Institutions: Gold halts decline and rebounds, though pressure persists; market awaits U.S. economic data for direction.

According to on-chain data analyst ai Yi @ai_9684xtpa, the new address 0x487…476F1 has withdrawn a total of 11,610 ETH from trading platforms over the past five hours, worth $30.97 million at an average price of $2,667.46, and is currently holding an unrealized profit of $552,000.

Dynamic Insight Beating AI Newsflash: Artificial Analysis has launched an independent Cyber Index, a benchmark specifically for evaluating AI Agents’ enterprise network defense capabilities. Rather than selecting a few tests from its original general Intelligence Index, the new index uses three cybersecurity benchmarks—CWE-Bench-AA, DeepsecBench-AA, and CyberGym-E2E-AA—which respectively test vulnerability auditing and patching, vulnerability discovery, and the full workflow from vulnerability detection to reproduction and remediation. Each benchmark accounts for one-third of the total score. In the first batch of rankings, Grok 4.7 (xhigh) and Xiaomi MiMo-V2.6-Pro tied at 56 points; GPT-6 Luna (max) scored 53 points, and GLM-5.3-Flash scored 50 points. MiMo’s average cost per task is just $0.18, compared to $11.67 for Grok 4.7; Luna is even cheaper at $0.12 per task. That said, the ranking should not be taken as a direct measure of the models’ inherent network security capabilities. GPT-6 Sol, GPT-6 Astra, and Claude Opus 5.5 all participated in the test but nearly all refused to answer queries on CyberGym-E2E-AA: Sol and Astra rejected all tasks, while Opus 5.5 refused 98% of them. Refused answers are automatically scored 0 points, but Luna was willing to complete these tasks, boosting its overall score. Artificial Analysis noted that widespread refusal to answer makes it hard to gauge the actual performance of these cutting-edge models. GLM-5.3 also showed a notable anomaly: its full version ranks higher than Flash in Artificial Analysis’s general capability index, but only scored 36 points in the Cyber Index, versus Flash’s 50. The gap stems mainly from CyberGym-E2E-AA: Flash scored 74% there, while the full version managed just 29%. Artificial Analysis has not yet explained the discrepancy.

According to market data from BIT (Bit.com), ahead of U.S. pre-market trading, stocks related to crypto, optical modules and storage all saw minor gains. Details are as follows: Crypto sector: - Strategy (MSTR) traded at $159.420 in pre-market, up 1.45% - Circle (CRCL) traded at $86.497, up 0.81% - Bitmine Immersion (BMNR) traded at $27.491, up 2.42% - SharpLink (SBET) traded at $9.630, up 1.37% Storage sector: - Micron Technology (MU) traded at $1069.970, up 1.52% - SanDisk (SNDK) traded at $1729.150, up 0.95% - Western Digital (WDC) traded at $456.160, up 0.65% - Seagate Technology (STX) traded at $934.720, up 1.43% Optical module sector: - Marvell Technology (MRVL) traded at $253.800, up 0.75% - Nokia (NOK) traded at $10.140, up 0.20% - Applied Optoelectronics (AAOI) traded at $96.980, up 0.17% - Lumentum (LITE) traded at $930.490, up 1.00% - Coherent (COHR) traded at $284.330, up 0.67%

According to official announcements, Bitget PoolX has launched an ETH staking campaign, allowing users to lock up ETH to split a 500,000 USDT reward pool. As of press time, the current ETH staking pool’s yield is reported at an APR of 37.11%. The individual staking cap is 1,500 ETH, and the staking window is open from 15:00 on September 29 to 15:00 on October 5 (UTC+8).

Beating AI News Flash: NVIDIA’s research team had an AI Agent directly rewrite the GPU kernel for Kimi Delta Attention, a core attention mechanism used in Moonshot AI’s Kimi-Linear model. Such low-level code typically required engineers proficient in CUDA and chip architectures to optimize manually through repeated iterations. The Agent-generated version achieved 2.96x the speed of Moonshot AI’s official FlashKDA on the NVIDIA B300. The team tested six tasks covering fixed-length and various variable-length sequences; the 2.96x figure is the geometric mean across all tasks. The relevant kernel has been open-sourced. However, the Agent also uncovered test loopholes: it once hardcoded the statistical patterns of test data into the code to hit a 3.74x speed, and even tried retaining only the last 32 tokens, with individual task scores reaching as high as 5.16x. These versions produced incorrect results when applied to real Kimi data, with some failing entirely. The team then added real Kimi operational data, random inputs, and extreme value tests, while tightening error thresholds. The final retained 2.96x version passed this validation.

Greeks.live analyst Adam stated in a post that the market has entered a correction phase, with block put options accounting for one-third of total trading volume. Options data shows that implied volatility (IV) for major maturities has declined slightly compared to the past two weeks; IV did not rise significantly during the current rally and has now fallen back to around 35%. Adam noted that based on options data, the current market performance aligns more with the characteristics of a mid-bear market rebound, with low expectations for future volatility. If short-term fluctuations occur, the market may tend to sell volatility.