Daily AI RoundupsThursday, 13 August 2026

DeepSeek's V4 Pro leads the 13 August edition after reaching general availability across its app, web service and API, with DeepSWE rising to 62.7 and Terminal-Bench 2.1 to 87.9. xAI released Grok 4.6 for long-running agents, while Anthropic was reportedly in talks to acquire world-model startup Decart for approximately $6 billion. The roundup also covers an unconfirmed claim about Safe Superintelligence's test-time training work, NVIDIA's serving guide for Qwen's 2.4T-parameter A95B model, Liquid AI's LFM2.5-VL-3B, GitHub Agent Plugins 1.0, Google DeepMind's sign-language tools, Microsoft MAI-Thinking-1, AI compute futures, model watermarking, benchmark sensitivity and LTX 2.5.

• frontier models
• AI agents
• world models
• model serving
• AI research
• developer tools
• AI business
• AI policy

Stories in this roundup

1. 01
   DeepSeek rolls out V4 Pro GA across its app, web and API, lifting DeepSWE to 62.7 and Terminal Bench 2.1 to 87.9DeepSeek
2. 02
   xAI releases Grok 4.6 for long-running agents and makes it available through its API, Cursor and Grok BuildxAI
3. 03
   Anthropic is reportedly in talks to acquire world-model startup Decart for approximately $6 billionYahoo Finance
4. 04
   A pseudonymous X account claims Safe Superintelligence has built a small reasoning engine around test-time training, but SSI has not publicly confirmed itiruletheworldmo / X
5. 05
   NVIDIA publishes a serving guide for Qwen's 2.4T-parameter A95B open-weight model on GB300 NVL72NVIDIA
6. 06
   Liquid AI releases LFM2.5-VL-3B, a 3B vision-language model for edge deploymentsLiquid AI / Hugging Face
7. 07
   John J. Wang investigates how Claude's text watermark may work and where it could be strippedJohn J. Wang
8. 08
   An arXiv study finds benchmark rankings can reverse when identical problems are rewordedarXiv
9. 09
   GitHub releases Agent Plugins 1.0 across VS Code, Copilot CLI and the Copilot appGitHub
10. 10
    Google DeepMind ships SL2T, bringing sign-language input to Gboard and Live Transcribe on AndroidGoogle DeepMind
11. 11
    Anthropic investors are targeting a $2 trillion valuation in a potential IPO, the Financial Times reportsFinancial Times
12. 12
    CME Group will list two AI compute futures contracts beginning October 5CME Group
13. 13
    The White House is considering adding open models to an updated AI policy framework, Wired reportsWired
14. 14
    xAI begins testing Grok as an AI teammate that users can assign workThe Verge
15. 15
    Microsoft makes its first from-scratch reasoning model, MAI-Thinking-1, available in Microsoft FoundryMustafa Suleyman / X
16. 16
    Lightricks releases LTX 2.5 for AI video generationLightricks