David Sacks stated that Anthropic’s IPO must be suspended until the "internal whistleblower" allegations are investigated.

Beating AI News Flash: Legal AI firm Harvey has secured a new $550 million funding round, lifting its valuation to $15.5 billion, with co-lead investors Diffusion and Lightspeed. Back in March, the company raised $200 million at a $11 billion valuation. Harvey provides AI tools for large law firms and corporate legal teams, covering contract review, due diligence, litigation support and more. The new capital will be primarily allocated to developing its proprietary models. In August, Harvey launched its first post-trained model, Tenet, built on Moonshot AI’s Kimi K3 base model; it is continuing to fine-tune Tenet with Fireworks for long-cycle legal tasks. Harvey is a portfolio company of OpenAI Startup Fund and has previously co-developed legal models with OpenAI. It now proactively opts for lower-cost base models when performance is comparable. Harvey’s annual recurring revenue (ARR) has exceeded $400 million, with over 3,000 clients, up from 1,300 in March. The company’s total cumulative funding raised surpasses $1.5 billion.

Crypto analyst Darkfost’s data analysis shows that despite Bitcoin’s roughly 45% recent rebound, liquidity remains weak, spot demand is still relatively soft, and the 90-day moving average of cumulative volume delta stays neutral. In futures markets, meanwhile, buyers have clearly gained the upper hand. Separately, stablecoin inflows to centralized exchanges (CEXs) have slowed. For instance, Binance’s stablecoin reserves have fallen nearly $70 billion from their peak; more inflows will be required for the market to reach higher levels. Darkfost also stated that Bitcoin’s technical outlook remains relatively optimistic, and a strong breakout above $80,000 would be a key threshold for liquidity to truly return.

Beating AI News reports, citing The Information and sources familiar with the matter, that DreamWorks co-founder Jeffrey Katzenberg is launching an AI video company with former OpenAI Sora lead Bill Peebles and ex-Dropbox CFO Sujay Jaswa. The startup will train its own video models to target filmmakers directly, and is currently in talks with investors including a16z for funding. The company’s name and funding size have not been disclosed. Peebles served as Sora’s core technical lead and is one of the authors of the Diffusion Transformer (DiT) paper. He left OpenAI shortly after the company shut down Sora’s consumer app and API this year, publicly stating at the time that exploratory research of this type needs space to operate outside a company’s core product roadmap. Katzenberg has already made an AI video bet this year: in May, his firm WndrCo participated in a $59 million funding round for real-time AI video startup Reactor. This time, he is going beyond investor status to co-found a company with the former Sora lead to build its own models. This marks Katzenberg’s most ambitious video venture since Quibi. The short-form video platform raised $1.75 billion but folded shortly after launching in 2020. Six years later, he is returning to the video space, though this time shifting from a short-form platform to developing AI models for filmmakers.

Solana core development team Anza announced on X that Mainnet Beta Transaction V1 will activate starting at Epoch 1035, with an estimated time of approximately 1:20 UTC on September 15. The ecosystem team noted that additional time is required to test and integrate V1 support; V1 is now live on the devnet for integration testing. Compute budget in V1 is set via new transaction configurations. While compute budget instructions remain functional in both legacy and V0 transactions, they will not take effect in V1. Application developers, even those not planning to submit V1 transactions, must verify the impact of these changes on their applications.

According to an official announcement, Bithumb will list the DEBIT token on its KRW market. Trading is scheduled to start at 3:00 PM local time on Thursday, September 10.

Beating AI Express News (Dongcha): DeepSeek’s V4.1 Flash is now rolling out on its app, merging its original three modes—“Fast”, “Expert”, and “Image Recognition”—into a single entry point. Users no longer need to manually select modes; the unified model handles daily conversations, image understanding, and complex queries simultaneously, with V4.1 Flash natively supporting multimodal capabilities. DeepSeek previously noted that the new model has fully outperformed V4 Pro across metrics including performance, cost, speed, and total processing time, and will take over all V4 Pro requests ahead of the launch of V4.1 Pro.