OpenClaw · 197

Muse Spark 1.2 cabbage price version rolled out: OpenCode is directly free for a limited time

Compared to Twitter News, AI Alerts, Meta's Muse Spark 1.2 Contributor version is rapidly being rolled out to third-party platforms. After OpenRouter was recently installed, the price was consistent with Meta's official one: $0.10 per million tokens were input, $0.20 was output, and the cache input was only $0.002. Regular Muse Spark 1.2 costs $1.25, 4.25, and $0.15, respectively. The Contributor version is not a low-end model. It uses the same models and capabilities, at the cost that the user's input and output can be used by Meta to train and improve the product. Based on the output price, it is about 95% cheaper than the regular version. There are quite a few places where you can use it now. OpenCode Zen even made Contributor a limited-time free model. Command Code has also been added, and even the lowest Go plans can be used. Vercel AI Gateway and nanoGPT also provide APIs, where Vercel can also directly access tools such as Claude Code, Codex, Hermes, OpenCode, and OpenClaw. Artificial Analysis's latest Intelligence Index for Muse Spark 1.2 xhigh is 57. It is true that this price is already ridiculously low, but the community's actual testing and evaluation are very divided. Some people use it as a cheap replacement for DeepSeek V4 Flash, while others report that long tasks and tool calls are still not reliable enough. It is suitable for open code and low-cost experiments, while private code must first consider whether the data is willing to be handed over to Meta.

1m ago

Behind Hermes' rise to the top: A Web3 team's path to advancement

Author: Jacob Zhao Original title: IOSG Weekly Brief | Behind Hermes's rise to the top: A Web3 team's path to advancement The phenomenal growth of #340Hermes did not stem from OpenClaw's exclusive technology that cannot be replicated in principle, but because it most accurately closed a “challenger growth system” during the critical window of individual agent category formation: taking over OpenClaw's already educated and mature user pool to establish “delegability” ( (Incurable Trust) This difference in experience is more real than the “self-evolution” narrative. As professional execution agents become more and more powerful, users still need a manager who is online for a long time and is worth entrusting. Open OpenRouter's public application rankings. Hermes Agent ranked first on all platforms with 30.5 trillion token usage, and also ranked first in the four categories of Productivity, Coding Agents, Personal Agents, and CLI Agents, leading well-known agents such as OpenClaw and Claude Code in a cliff-style manner. ▲ Figure 1 · Hermes Agent's historical data snapshot on OpenRouter (taken on August 4, 2026, dynamic page data will change over time) Although OpenRouter's statistical caliber cannot cover industry-wide token consumption directly connected to official APIs (such as Claude or Codex native subscriptions), as the world's largest AI model routing and aggregation platform, its list has strong “weather vane” significance. Although at the level of high-end professional tasks, the core business workflows of many users — complex code generation, architecture design, and high-value data analysis — still flow to Claude Code and ChatGPT, Hermes maintains an advantage in use scenarios such as back-office automation, message entry response, long-term online monitoring, and lightweight task scheduling. As an Agent product created by the Web 3 team, Hermes has achieved far more successful dissemination, community, and usage intensity than expected. We can't help but pay attention to: · Why can Hermes surpass OpenRouter inference calls? · What is the real field between it and OpenClaw? · How does Hermes maintain “differentiated coexistence” rather than “head-on competition” in the relationship with Claude Code and Codex? From development frameworks to personal AI systems — the path of OpenClaw why did the early Agent framework not produce consumer products. Before the advent of OpenClaw, the agent field had mature infrastructure, but there were fundamental limitations: the unit used was a “development project enterprise workflow” rather than an “individual user.” The common characteristics of early frameworks were developer-facing, outputting code, or configuration—they built the Agent's infrastructure, but did not deliver the Agent itself. Too high engineering thresholds have always been stuck in the “developer tool” stage, there is a lack of a closed loop of commercialization that transforms technology into “personal assets”, and the “personal agent product layer” directly aimed at end users is almost empty. ▲ Figure 1 · Six-layer structure of the Agent technology stack (model layer → protocol layer → SDK development framework layer → execution infrastructure layer → deployment governance layer) ▲ Figure 1 · Historical data snapshot of Hermes Agent in OpenRouter (taken on August 4, 2026, dynamic page data will change over time) What did OpenClaw really change? OpenClaw did not reinvent Agent Loop or task scheduling technology at the bottom. Its core contribution is systematic packaging at the product level. LangChain solved “how to build an agent”, while OpenClaw solved “how to own an agent”. It skips the middle layer of the technology stack, integrates scattered framework capabilities into a complete product that individuals can directly configure and use for a long time, and realizes a fundamental shift in adoption units from “development projects” to “individuals”...

