AI DAILY / DEV
←
MONDAY
April 20, 2026
→

    Anthropic Launches Claude Design
    • New Anthropic Labs product for prototypes, slides, one-pagers, and mockups.
    • Onboarding reads your codebase and design files to auto-build a team design system.
    • Powered by Opus 4.7; research preview for Pro/Max/Team/Enterprise.
    • Exports to Canva, PDF, PPTX, or standalone HTML.
    • HN reaction thread: 375 points, 243 comments.
    tools anthropic.com

    Opus 4.7's 'Stealth Price Hike': New Tokenizer Burns 35–47% More Tokens
    • Per-token price unchanged ($5/$25 per million) but the new tokenizer produces ~37.4% more tokens per request (483-submission sample at tokens.billchambers.me).
    • Anthropic's migration guide cites 1.33× for Claude Code content, 1.45× for CLAUDE.md, 1.47× for technical docs.
    • MRCR long-document recall collapsed from 78.3% on 4.6 to 32.2% on 4.7.
    • HN: 674 points Apr 17, 606 points follow-up Apr 18.
    models news.ycombinator.com

    Cloudflare Ships Dynamic Workers — 100× Faster Than Containers
    • Dynamic Workers: isolate-based sandboxes that spin up in milliseconds, 100× faster and 100× more memory-efficient than containers.
    • Priced at $0.002 per unique Worker loaded per day — waived during beta.
    • Week also shipped Agent Memory, Cloudflare Mesh (private network for agents), and AI Gateway across 14+ providers.
    • 'Project Think' Agents SDK adds durable execution, sub-agents, and self-authored extensions.
    frameworks cloudflare.com

    xAI's Weekend Drop: Grok 4.3, Desktop Agent, and Voice APIs
    • Grok 4.3 beta appeared in the model selector Apr 17 — $300/mo SuperGrok Heavy only, adds native PDF/PowerPoint/spreadsheet generation and video input.
    • Grok STT/TTS APIs went GA Apr 18 across 25+ languages ($0.10/hr batch STT, $4.20 per 1M chars TTS).
    • Grok STT claims a 5.0% phone-call entity error rate vs. 12–21% for ElevenLabs, Deepgram, and AssemblyAI.
    • Grok Computer private beta also live — desktop agent processing the last 5 seconds of screen video.
    tools marktechpost.com

    Simon Willison Diffs Claude's System Prompt from 4.6 to 4.7
    • 'Donald Trump is the current president' block removed — 4.7's knowledge cutoff is January 2026.
    • Claude is explicitly told not to ask users to continue the conversation once they're done.
    • Ambiguity handling rewritten: make a reasonable attempt, or call a tool, before asking the human.
    • HN: 310 points, 177 comments (Apr 19).
    research simonwillison.net

    Claude Code Game Studios Hits 12.4k Stars
    • Donchitos/Claude-Code-Game-Studios: 12.4k stars, 1.8k forks — GitHub's breakout AI repo of the weekend.
    • 49 specialized Claude Code agents organized as a studio (directors, department leads, hands-on specialists).
    • 72 workflow skills and slash commands for every phase of game development.
    • Hooks validate commits, pushes, and asset changes automatically.
    open-source github.com