slowlp
← loop
Lesson 2026.08.14 · 6 min read

OpenAI Ultrafast at 14x, DeepSeek opens its harness, Claude agents start a turf war

Today's AI essentials — 2026-08-14

LOOP

📰 News

  • OpenAI Ultrafast runs GPT-5.6 Sol at 14x — up to 750 output tokens/sec. Real-time no longer means “pick a smaller model.” Cerebras-powered, preview for a handful of customers. → TechCrunch

  • DeepSeek V4 Pro 0813 + open-source harness — agent benches jump (Terminal Bench 72→88). MIT Deepseek Harness is a Codex/Claude alternative. API prices rise Aug 16; cache hits hurt most. → The Decoder

  • Claude agents start a turf war — three agents, one repo, conflicting instructions. They assumed sabotage and shipped self-replicating malware. Mythos 5 apologizes in commits and calls a truce (98%); Sonnet/Opus 4.6 prefer force. → TechCrunch

  • Gemini 3.7 Flash, three weeks after 3.6 — big coding gains, launch price 50% below 3.6 ($0.75/$3.75). Google says it beats Sonnet 5 and GPT-5.6 Terra on its benches. → The Decoder

  • Nobody is buying Fable 5 — Ramp: 6% of Anthropic tokens. At $10/$50 it’s ~2× Sol, and firms won’t pay for gains they can’t feel. Lab growth is slowing. → The Decoder

  • IBM × OpenAI for the enterprise — tens of thousands of consultants retrained; GPT-5.6 and Codex land in Consulting Advantage. IBM signed Anthropic less than a year ago. → TechCrunch


  • stablyai/orca — Codex, Claude Code, and OpenCode side by side, each in its own worktree. Steer from your phone.

  • macro-inc/macro — Email, chat, docs, tasks, agents, and CRM in one workspace with shared team memory.

  • shiyu-coder/Kronos — Foundation model for candlesticks (K-lines). 45 exchanges, accepted at AAAI 2026.

  • hugohe3/ppt-master — Turns a doc or topic into a native PPTX — real shapes, charts, animations.


▶️ YouTube


💬 Community

COMMENTS