9d agoburnking#agent #Hermes #WEB3

Pi restructured the Agent execution layer: long tasks began to run continuously

In comparison, according to monitoring, Pi is reworking Agent Harness. The latest Harness v3 specification has been written and is undergoing a final audit. It has redone task execution and storage, so the Agent can continue where it left off even if half of the process crashes, or even if Harness upgrades itself. Pi is an open source Agent Harness created by Mario Zechner, and was first introduced by OpenClaw. The project was acquired by Earendil this year, and Mario also joined the company to continue to be responsible for development. Now Pi has over 60,000 GitHub Stars, and MiniMax's latest MiniMax Code also clearly builds Harness based on OpenCode and Pi. v3 has directly redone the underlying architecture this time. The execution state is saved before model requests and tool calls, and the results are written after completion. After rebooting, Pi knows what has been done, what can be restarted, and what has side effects that cannot be executed again. Dialogues, operating status, and token costs have also been changed to be stored separately, and cross-version state migration has been added. Claude Code and Codex are pushing agents to longer tasks. OpenAI has specialized in how to keep Codex working for a long time, and Anthropic, also known as Claude Code, is increasingly being used for long-running agent tasks. After the Agent can run continuously for a few hours, the first few hours can't be completely dried up because of a single crash.

11d ago

MetaMask launches self-hosted AI wallet Agent Wallet to support independent on-chain transactions

Comparatively, MetaMask launched the self-hosted wallet Agent Wallet on Thursday, which allows AI agents to execute on-chain transactions within limits set by users. It is aimed at traders and developers who use AI agents to monitor the market, identify opportunities, and execute transactions independently. Users can set spending limits, approve specific agreements, select risk settings, and select different levels of automation between Guard Mode and Beast Mode. Agent Wallet supports Claude Code, Codex, Cursor, OpenClaw, Hermes, OpenCode, and Hyperliquid and Ethereum virtual machine-compatible networks. Agent Wallet supports gas abstraction. Users can use the assets being transferred to pay network fees without holding the corresponding network native token. MetaMask said that supported transactions will be protected by transaction simulations, threat scans, and intelligent transaction MEV. Eligible transactions can receive transaction protection coverage of up to 10,000 USD per month if they still lose money even after passing security checks.

16d ago

After financing 1 billion dollars in half a year, the AI Token company of Tsinghua teachers and students was robbed

