AI & design briefing for Thursday, July 2, 2026

tldr

• Anthropic's Fable 5 returns after 18-day government shutdown; Claude Sonnet 5 delivers near-Opus reasoning at 40% cost; Together AI raises $800M at $8.3B.
• 39 principles for human-AI interaction design; CMU paper shows most 'AI agents' are just scaffolding; Senior SWE-Bench benchmarks senior-level coding.
• xAI Voice Agent Builder launches; Google ships agent stack with Genkit + ADK 2.0; OpenAI Tax AI debuts in Codex.

Top Stories

• Anthropic's Fable 5 returns after 18-day government-mandated shutdown. Anthropic redeployed Fable 5 with a new cybersecurity classifier after US Commerce Department lifted export controls. Blocks reported jailbreak in >99% of cases. Paid access through July 7 capped at 50% of weekly limits. Flagged requests fall back to Opus 4.8. ^{[1] [2]}
• Claude Sonnet 5 ships with near-Opus reasoning at ~40% of the cost. Claude Sonnet 5 scores 63.2% on agentic coding (vs Opus 4.8's 69.2%) with 1M context window and 128K output limit. Introductory pricing $2/$10 per million tokens through August 31. ^[20]
• Meta explores cloud business for spare AI compute. Meta reportedly building cloud service to rent spare data-center capacity. Stock up 9.3%. Options range from raw compute to hosted model inference. ^[4]
• Together AI raises $800M at $8.3B valuation. Funding round expands open-model infrastructure, positioning Together as a major alternative to closed API providers. ^[5]
• AI climbs the freelance value chain. Remote Labor Index: Fable 5 matched/beat human pros on 16.1% of 240 real freelance tasks, up from GPT-5.2's 2.5% in Oct 2025. Opus 4.8 second at 8.3%, GPT-5.5 third at 6.3%. ^[6]
• NVIDIA splits a 30B model in two for 2.42x faster generation. New technique partitions a single large model across devices for nearly 2.5x text-generation speedup for multi-GPU inference. ^[3]

Deeper Cuts

• 39 principles for designing human-AI interaction. Applied framework for AI interfaces covering appropriate reliance, user control, transparency, and responsible autonomy. ^[7]
• Map the fog before you build. Matt Pocock's /decision-mapping prompt pattern: before greenfield projects, ask AI to separate fixed decisions from unknowns and parallel work tracks. ^[8]
• CMU paper: most 'AI agents' are just scaffolding. Argues that many agentic systems are brittle wrappers around LLMs rather than genuine autonomous agents.
• Senior SWE-Bench: benchmark for senior-level coding tasks. Open-source benchmark testing coding agents on ambiguous, multi-step senior-engineer-level tasks. ^[9]

Tools and Launches

• xAI Voice Agent Builder. No-code Grok voice agents for support/sales/scheduling at $0.05/min. ^[10]
• Devin Security Swarm. Autonomous codebase scanning, sandboxed exploit testing, and remediation PRs. ^[11]
• Factory AI Droid Shield 2.0. Learned secret-detection for safer autonomous engineering agents. ^[12]
• Google's agent stack (Genkit + ADK 2.0). Agentic full-stack development with Genkit, ADK 2.0, and cloud-local ML in VS Code. ^[13]
• Google Design.md standard. Agent-friendly design briefs as markdown, compatible with Claude Code. ^[14]
• OpenAI Tax AI (Codex). Self-improving agent with expert-feedback loops for complex tax preparation. ^[19]

Prompt of the day