SpaceXAI (xAI) ships Grok 4.6, tuned for long-running agents: The Grok 4.5 successor is built to carry complex, multi-step work to completion — research, codebase analysis, or turning an idea into a working app. SpaceXAI says it matches GPT-5.6 Sol with a 61 on the Artificial Analysis Intelligence Index, and it is available today in Cursor and Grok Build. SpaceXAI
Qwen3.8-2.4T-A95B brings a Max-class model to open weights: Alibaba’s Qwen released a 2.4T-parameter, 95B-active MoE model — the first Qwen-Max-class model available openly. It improves coding, professional work, research, and long-horizon agentic tasks, with 262K native context extensible past 1M tokens. Hugging Face
DeepSeek V4 Pro 0813 reaches general availability: The GA release of DeepSeek’s large-scale MoE model, priced at $0.435 input / $0.87 output per 1M tokens, with a 1M-token context window and up to 384K output tokens. OpenRouter
Lovable confirms a $13.3B valuation with a $400M Series C: The app-generation startup raised $400M led by Menlo Ventures and EQT’s Scaleup Europe Fund, after passing $500M in annualized run-rate revenue in June. TechCrunch · Lovable
Made by Google ‘26: Pixel 11, Pixel Watch 5, Pixel Tag, and a Gemini push: Google announced the Pixel 11 lineup, a new AirTag rival, and a broad set of on-device Gemini features. TechCrunch
Frontier capability is commoditizing fast — and going open: Qwen3.8’s open Max-class release plus DeepSeek V4 Pro’s sub-$1 token pricing show frontier performance is no longer locked to a few closed labs. Teams should design for model portability rather than vendor lock-in. Hugging Face · OpenRouter
Training-data consent shifts to opt-out by default: Amazon will train on Twitch streamers’ content unless creators opt out. Twitch’s CPO admitted “if this was opt-in, nobody would opt in.” It signals how platform-scale training data is increasingly assumed rather than requested. TechCrunch · The Verge
Prototype long-context agentic apps on DeepSeek V4 Pro: $0.435/$0.87 per 1M tokens with a 1M context window makes it cheap to ship multi-step agents that ingest an entire codebase or corpus at once. OpenRouter
Self-host a Max-class model with Qwen3.8: With 95B active parameters and vLLM/SGLang compatibility, teams can now run frontier-grade inference in-house for data-sovereign professional and enterprise workloads. Hugging Face
Is AI-coding funding getting frothy?: Cognition is reportedly in talks to raise at a $40B valuation just months after $26B, and Blacksmith jumped ~10x to $550M in under a year. Nothing is closed, and the revenue multiples behind these numbers are unproven. TechCrunch · TechCrunch
Frontier benchmark claims are largely self-reported: Grok 4.6’s “matches GPT-5.6 Sol” claim rests on vendor-published system cards and leaderboards, not independent verification. Test models against your own workloads before committing. SpaceXAI