Key Takeaways
- OpenAI releases GPT-6.2 Astra and Sol, with mixed early test feedback and strong efficiency claims.
- Perplexity's ARR reportedly passes $1B as it ships an Orchestrator API and Context API.
- Tesla unveils Powercore, a solid-state medium-voltage transformer for 800 Vdc data centers.
- SpaceXSI reports Grok Bots cutting accounting work from hours to minutes.
- Runway halves the GPUs needed to run its 57B-parameter video model.
- Qwen3.8-27B-pi-adapted posts large gains on coding-agent benchmarks with open weights.
1. AI Models and Releases
- OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, which OpenAI described as efficient and focused on real-world performance; Greg Brockman said both perform well on professional work and scientific research at substantially lower per-task cost 1 2. Early testers were divided: Ethan Mollick called Astra uneven and harder to steer but impressive on large projects, while another account said it performed poorly across reviewer evaluations 3 4 5. Perplexity reported that both models are available in its Search and Computer products, with Astra scoring 0.783 on WANDR at about $12.54 per task and Sol scoring 0.746 at about $2.65 6 7. A separate post said Sol scored 82% on browser use versus 81% for GPT-6.1 Sol, at 35% lower cost and twice the speed 8.
- Perplexity released an Orchestrator API that routes across more than 45 open-weight and proprietary models with unified billing and KV-cache-friendly routing; routing requests that one model often refuses to gpt-6.1 sol raised terminal-bench 4.1 solves from 38/63 to 41/63 1 2 3. It also launched a Context API reading up to 10 URLs per call at $1 per 1,000 URLs, with cited figures of 92.3% usable content and 87.9% fact recall 4. Yann LeCun's account said Perplexity's ARR passed $1B with strong margins and healthy growth 5.
- Hugging Face highlighted Qwen3.8-27B-pi-adapted, tuned with reinforcement learning and LoRA adapters for the pi coding agent; SWE-bench Pro (Hard 51) rose from 29.4% to 62.7% and Terminal-Bench 2.1 from 39.5% to 55.8%, with 55–62% fewer tool calls per task and open weights 1. Midjourney added Moodboards, Style Codes and Style References to its editor model, letting users attach images to the prompt area without typing a prompt 2. Mistral Vibe launched desktop versions for macOS and Windows with local or remote agents 3.
2. Infrastructure and Enterprise
- Tesla unveiled Powercore, a solid-state medium-voltage transformer integrating transformer, inverter and breaker, with 98.86% peak efficiency, 9 MVA/7.4 MW rating, 10–35 kV AC and 800–2000 Vdc, a 25-year warranty and planned 2027 deployment for 800 Vdc data centers 1. A reposted claim said it halves losses versus oil-immersed copper-coil transformers 2.
- SpaceXSI uses five dedicated Grok Bots in accounting: bill processing fell from about 24 hours to minutes, payment reconciliation from 3–4 hours to about 30 minutes, and the company estimated it would need more than double its accounting team without them; refunds still require human approval 1 2. Musk said the bots can review every expense and call suppliers directly 3.
- Runway said it cut the GPUs needed to run its 57B-parameter video model at 24fps from 8 to 4 and published a research article on the optimizations 1. OpenAI launched Daybreak for NATO and allied networks and GPT-6 Cyber for internet security 2 3.
3. Research and Industry Signals
- A reposted Energy-Based Reasoners study reported 345x less inference compute than the best baseline, 2.6x fewer training FLOPs, 7–43x better sample efficiency, 99.9% on Sudoku-Extreme and near-100% on multi-solution benchmarks 1. A reposted Cambridge AIxGEO analysis split the AI supply chain into 14 layers and 137 elements, finding no single country could control the whole chain and that over a third of the supply chain went unmentioned in the policy documents analyzed 2.
- Ethan Mollick argued that Astra's uneven reception shows the continued value of "vibe benchmarks" beyond numeric scores, and cautioned against over-extrapolating from any single model release 1 2. swyx criticized the Astra launch for lacking someone with authority to veto the ad, while calling the model a solid incremental improvement and world-leading in efficiency 3.
- A reposted view said Cloudflare followed Jev with Clef, attributing the pace to asymmetric effort: new ideas are hard to invent and easy to copy, and smaller, faster, cheaper models can take over some work from LLMs 1. Hamel Husain's company raised a $57.5M Series B to help researchers turn petabyte-scale video, image and audio data into better models 2.
