📰 News
-
Fable 5 tops every industry benchmark — but the price tag is brutal — Artificial Analysis released six new domain-specific benchmarks across Finance, Legal, Healthcare, Strategy, Engineering, and Economics. Fable 5 swept all six, with Opus 4.8 in second and GLM-5.2 leading open-weights. The catch: a single Strategy & Ops task costs $3.48 with Fable 5 versus $0.03 with DeepSeek — over 100× difference for a 12-point lead. → The Decoder
-
Anthropic’s answer to Fable 5’s price: delegate everything to Sonnet 5 — Two official patterns are now documented. ①Advisor: Sonnet 5 handles execution and only calls Fable 5 when stuck → 92% of Fable 5’s performance at 63% of the cost. ②Orchestrator: Fable 5 plans and distributes tasks to multiple Sonnet 5 worker agents → 96% performance at 46% cost. This reads as a direct response to pricing pressure from Chinese open-source models. → The Decoder
-
Grok 4.5 drops at $2 input / $6 output per million tokens — SpaceXAI’s new model is dramatically cheaper than Fable 5 ($10/$50) and GPT-5.5 ($5/$30), and uses 4.2× fewer tokens than Opus 4.8 on benchmark tasks. Terminal Bench 2.1 scores put it within one point of Fable 5 (83.3% vs 84.3%). But on real GitHub issue resolution (DeepSWE 1.1), the gap widens: Fable 5 at 70%, Grok at 53%. Elon called it “Opus-class” — fair in some contexts, a stretch in others. → The Decoder / TechCrunch
-
OpenAI launches GPT-Live: full-duplex voice that speaks and listens at the same time — The model handles interruptions naturally, drops in filler words like “mhmm” to signal it’s following along, and offloads complex queries to GPT-5.5 in the background. The result is a massive jump on agentic web search (BrowseComp): from 0.7% in Advanced Voice Mode to 75.2%. Paid users get GPT-Live-1; free users get the mini version. → The Decoder / TechCrunch
-
Google Gemini API adds background execution and MCP support for managed agents — Four new features: asynchronous background execution (no open HTTP connection required), remote MCP server connections, custom function support alongside built-in sandbox tools, and mid-session credential refresh. Developer quality-of-life improvement across the board. → The Decoder
-
Mistral enters robotics with an 8B model that needs just one camera — Robostral Navigate steers wheeled, legged, and flying robots through unknown environments using a single RGB camera. Trained entirely in simulation (400K paths across 6,000 virtual spaces), it hits 79.4% on the R2R-CE navigation benchmark — outperforming depth-sensor systems. Reinforcement learning is already pushing the number higher. → The Decoder
-
Prime Intellect raises $130M Series A to let enterprises train their own AI agents — The startup offers a full stack — compute, reinforcement learning framework, and eval tools — so companies can build domain-specific agents without relying on frontier labs. Already at $100M ARR. Ramp used the platform to build a spreadsheet agent that beats frontier models on accuracy, speed, and cost. → TechCrunch
🔥 GitHub Trending
-
addyosmani/agent-skills — Production-grade engineering skills for AI coding agents (Claude Code, Cursor, and 70+ more). Maps the full dev cycle — spec → plan → build → test → ship — to slash commands.
-
asgeirtj/system_prompts_leaks — Collected system prompts from Claude Fable 5, Opus 4.8, GPT-5.5, Gemini, Grok, and more. Includes a diff of exactly what changed between Opus 4.8 and Fable 5.
-
Zackriya-Solutions/meetily — Privacy-first AI meeting assistant: real-time transcription, speaker diarization, and summaries — all processed locally, no cloud required. Built for privacy-conscious teams.
-
steipete/CodexBar — macOS menu bar app that shows live usage stats and reset timers for Claude Code, Codex, Cursor, and 20+ other AI coding providers. Helpful when you’re burning through limits mid-task.
-
TencentCloud/CubeSandbox — High-performance secure sandbox for AI agents: sub-60ms boot, hardware-level isolation, E2B SDK compatible. Open-sourced by Tencent Cloud.
▶️ YouTube
-
DeepSeek’s New AI Speed Hack Is Amazing — Two Minute Papers | Breakdown of DeepSeek’s new inference speed optimization paper. Worth watching if you care about efficiency at scale.
-
Fable 5 Extended — Matthew Berman | Quick take on Fable 5’s extended capabilities. Good companion to today’s benchmark news.
💬 Community
-
Kenton Varda banned AI-written PR descriptions across his team — Simon Willison | The Cloudflare engineer’s take: “AI writes descriptions that outline code details visible from the diff itself, while omitting the higher-level framing needed to understand what the code is doing.” A lot of teams are hitting exactly this wall.
-
A global workspace in language models — Lobsters | Anthropic’s research finding a structure analogous to the “global workspace” from consciousness theory inside LLMs. Related to the J-Space work from yesterday.
-
Google’s exponential path to climate-wrecking digital bloat — Lobsters | Analysis of how AI infrastructure expansion is driving Google’s carbon footprint on an exponential curve. The other side of AI convenience.