Thursday, August 14, 2026. A packed 24 hours in AI: DeepSeek shipped the stable version of V4 Pro, SpaceX AI’s Grok 4.6 is back in the top tier at aggressive prices, and Claude quietly cleared a long-standing math problem. Google says Gemini just passed 1 billion monthly users, and Anthropic started watermarking AI-generated text. Here’s what matters.
Top 3
1. DeepSeek V4 Pro stable version is live
DeepSeek updated its API docs overnight with DeepSeek-V4-Pro-0813, completing the official rollout of both V4 variants (Flash went stable in late July). The Pro release keeps 1M context and 384K max output, priced at ¥3/M input and ¥6/M output tokens. Developers call it with the existing deepseek-v4-pro name — no code changes needed. [36Kr]
2. Grok 4.6 returns to the top tier — and undercuts rivals
SpaceX AI released Grok 4.6 on August 12, priced at $2/M input and $6/M output tokens — below Fable 5 — and early reviews say it is back among the frontier leaders. The Cursor acquisition and the new agent-style work features are cited as the reasons the gap closed. [QbitAI]
3. Claude cleared every Hadamard matrix below order 2000
Anthropic’s Claude swept the list of unsolved Hadamard matrix cases under order 2000 — a combinatorial math problem that had resisted automated attacks. It is the latest sign of LLMs moving from doing math to doing mathematics: not just solving known problems faster, but knocking items off the open-problems list. [QbitAI]
More news
- Gemini passes 1B monthly users. Pichai calls it Alphabet’s fastest-growing product ever — the 14th Google product to hit the milestone. [36Kr]
- Anthropic starts watermarking AI text. Invisible, machine-readable watermarks are now embedded in Claude outputs (models released from Aug 2); copy-paste and light editing will not remove them. [36Kr]
- Google’s co-founder is back in the model war. Sergey Brin is pushing Gemini teams hard, reportedly steering resources toward recursive self-improvement, after internal tests showed new Gemini versions still lagging rivals on coding; the flagship update slipped two months to August. [36Kr]
- Tencent Q2: AI capex up 176% YoY. Revenue hit ¥204.9B (+11%); cloud growth is being driven by AI demand, though new AI product investment trimmed operating profit growth to 9%. [36Kr]
- Seedance 2.5 reviewed: 30s clips, 50 reference assets, local editing. A seven-dimension test found cinematic camera work and near-live-action visuals — but character consistency in long shots remains the biggest unsolved pain point. [36Kr]
- Samsung verified chip designs with Claude: one month → two days. After rolling out Claude Code to software developers, Samsung’s custom SoC validation tasks that once took over a month now finish in two days — roughly 15x faster. [36Kr]
Trend watch
Two shifts worth tracking. First, price is the new battleground: DeepSeek is now shipping stable V4 models at sub-$1/M input while Grok 4.6 undercuts the incumbents — frontier-quality inference is getting cheap fast, and that changes who can build on top of it. Second, AI is moving from generation to judgment: Anthropic’s watermarking, Agent-as-Judge evaluation systems, and the ChinaJoy debate over who plays the AI games once anyone can make them — the scarce resource is no longer “can you make it,” it is verification, distribution, and quality. The Hadamard milestone is the wildcard: when models start clearing open-problems lists, the “AI scientist” story stops being a demo.
Disclaimer: This briefing is compiled by AI from public sources; verify details at the linked originals. Prices and facts are as reported at publication time.