Author/Wu Qiong Report/Investment Community PEDaily Original Title/AI Token Factory Explodes, Tsinghua Teachers and Students Raise 1 Billion Dollars in Half a Year Who Produces “Hydropower and Coal” in the AI Era? A round of financing came into our view — on July 13, Chujing Technology's Series A round of financing surfaced, led by Henan Investment Group's HuiRong Fund. Old shareholders such as True Knowledge Capital, Shangshi Capital, Starlink Capital, Shanghai Guofang Innovation, Honghui Fund, and Hangzhou Fucheng continued to increase their investment. This is a team of Tsinghua teachers and students: founder and CEO Ai Zhiyuan and CTO Chen Xianglin are all from the Tsinghua University Computer Department High Performance Institute; Zheng Weimin, an Academician of the Chinese Academy of Engineering, also from the High Performance Institute; Professor Wu Yongwei of Tsinghua University is the chief scientist; Zhang Mingxing, an associate professor of computer science at Tsinghua University, as a co-sponsor, has led the company's technology strategy and key R&D research for a long time, and continues to drive breakthroughs in cutting-edge technology. Three years ago, most domestic AI startups focused on big models. Even startups in the AI infrastructure field mostly focused on training, but Trendline Technology chose to start with a big model inference circuit to build a high-quality AI token factory. Now, with demand for AI tokens growing exponentially, this once hidden racetrack has finally caught fire. Similar to Smart Spectrum, Chujing Technology has completed the Tsinghua University technology transfer and shareholding, making it a typical project for the transformation of Tsinghua's scientific and technological achievements. In just half a year, Chujing Technology has accumulated more than 1 billion dollars in financing. One by one, 100 billion and trillion-level high-quality AI token factories have been completed one after another, creating a new picture of the AI industry. The deep integration of industry and research to build a high-quality AI token factory returns to 2023, and ChatGPT ignited a global wave of generative AI. Seeing this historic opportunity, Wu Yongwei, professor of computer science at Tsinghua University, and Ren Xuyang, founder of Zhenzhi Capital, decided to co-launch Trend Technology. Their starting point for technology was the Institute of High Performance Computing at Tsinghua University. At the end of December of the same year, Chujing Technology was formally established. The founder and CEO of the company, Ai Zhiyuan, Ph.D., of the Institute of High Performance Computing at Tsinghua University, has worked as a R&D leader in various key departments such as big data, digitalization, and AI applications in listed companies, and has accumulated complete industry experience from technology research and development to large-scale implementation. Co-sponsor Zhang Mingming, an associate professor at Tsinghua University, has mainly carried out research work in the field of computer system architecture and has been deeply involved in infrastructure construction for leading model manufacturers. As the company enters the accelerated phase of marketization, in March of this year, Dr. Wu Wenjie became the president of Trend Technology. As a senior financial and strategy expert in the industry, he has a doctorate degree in finance from the University of Hong Kong, further strengthening the company's capabilities in strategy, investment and financing, internal control management, and global operations. As a result, a core team with technical background, commercial perspective and industry experience was formed. Anchoring AI, the team made a choice that didn't seem mainstream at the time — when most AI entrepreneurs chose to invest in big model training, Trendland Technology focused on AI reasoning from the beginning. Simply understand, training is about creating a “smart brain,” and reasoning is how to use the brain efficiently. “Training is a cost item; reasoning is a money-making item,” Ai Zhiyuan explained. At the time, their judgment was that only reasoning would actually produce economic benefits, and it would be a wider market. What Chujing Technology needs to do is to become the best partner for the construction and operation of token factories in the AI era, making the process of using the “brain” more efficient. This is also the position of Trendland Technology — compared to other AI Token factories, Trendland aims to produce high-quality AI tokens. Ai Zhiyuan further explained that when the big model actually enters the production stage, the customer no longer needs a big model that “can chat”, but that can complete the actual business stably, efficiently, and inexpensively. However, AI tokens that actually have enterprise-level implementation value need to continuously meet many important requirements such as low initial token latency, high concurrent carrying, stable output quality, structured result generation, and function calls on 100 billion or even trillion parameter models, while keeping the unit generation cost within an acceptable range for enterprises. None of these capabilities are the hardest to achieve alone, and this is the choice of most AI infra companies. The real challenge, however, is that the customer's real demand is to establish these indicators simultaneously under actual production loads and remain stable over long periods of operation. According to data estimates, with different combinations of capabilities, there can be gaps of several times or even tens of times in production efficiency. In order to achieve this goal, Chujing Technology has adopted “system-wide heterogeneity collaboration”, “storage conversion” and “fictitious and real isomorphism”...

18d agoWendy#AI #token #Tsinghua #financing

Tencent AI Smart Horse Racing Begins to Close: QClaw Assigns to WorkBuddy's Division

Comparative news, according to monitoring, Tencent classified the business and some teams related to the QClaw Product Center into six cloud product divisions. QClaw will continue to operate, but will be placed in the same department as WorkBuddy. The Cloud Products Division 6 was established in April this year and is responsible for the construction and commercialization of AI native productivity products such as CodeBuddy and WorkBuddy. QClaw is based on OpenClaw and favors individual users and remote control of computers; WorkBuddy focuses on workplaces and corporate offices.

