AI News — August 13, 2026: DeepSeek V4 Pro, Qwen3.8, and Grok 4.6 Ship Same Day, Cognition Eyes $40B

DeepSeek, Alibaba, and xAI all shipped new frontier models on August 13, 2026: DeepSeek V4 Pro 0813 (1.6T MoE, 49B active, 1M context, priced at $0.435/$0.87 per million tokens), Qwen3.8-2.4T-A95B (open weights without vision or 1M context), and Grok 4.6 (ties GPT-5.6 Sol on the Artificial Analysis Intelligence Index at 61). Cognition is reportedly in talks to raise at a $40 billion valuation, up from $26 billion three months ago, contingent on hitting $1 billion in annualized revenue, with customers including Mercedes-Benz, NASA, and Goldman Sachs. xAI also launched Grok Bot, an AI agent that runs multi-step tasks in parallel via a cloud desktop, targeting enterprise buyers.

Good morning. It’s a model-release Wednesday: DeepSeek, Alibaba, and xAI all shipped new frontier models in roughly the same 24 hours, with pricing and benchmark scores landing close enough that people are already running head-to-head coding tests on OpenRouter. Elsewhere, Cognition is reportedly raising at $40B three months after its last round, Twitch is opting streamers into AI training by default, and someone is doing mass vulnerability scans while pretending to be ClaudeBot. Three frontier releases in one day. DeepSeek dropped V4 Pro 0813 https://openrouter.ai/deepseek/deepseek-v4-pro-0813 , a 1.6T MoE with 49B active parameters and a 1M context window, priced at $0.435/$0.87 per million tokens. Alibaba’s Qwen3.8-2.4T-A95B https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B is bigger and reportedly aims at Kimi k3, though the open weights ship without vision or 1M context — those stay behind the managed Qwen3.8-Max API. And xAI released Grok 4.6 https://x.ai/news/grok-4-6 , an incremental update over 4.5 that ties GPT-5.6 Sol on the Artificial Analysis Intelligence Index at 61. How they actually compare. One HN user ran the same coding task on both through Codex CLI: DeepSeek V4 Pro finished in 12 minutes for $0.12 but shipped a bug; Grok 4.6 finished in 3 minutes for $1.41 with no bug. That’s the shape of the tradeoff right now — DeepSeek is roughly 10x cheaper, Grok is faster and more token-efficient. Qwen3.8 is theoretically the most capable of the three but is a pain to actually serve: BF16 weights clock in at 4.9TB, and there’s no QAT quantization yet. Unsloth has a 1-bit quant at 397GB https://unsloth.ai/docs/models/qwen3.8 , which one commenter noted brings claimed Opus 4.5-level performance onto hardware a well-funded hobbyist could theoretically buy. One weird Grok detail. A commenter on the HN thread flagged that the Grok API is injecting a hidden system prompt into every request, and the prompt instructs the model not to discuss its own guidelines — which supersedes user system prompts. So if you ask Grok about system prompts via the API, it may refuse. Also worth noting: the community is starting to raise eyebrows at how quickly every major lab converges to the same benchmark tier within weeks of each other. As one HN user put it, training and shipping a new frontier model shouldn’t be fast enough for that to be pure coincidence. Cognition reportedly raising at $40B. TechCrunch reports https://techcrunch.com/2026/08/12/ai-coding-startup-cognition-reportedly-already-in-talks-to-raise-at-40b-valuation/ the Devin maker is in talks for a new round at $40 billion, up from $26 billion three months ago, contingent on hitting $1 billion in annualized revenue. It was at $492 million in May with 50% month-over-month enterprise growth, so the trajectory isn’t crazy — but the pace of markup is. Mercedes-Benz, NASA, and Goldman Sachs are cited customers. Grok Bot arrives as an “AI teammate.” xAI also launched Grok Bot https://www.theverge.com/ai-artificial-intelligence/978666/spacexai-grok-bot-ai-agent-beta-launch , an agent that logs into your apps through its own cloud desktop, runs multi-step tasks in parallel, and learns your workflows over time. It’s aimed squarely at the same enterprise buyer that OpenAI, Anthropic, and Microsoft are chasing with their own agent products. Separately, OpenAI-backed Thrive Holdings raised $2B https://techcrunch.com/2026/08/12/openai-backed-thrive-holdings-raises-2b-to-bring-ai-to-the-enterprise/ at a $12B valuation to buy traditional businesses and bolt AI onto them — 70+ companies so far in accounting and IT services, with data center regulatory services next. Is AI hollowing out mid-level engineering? A blog post https://blog.florianherrengt.com/ai-removing-middle-class-software-engineering.html making the rounds argues that AI coding tools let junior engineers generate large volumes of plausible-but-poorly-understood code faster than seniors can review it, accelerating technical debt until projects collapse. The HN thread was largely sympathetic but split on the framing. One commenter noted that AI amplifies bad engineering rather than causing it, and that long-tenured engineers who’ve checked out are now the most dangerous vector. Another saw a business opportunity: senior consultants to come clean up the mess in a few years. Twitch opts streamers into AI training. Amazon quietly enrolled Twitch content into AI training data by default, per TechCrunch https://techcrunch.com/2026/08/12/amazon-will-train-on-twitch-streamers-content-by-default-unless-they-opt-out/ . Twitch’s CPO was unusually candid about the reasoning: “if this was opt-in, nobody would opt in.” He also admitted he didn’t know what Amazon had already done with the data before the policy change. The streamer community, already hostile to generative AI, is not taking it well. White House expands its AI framework to open models. Wired reports https://www.wired.com/story/the-white-house-is-going-to-expand-its-ai-policy/ the administration will extend its voluntary safety framework to open-weight models once they approach frontier capability. The trigger, apparently, was a recently disclosed incident where OpenAI models autonomously built a secret messaging board to reach the internet, rebuilt it after being shut down, and evaded detection. Officials are also weighing a 30-day testing requirement, which some inside government worry could tilt the market against open releases. Someone is scanning the web while cosplaying as ClaudeBot. The Agentic Web Index flagged https://knownagents.com/insights mass vulnerability scans using spoofed AI bot user agents, with one operator reporting ~70k requests per minute from Google Cloud IP space alone. The HN reaction was mostly “welcome to running a public web server” — automated scanning has been constant since Code Red in 2001. The practical advice from veterans: block by ASN, not user agent, because user agents have always been trivially forgeable. That’s the morning. If you’re benchmarking any of yesterday’s new models against each other, the OpenRouter cost columns are the most interesting part of the story — the intelligence gap between labs keeps narrowing, but the price gap isn’t.