xengineer.com/blog/claude-fable-5-2
Claude Code's "Don't Ask Again": Secrets in 1,122 Public Files
Apiiro's November 5 report found that Claude Code's “Yes, and don't ask again” saves the exact approved command as an allow rule in .claude/settings.local.json, including any password in that command. In its November 1–3 scan, Apiiro found 104,185 public copies of the file and classified 1,122 as holding real secrets; gitleaks and TruffleHog both missed 215 of those files. If you use Claude Code, I'd search repositories and history, rotate every credential the file held, untrack it with git rm --cached, add it to each repository's .gitignore, and reference secrets through variables.
Full breakdown with sources: 500xengineer.com/blog/dont-ask-again-leak
Claude Code Can Now Tap Through Your Real iPhone (No Jailbreak)
Phone Use, Aleksey Smolenchuk's free MIT-licensed alpha first released November 10, lets Claude Code, Codex, and other agents read an unmodified iPhone's screen through AirPlay Screen Mirroring and tap, type, and swipe through Bluetooth mouse and keyboard input, using a CLI and MCP server without a jailbreak or cable. The Mac setup needs macOS 15+ and an ESP32 LilyGO T-Display bridge board, Linux can use the computer's own Bluetooth, and the iPhone app needs iOS 17.2+. If you want an agent to operate your iPhone, I'd check those requirements first and review the access you're granting: you're giving an agent access to your phone.
Full breakdown with sources: 500xengineer.com/blog/phoneuse
Claude Dreaming: Anthropic's Fix for Stale AI Agent Memory
In its November 6 post “The components of agent memory,” Anthropic describes dreaming as Claude rereading past session transcripts, checking memory notes against what happened, and rewriting them; the Claude API's Dreams feature in Managed Agents takes a memory store plus 1 to 100 past sessions and writes a new store without changing the input. Anthropic reports that a wiki built over about 350,000 files rose from 13% to 48% of graded criteria over 128 dreaming sessions. If you're maintaining agent memory, I'd use that as a reason to evaluate transcript-based revision, and keep Claude Code's MEMORY.md as an index because only the first 200 lines or 25KB load at session start.
Full breakdown with sources: 500xengineer.com/blog/claude-memory-dreaming
Brave Search: AI Agents Pay With x402, No Account Needed
Brave announced x402 support for its Search API on November 12, letting agents pay per request in USDC on Base without an account, plan, or API key; an unpaid request returns HTTP 402 with the price. Our own unpaid test on November 13 got a 402 quoting $0.005 per search, or $5 per 1,000, matching the regular rate. If you're adding search to an agent, I'd check that its wallet has USDC on Base and look at Stripe's purl CLI, which Brave points to; Answers, Autosuggest, and Spellcheck still require an API key.
Full breakdown with sources: 500xengineer.com/blog/brave-x402-agent-search
Daily email
Get each breakdown in your inbox.
One email each evening, right after the Short goes live: the AI tool it covers, the full write-up with sources, and how to get paid for it. Free, and you can unsubscribe anytime.
Sources
- Claude Fable 5.2: Pricier Tokens, Lower Task Costs in Anthropic's Tests: https://www.500xengineer.com/blog/claude-fable-5-2/
- Claude Code's Don't Ask Again Settings Can Expose Passwords on GitHub: https://www.500xengineer.com/blog/dont-ask-again-leak/
- Claude Code Can Use Your Real iPhone. Here's the Catch: https://www.500xengineer.com/blog/phoneuse/
- Claude Dreaming: How Anthropic Fixes Stale AI Agent Memory, and the Catch: https://www.500xengineer.com/blog/claude-memory-dreaming/
- Brave Search Lets AI Agents Pay With x402 Without an Account: https://www.500xengineer.com/blog/brave-x402-agent-search/