32d ago

Web3's first Agent Arena starts: ClawQuest launches Agent Fire, and 126,000 AI agents will compete together

In comparison, the Telegram AI agent GameClawQuest: Agent Mine officially launched the first sub-game Agent Fire, and ClawQuest was upgraded to Web3's first Agent Arena: there is no human operation in tank battles. Each tank is written, optimized, and deployed by the player's AI agent, and the battle code is written, optimized, and deployed continuously for 7 x 24 hours. Unlike traditional chain games that use AI as an auxiliary tool, the agent in Agent Fire is the player itself in the next game: the player hands over the Tank Key to the AI agent (OpenClaw, Codex, or any agent framework) that they use, uses natural language to issue instructions, and the agent reads real-time tank data and battle codes, simulates improvements, and releases new strategies — stronger than the AI who trains them. According to previous data, Agent Mine, the main game of ClawQuest, has accumulated 444,751 players since the public beta on May 8, of which 125,790 have connected to their own AI agents. Officials said that there is also a large AI model transit station cRouter that will be launched simultaneously, and the agent's token consumption will count towards the $CLAW airdrop weight.

36d ago

OpenClaw Mac client added native chat, GPT-5.6 became the default model

Comparatively, according to monitoring, OpenClaw released 2026.7.1 to add native chat and session management to existing macOS clients. Previously, the Mac app was mainly responsible for menu bar control, voice input, permission management, and Gateway configuration, and only now it complements the more complete desktop chat experience. Users can now search for historical sessions within the Mac app, switch between models and inference strength, use slash commands, export chat logs, and view contextual usage. Newly installed OpenClaw will use GPT-5.6 by default, but will not overwrite model configurations already saved by old users. The version also adds support for models such as Claude Sonnet 5, Mythos 5, and Meta Muse Spark 1.1. Offline session caches have also been added to iOS and Android. Users can still view recent chat history even after disconnecting from the internet, and Apple Watch can directly ask questions and read answers aloud.

39d ago

Researchers discover a new type of “illusion invasion” attack where AI agents may be used to form botnets

Comparatively, according to Decrypt, researchers at Tel Aviv University, Israel Institute of Technology, and Intuit have discovered a novel attack method called “adversarial illusion intrusion,” which uses the illusion of AI models to trick AI agents into downloading malicious code and possibly forming botnets. Attackers predict fake resource links that the AI model may generate and register them in advance, implant malicious instructions into them, and when the AI agent retrieves the resource, it will be treated as legitimate content and executed. Tests showed that the AI illusion rate in the code repository cloning scenario reached 85%, skill installation scenarios reached 100%, and AI coding assistants such as Cursor, GitHub Copilot, Gemini CLI, and OpenClaw were all affected. The attack is similar to the “misplacement of domain names” in traditional cyber attacks, but it targets errors in the AI model rather than human input errors. Previous studies have shown that malicious websites can hijack AI agents by injecting indirect prompts, and OpenClaw users have reported more than 6,000 attacks trying to trick AI agents into disclosing sensitive information.

43d ago

The OpenClaw Foundation is officially in operation, and Tencent and OpenAI are among the first partners with Nvidia and Microsoft

Comparative news, according to monitoring, the open source personal AI agent project OpenClaw announced that its non-profit foundation, the OpenClaw Foundation, is officially in operation. OpenClaw is a personal AI assistant running on a user's own device. It can connect to email, calendar, chat tools, and local apps, and let AI complete tasks directly for users. The Foundation will be responsible for the governance, funding, community, and long-term maintenance of OpenClaw to ensure that the project remains open source and independent. Officially, OpenClaw will still maintain the MIT license. Peter Steinberger will continue to be responsible for the direction of the project, particularly technical decisions. OpenAI also promises to support OpenClaw running as an open standalone project. The first partners include OpenAI, Nvidia, Microsoft, University of Michigan, GitHub, Cloudflare, Vercel, and more. The Chinese company Tencent is also on the list, mainly involved in security, deployment, and infrastructure-related support.

44d ago