DeepSeek V4 Pro & Grok 4.6 Launch: AI News Aug 13, 2026

I. Today's Headlines

• DeepSeek V4 Pro goes stable. On the night of August 12, DeepSeek flipped its API node to DeepSeek-V4-Pro-0813. The flagship keeps a 1M-token context and adds 384K max output, thinking plus non-thinking modes (thinking on by default), and Anthropic-API compatibility. Subject to vendor confirmation.
• Alibaba open-weights its strongest model. The Qwen team released Qwen3.8-2.4T-A95B weights — the first time a Qwen-Max-tier model is open. 2.4T total parameters with 95B active, 256K native context expandable to roughly 1M. Subject to vendor confirmation.
• SpaceX AI ships Grok 4.6. The new model scores 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol and trailing only Fable 5 (62), at USD 2 input / USD 6 output per million tokens. Subject to vendor confirmation.
• Claude gets an invisible watermark. Anthropic signed the EU AI Act Article 50 transparency code; from August 2 Claude embeds machine-readable watermarks in generated text and C2PA provenance metadata on files, applied globally across all Claude products and cloud partners. Subject to vendor confirmation.

II. Model Releases and Product Updates

• DeepSeek V4 Pro. 1.6T total / 49B active parameters. Pricing: RMB 3 per 1M input (RMB 0.025 cache-hit), RMB 6 per 1M output — three times V4-Flash. Concurrency limit 500 (Flash is 2500). Adds Responses API, Codex integration, JSON output and tool calls. Internal agent benchmarks jumped: DeepSWE 12.8 to 62.7, TerminalBench 87.9 (near Kimi-K3's 88.3). Performance nears Claude Fable 5. Subject to vendor confirmation.
• Qwen3.8-2.4T-A95B. A mixture-of-experts model with 512 experts per layer, 10 routed plus 1 shared per token. The cloud Qwen3.8-Max adds vision input, non-thinking mode and a 1M default context. Official scores: PaperBench 93.0, TerminalBench 2.1 86.6, FrontierSWE 73.5, CoWorkBench 74.8. Released under a custom Qwen license (not Apache 2.0); Qwen3.8-27B is next. Subject to vendor confirmation.
• Grok 4.6. Built for long-running agents, complex interaction and vision tasks. A faster tier doubles the price. SpaceX AI also introduced GrokBot, a "digital coworker" with its own cloud Linux VM and 24/7 autonomy, at about USD 200 per month (personal) or USD 120 per month (team). Subject to vendor confirmation.

III. Industry and Capital

• NVIDIA mobilizes USD 500 billion. Together with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, NVIDIA is setting up financing platforms to channel more than USD 500 billion of third-party capital into AI infrastructure. Subject to vendor confirmation.
• OpenAI widens free access. Free and Go users get unlimited text chat and a "Think" button, with GPT-5.6 Luna as the new default; weekly limits migrate to monthly on August 15; the DALL·E GPT retires on August 30. Subject to vendor confirmation.
• Microsoft ships MAI-Code-1.1-Flash. Now powering GitHub Copilot, the model cuts cost and raises code-survival rates; Microsoft also added MAI-Image-2.5-Pro, MAI-Voice-2-Flash and its first purpose-built security model MAI-Cyber-1-Flash, and co-signed NVIDIA's open-weights letter (Anthropic is the lone major holdout). Subject to vendor confirmation.

IV. China Power

• A Hangzhou double shot. DeepSeek V4 Pro and Alibaba Qwen3.8 landed the same day, together showing that Chinese labs now ship both cloud-scale APIs and top-tier open weights. Subject to vendor confirmation.
• Pricing as product. DeepSeek's three-tier spread (Flash at RMB 1 / 2 versus Pro at RMB 3 / 6 per 1M tokens) mirrors the industry's shift from "who is smartest" to "cost per unit of work." Subject to vendor confirmation.
• Local silicon. Moore Threads reported first-half revenue up 147% and filed for a Hong Kong IPO; ChangXin (CXMT) drew a rare proactive Apple inquiry for LPDDR5X memory. Subject to vendor confirmation.

V. Today's Observation

Three flagship-class releases in a single day — DeepSeek V4 Pro, Qwen3.8 and Grok 4.6 — mark the moment the contest stopped being about leaderboard points and became about agent infrastructure: context length, tool use, throughput and price per task. Open weights kept winning (Qwen, DeepSeek, Meta's Muse Glimmer, NVIDIA Nemotron), while Anthropic's global watermark signals a parallel push for AI-content traceability. The frontier is now a platform, not a leaderboard.