GPT-5.6 and Codex Surge, Grok 4.5 Leads Benchmarks, Perplexity Open-Sources WANDR

Key Takeaways

1. OpenAI's GPT-5.6 Sol sees rapid growth; Codex and ChatGPT Work reach 8M users with rate limits removed.
2. Grok 4.5 tops FrontierSWE and Long-Horizon Terminal-Bench, positioning as a low-cost workflow OS.
3. Perplexity open-sources WANDR, a 500-task benchmark for deep and broad research.
4. Yann LeCun warns US chip export controls will backfire, boosting China and open models.
5. AI startups report revenue run rates: Cursor $4B, Scale $2B, highlighting market expansion.
6. Hugging Face releases OvisOCR2 (top OCR), Bonsai 27B (quantized to 3.9GB), and NVIDIA ARDY for real-time AI animation.

1. AI Model Releases and Benchmarks

1. Grok 4.5 from xAI tops FrontierSWE and Long-Horizon Terminal-Bench, outperforming Claude Opus and GPT series. It features low cost, fast processing of long-horizon engineering tasks, and new Stripe/Calendly connectors, becoming a workflow OS. — via 1 2 3 4 5 6
2. OvisOCR2 from Hugging Face achieves 96.58 on OmniDocBench v1.6, the first end-to-end model to beat pipeline systems, open-source and available on HF Jobs. — via 1
3. Bonsai 27B is quantized from 54GB to 3.9GB (-93%) via 1-bit quantization, retaining 90% performance, runnable on phones/browsers, Apache 2.0. — via 1 2
4. NVIDIA ARDY enables real-time AI animation from text, open-source and playable on Hugging Face Spaces. — via 1

2. Platform and Tool Updates

1. OpenAI reports GPT-5.6 Sol growth 5.6x, with reasoning team doing great work; Codex and ChatGPT Work reach 8 million active users with all rate limits reset and 5-hour caps removed. — via 1 2. Greg Brockman notes the model offers best price for any task and invites contact for better value. — via 3
2. Perplexity open-sources WANDR, a 500-task benchmark for measuring deep and broad research capabilities based on real knowledge work. Wide Research is now available via Agent API. — via 1 2
3. Vorflux (AI coding agent) achieves autopilot mode and announces a $15M seed round. — via 1
4. AI Engineer Conference identifies five trends: focus shift from agents to systems, circuit engineering as new control layer, AI engineering going enterprise, coding agents replacing IDEs, and agent platforms built around skills. — via 1
5. AI startup revenue run rates: 22 companies (excl. top 3 labs) exceed $500M ARR, with Cursor at $4B and Scale at $2B. — via 1

3. Policy and Industry Insights

1. Yann LeCun harshly criticizes US chip export controls, arguing they will destroy US market dominance and stimulate Chinese chip rise. He warns that further isolation will lead to the world standardizing on Chinese hardware/software, while the US falls behind due to closed-source protectionism. He stresses open source will win, like Linux. — via 1 2 3
2. swyx defines AI engineering principles: system ownership over titles, taste as ability for high-quality judgments, evidence-backed judgment as scarce resource when generation outpaces understanding, and new rule: "explain or don't release." — via 1
3. Elon Musk announces open-sourcing X's entire codebase after security review, inviting third-party audits to ensure consistency with running code. — via 1
4. SpaceX launches two Falcon 9 rockets from two launch sites within 8 hours, deploying 56 Starlink satellites; Starship still on track for tomorrow. New Starlink V5 terminal is 48% smaller, 62% lighter, supports 375+ Mbps, US-made. — via 1 2 3