A trader spent $2,972 to buy MEME, reaping a 713x return.

Dongcha Beating AI News Flash: OpenAI may soon update its image generation model. Community researchers unearthed strings including "imagegen-25-announcement-modal" and "imageGen25Announcement" in the latest Codex Desktop frontend, and forced the hidden announcement pop-up to display. The pop-up reads "Major upgrade for image creation", highlighting higher image quality, faster generation speeds, and smarter creation tools.

According to an official announcement, Binance will add watch tags for AVA (AVA), Gains Network (GNS), Scroll (SCR), and Towns Protocol (TOWNS). Compared to other listed tokens, watch-tagged assets have higher volatility and greater associated risk. Trading these tokens is risky, as they no longer meet the exchange’s listing standards and are subject to potential delisting.

Crypto KOL Ansem published a post stating that over the next 12 to 18 months, the crypto industry will see more than 10 protocols grow from scratch to a valuation of over $1 billion. He attributes this trend primarily to the ongoing migration of real-world assets (RWA) onto blockchains. Meanwhile, more skilled developers are coming to recognize that the crypto sector enables immediate incentive alignment, boosting their willingness to participate; continuously evolving AI models will also lower the barriers to protocol development.

According to GMGN data, the stock-meme project GME (Greatest Meme Ever) on Robinhood Chain saw its market cap briefly surge past $11 million within five hours of launch, before pulling back to $7.1 million, with trading volume reaching $8.7 million. GME uses stock trading platform Robinhood, pairing its liquidity pool with the tokenized version of U.S. stock GameStop (ticker: GME), and provides liquidity via a GME/GME trading pair. BlockBeats Note: Stock Meme is an emerging concept that combines traditional meme coins with tokenized U.S. stocks: instead of pairing meme coins with USDT or ETH, they are directly matched with on-chain U.S. stock tokens (such as NVDA, TSLA, AAPL, etc.). This model retains meme coins' high volatility and community-driven speculative attributes, while leveraging the popularity and narrative of real stocks. A portion of transaction fees is often channeled back to the community treasury to accumulate corresponding U.S. stock tokens, forming a dual-driven model of "sentiment speculation + real asset anchoring". Prices are highly volatile; investors should exercise caution.

Per official announcement, Huobi HTX launched GPRO/USDT, LMT/USDT, and SMIC/USDT perpetual contracts on September 4, supporting 1x to 20x long and short positions. Meanwhile, from now until 15:00 UTC+8 on September 8, Huobi HTX is hosting a new contract token trading contest: users who complete registration, trade the eligible token contracts, and meet the specified thresholds will have the chance to split the total prize pool of 1 billion $HTX.

From Beating AI Express: Perplexity used its proprietary agent benchmark WANDR to test GPT-6 Astra, which scored 0.682 — the highest mark of any model tested to date. The average cost per task came to $11.98. Compared to Claude Fable 5.1, Astra’s score is 13.5% higher, while its cost is 6.1% lower; versus Opus 5, the score is 27% higher, with a cost increase of only 3.3%. WANDR is tailored to evaluate "broad and deep" research tasks, featuring 500 real-world research assignments that require agents to not just find partial answers, but to identify all eligible entities, verify each entity’s information and identity, and attach verifiable sources to every result. Typical tasks include competitor research, due diligence, literature retrieval, market analysis, and talent search. Prior to this, Fable 5.1 held the top spot with a score of 0.601 and $12.76 per task, while Opus 5 scored 0.537 at $11.60 per task. Astra’s performance has notably lifted the benchmark score, and this gain was not driven by inflated costs.