CZ personally attended the Hong Kong meet-and-greet for the book *Binance Life*.

Beating AI News Flash: Google has launched Gemini 3.5 Transcribe, its most accurate speech-to-text model to date, and the first dedicated transcription model from Google to feature a Smart mode. The model comes in two versions—real-time and audio recording transcription—and is now open for public beta testing via the Gemini API. In an evaluation by Artificial Analysis, the non-streaming Word Error Rate (WER, lower is better) stands at 2.6%, while the real-time version hits 4.0%. Final transcription latency is 70% shorter than that of its predecessor, Chirp 3. The Smart mode goes beyond literal word-for-word transcription: for example, if a user says “Meeting on Tuesday, no, Wednesday,” it will retain only “Wednesday”; filler words like “um” and “ah” are automatically removed, and spoken content can be organized into paragraphs, lists, dates, and numbers. For verbatim meeting notes, users can switch back to the default literal transcription mode. The model supports over 85 languages, allows mid-transcription language switching, and supports adding custom domain-specific vocabulary. Pricing is approximately $0.005 per minute for audio transcription (equivalent to $5 per 1,000 minutes) and $0.009 per minute for the real-time version. In independent benchmarks, Gemini 3.5 Transcribe is more accurate than OpenAI’s GPT Transcribe (3.3% WER) but lags behind ElevenLabs’ Scribe v2, which has a 2.2% WER and costs roughly $3.67 per 1,000 minutes. The model is already integrated into Android’s Rambler and the macOS version of Gemini, and will soon roll out to Chrome, enabling direct voice-to-text input in web text boxes.

Glassnode’s market analysis shows that during Bitcoin’s current rebound window, U.S. spot Bitcoin ETFs recorded a net inflow of $2.23 billion, marking the strongest weekly inflow since 2026. Bitcoin continues to flow out of exchanges, while wallets of all sizes are accumulating holdings. Over the same period, Bitcoin futures open interest fell 11% in coin terms, with funding rates remaining largely neutral, indicating no significant reaccumulation of market leverage. From a market structure perspective, Bitcoin is currently in a recovery phase, but upward supply pressure is building. Data shows cost-base resistance from self-custody holders starting around $80,800, a dealer gamma flip point near $82,300, and multiple resistances in the $82,000–$86,000 range including short liquidations, long-term holder supply, and order book sell orders. Glassnode notes that the $83,000–$86,000 range is the core supply zone facing this rebound; if Bitcoin can hold above $83,300 while ETF inflows continue, it may signal the market is gradually absorbing this supply wall. On the downside, the cost base for short-term holders is around $70,000, while the $62,000–$65,000 range forms a more critical bottom support zone. The options market is not pricing in a clear directional breakout expectation; as of the September 25 expiry, the options-implied 70% middle outcome range is roughly $69,000–$89,700, overall pointing to Bitcoin continuing to oscillate between key support and resistance levels. Glassnode believes a break below $70,000 would first test the $62,000–$65,000 support zone; a drop back to near $62,900 would likely mean this rebound has been fully reversed.

According to monitoring by TradingBeats (formerly Hyperinsight), the 0x0911 whale that opened a "buy ETH, sell BTC" pair position two days ago is still adding to its holdings. The nominal value of both legs has expanded from an initial ~$20 million to ~$29.38 million, a rise of ~46.9%. As of press time, the address holds a long position of 5,943.61 ETH worth ~$14.819 million, with an average entry price of $2,486.64, generating an unrealized profit of ~$39,000. It also holds a short position of 184.95 BTC worth ~$14.558 million, with an average entry price of $79,804, for an unrealized profit of ~$201,000. The sizes of both legs remain roughly balanced, with a total unrealized profit of ~$241,000. When the position was opened on August 25, BTC outperformed ETH by ~2.2 percentage points in the prior 24 hours, and the average execution price of the two TWAP trades corresponded to a BTC/ETH ratio of ~32.20. The current BTC/ETH ratio has dropped to ~31.57, a decrease of ~2% from that level. Currently, both the ETH long and BTC short positions are generating unrealized profits: if the BTC/ETH ratio continues to decline, the pair position will keep benefiting; if it rebounds back to around ~32.20, the existing unrealized profit will face a pullback, excluding transaction fees and funding rates for now. Previous news: BTC/ETH ratio falls back to 32.2, how does the $20 million whale expect the ratio to move?

Data shows that despite ongoing geopolitical tensions between the U.S. and Iran, shipping volumes through the Strait of Hormuz have seen a slight rebound. Kpler data indicates that 10 tracked commodity transport vessels passed through the Strait of Hormuz on Wednesday, a minor increase from 8 on Tuesday, but still below the 10-day moving average of around 15 vessels. Two medium-sized product tankers, one liquefied petroleum gas (LPG) carrier, one Panamax tanker, and three handysize tankers entered the Strait of Hormuz from the Gulf of Oman. One medium-sized product tanker, one bitumen carrier, and one bulk carrier exited the waterway from the Gulf side. Meanwhile, shipping volumes through another key waterway, the Bab el-Mandeb Strait, slowed for the second consecutive day. Kpler data shows that a total of 19 commodity transport vessels transited the Bab el-Mandeb Strait on Wednesday, including 6 exiting tankers, one of which is a very large crude carrier (VLCC). The total number of transiting vessels was lower than the 24 recorded the previous day.

Beating AI Express News: Zhipu AI’s flagship model GLM-5.3 will be open-sourced at 10 AM Beijing Time on August 28. GLM-5.3 was already launched on Coding Plan on August 14, with its API opened shortly after. The model weights are finally available for download tomorrow, two weeks behind schedule, for a specific reason. Zhipu AI stated that GLM-5.3’s cybersecurity capabilities improved “faster than expected” during post-training, with the model not only becoming more adept at identifying vulnerabilities but also learning to plan full multi-step exploit chains. The official thus conducted security assessments and hardening work before releasing the weights. In official evaluations, GLM-5.3 scored 84.5% in the vulnerability detection test CyberGym, slightly higher than Mythos 5’s 83.8% and GPT-5.6 Sol’s 83.6%. On ExploitBench, a test focused on attack exploitation, its score jumped from 24.4% for GLM-5.2 to 54.4%—though it still lags far behind Mythos 5’s 78.0%. Zhipu AI added that the model has already found 2,436 vulnerabilities across 269 open-source projects, with 1,097 classified as critical or high-severity. GLM-5.3 retains the same base model as GLM-5.2, with all performance improvements coming from post-training.

According to Bitget's market data, the share price gain of Hong Kong-listed Zhipu has widened to 10%.