Meme crypto token project MEME has exceeded a market capitalization of $40 million, surging 36.7% in the past 24 hours.

Binance Wallet announced in a statement that from September 23 to December 22, 2026, users transferring USDT and other TRC-20 tokens via Binance Wallet to the TRON network will be eligible for a 0 gas fee promotion. The campaign is backed by TRON DAO, and after its conclusion, transfer fees will be reduced to 1 USDT per transaction.

Beating AI Express: Open-source computer-use infrastructure project Cua has open-sourced two S1 models specialized for computer use decision-making, and launched Cua-Bench-S1, a benchmark that evaluates models like Jev separately. This benchmark does not allow models to plan on their own. Each task provides the current interface and a set of pre-selected actions (e.g., "fill email", "check agree", "click submit", "skip"), with the model making only one selection per task. No re-screening of the interface or failure retries are permitted. This setup tests one core question purely: what is the next step? The two open-sourced models take different approaches: The Nano version has only ~855,000 trainable parameters, trained from scratch, and is specialized for scoring candidate actions. The 4B version uses Qwen3.5-4B as its base model, fine-tuned into the same "multiple-choice model" framework. The former prioritizes minimal size and speed, while the latter leverages the large model’s inherent comprehension for stronger generalization. Beyond common tasks like form filling, login, and search, the benchmark also includes security judgment, chess, Doom, and JevBench. The latter tasks are intentionally excluded from the models’ training data, to test their ability to make correct decisions outside familiar scenarios.

Insight Beating AI Express: The Information, citing people familiar with the matter, reports that Anthropic is in talks with Stream Data Centers to directly lease up to 1GW of data center computing capacity. The two sides are in early-stage negotiations, with no final size confirmed, and the deal could still fall through. If signed, Anthropic will become a direct data center tenant, deploying its own servers and chips—including Google and Broadcom’s TPUs, as well as NVIDIA GPUs. Google is also discussing providing credit guarantees for some of the projects. This differs from renting servers directly from AWS or Google Cloud: Anthropic will control its own chips, racks and network equipment, cutting out a layer of cloud provider markup, though it will bear the costs of equipment procurement and data center operations. Per estimates from data center developers interviewed by The Information, a full 1GW computing infrastructure would require at least $40 billion in capital investment. Anthropic has already directly leased around 401MW of data center capacity from TeraWulf this year, and last year announced a $50 billion investment in a custom data center in partnership with Fluidstack. If the Stream deal reaches 1GW, it will become one of Anthropic’s largest direct data center leases to date.

According to Coinglass data, Bitcoin has firmly broken through $86,000 and is now consolidating in a range. Altcoins and meme coins have also seen rotational trading activity, while the Crypto Fear & Greed Index has climbed above 70, entering the "Extreme Greed" territory for the first time in a year. However, current funding rates on major centralized (CEX) and decentralized exchanges (DEX) show no signs of market overheating. Funding rates for mainstream crypto asset contracts remain neutral to bearish; specific rates are detailed in the attached chart. BlockBeats Note: The funding rate is a mechanism set by cryptocurrency trading platforms to align contract prices with underlying asset prices, typically applied to perpetual contracts. It functions as a fund exchange between long and short traders—platforms do not collect this fee, and it adjusts the cost or return of holding contracts to keep contract prices closely tied to underlying assets. A 0.01% funding rate is the benchmark. When the rate exceeds 0.01%, it signals widespread bullish sentiment; when it falls below 0.005%, it indicates broad bearish sentiment.

Beating AI Insight News Brief: Meta’s personal AI agent Muse announced four upcoming connectors from Shopify, PayPal, Expedia, and Instacart within 24 hours. Once integrated, Muse can directly leverage these services for shopping, payments, hotel bookings, and grocery runs, no longer fully dependent on opening web pages itself. Shopify will integrate Shop Pay one-click checkout across all its stores; PayPal enables Muse to shop and complete transactions at PayPal merchants globally; Expedia allows Muse to select and book hotels; Instacart offers a streamlined experience: users simply say “Tacos for Tuesday,” and Muse will generate a shopping cart based on store inventory, then hand it off to Instacart for checkout and delivery. These partnerships address a critical gap for Muse: shifting the agent from “simulating website operations” to “websites proactively connecting their services to it.” The former approach (browser click simulation) can be blocked by sites at any time—Amazon has already implemented this—while the latter means product, inventory, payment, and fulfillment are all provided by partner platforms, with Muse only tasked with understanding user needs and initiating tasks. More integrations will reduce Muse’s reliance on browser click simulation.

Beating AI Express: Robotics data firm Midcentury has emerged from stealth mode and closed a $15 million seed round. Instead of manufacturing robot hardware, the company sells training data and simulation tools for robots. Midcentury has collected over 2 million hours of first-person human videos, covering more than 50 environments and 20,000 tasks. 90% of the footage shows human hands, annotated with 3D hand and body poses, depth, and other metadata. Simply put, robots learn by watching humans perform actions like grasping objects, opening doors, and organizing items. Unlike large models that can scrape the internet for text, robots lack an existing "action internet"—a gap Midcentury aims to fill. The firm also built a simulation platform called Matrix, letting robots run thousands of tests and training runs in virtual environments before operating on real hardware, cutting costs from repeated real-world trial and error.