1. Anthropic Details Claude Text Watermark: No Hidden Characters, Detectable Pattern Comes From Random Word Choice Anthropic says future Claude models will watermark generated text without inserting hidden characters or attaching extra information.

2. Apple Reportedly Trained Its Own China AI Model With Alibaba as Apple Intelligence Clears a Regulatory Hurdle Apple has reportedly worked with Chinese technology giant Alibaba to train a custom large language model for China, changing its approach as it prepares to introduce Apple Intelligence in the country.

3. Cerebras and OpenAI Preview GPT-5.6 Sol Ultrafast Service, Claiming Up to 750 Output Tokens per Second Cerebras and OpenAI have previewed Ultrafast Mode, a new OpenAI API service tier that runs GPT-5.6 Sol on Cerebras hardware.

In Brief

• LiteLLM Supply-Chain Attack Exposed Credentials From More Than 2,500 Organizations Security firms CloudSEK and Hudson Rock said compromised versions of LiteLLM, an open-source AI development tool distributed through PyPI, extracted cloud keys, repository tokens, SSH keys, and other secrets during a 40-minute window in March. Hudson Rock analyzed a 195TB file, while independent researcher Kevin Beaumont said he confirmed data belonging to multiple victims.

• OpenAI and Anthropic Cut Model Prices as Chinese Rivals Gain Customers OpenAI reduced GPT-5.6 Luna prices by 80%, while Anthropic launched Claude Opus 5 at half the price of Fable 5. Silicon Data’s index indicates prices paid for leading US models have fallen almost 25% since mid-July as companies including DoorDash and Airbnb test lower-cost Chinese models.

• Microsoft Consolidates Copilot Apps and Retires Deep Research, Group Chats, and Mico Microsoft is combining its consumer Copilot and Microsoft 365 Copilot apps, with generated files migrating to OneDrive. Group Chats, AI-generated podcasts, Copilot Labs experiments, consumer Deep Research, and the animated Mico character are scheduled for removal by August 18.

• Google Makes Visible Watermarks Optional for AI-Generated Media Google will let users remove visible watermarks from images, videos, and songs produced with its Nano Banana, Omni, and Lyria models, while retaining invisible SynthID markers and C2PA metadata. The setting is rolling out in Gemini and the Flow video editor, with Search support planned.

• Court Sanctions Plaintiff Who Hid Prompt-Injection Instructions in Filings Connecticut judge Walter Spader Jr. said plaintiff Matthew Elliott embedded white-on-white instructions intended to make any AI reviewing his filings favor his arguments. The judge said the prompts had no effect on the decision but imposed modest sanctions after Elliott continued inserting hidden text despite a warning.

• Anthropic Investors Model a Potential $2 Trillion IPO Valuation Six Anthropic backers told the Financial Times they expect the Claude developer could seek a valuation of at least $2 trillion in a planned October flotation. Senior Anthropic executives have not fixed a valuation target, according to the report, and the projections are investors’ own estimates.

• Kog Targets Faster LLM Inference on Standard Data-Center GPUs French startup Kog says its software-level GPU optimization produced 3,000 tokens per second with its open-source, two-billion-parameter Laneformer model on conventional AMD and Nvidia hardware. The company still needs to demonstrate the approach on larger models and aims to show a major model running 10 times faster in September.

• ShieldFont Serves AI Scrapers Altered Text While Keeping Pages Human-Readable Designers Isaque Seneda and Gabriel Abrucio created ShieldFont, which uses font ligatures to display intended words to readers while leaving substituted words in a webpage’s underlying text. The project is intended to reduce the value of content collected by plaintext-based AI scrapers.

• LLMRouter Open-Sources a Common Toolkit for Cost-Aware Model Selection Researchers released LLMRouter, modular infrastructure containing more than 16 routing methods, alongside the xRouteBench benchmark for text, vision, time-series, memory, and personalized tasks. Their experiments report that learned routers outperformed the strongest fixed-model baseline by 14.6% relatively.

• DarwinX Evolves Agent Prompts, Tools, and Control Flow Without Retraining Models The DarwinX paper describes selecting and recombining populations of agent harnesses while keeping model weights frozen. Its authors report improvements across terminal, web, and software-engineering benchmarks, including a zero-shot transfer from Terminal-Bench to SWE-bench Verified.

• OpenAI Names Dali Rajic Chief Revenue Officer OpenAI appointed Dali Rajic to lead its global revenue organization and work with businesses adopting the company’s AI products.

Read the full edition →