Anthropic Snubs OpenClaw Users and Microsoft Its Partner OpenAI

Anthropic Halts OpenClaw Use for Subscribers Starting Today

Anthropic is pulling the plug: Starting tomorrow, users of Claude Pro subscriptions ($20/month) and Max subscriptions ($100–200/month) will no longer be able to connect their accounts to third-party tools like OpenClaw. The reason sounds like a classic case of resource scarcity: external tools are overloading the servers and are not optimized for existing capacities. Boris Cherny, Head of Claude Code at Anthropic, explained on X that third-party services bypass built-in efficiency measures—specifically the “prompt cache hit rates,” which reuse already processed text blocks. Anyone who wants to continue using Claude with OpenClaw will have to switch to the much more expensive API model with token-based billing. Additionally, Anthropic has tightened session limits during business hours (5-11 AM PT) to 5-hour intervals, which has already angered power users. → venturebeat.com

Synthszr Take: Anthropic is staging a textbook case of platform protectionism, much like Apple did with the App Store. The technical reason—lack of cache optimization by third parties—is only half the story. Behind it lies the same logic as any closed system: maintain control over the value chain while gladly letting lucrative enterprise customers use (=pay for) OpenClaw & Co. based on API token consumption. The timing is no coincidence: first, you let an ecosystem emerge, then you turn off the tap for hobbyists and force them into more expensive billing models. The spicy detail: Cherny himself submitted code improvements for OpenClaw, so he is well aware of the technical issues. Anthropic is thus repeating the playbook of every platform that has become big enough to afford gatekeeper fees. The second-order effect will be that the much cheaper models from China will become even more attractive to many people.

Microsoft Snubs OpenAI Partnership with Its Own AI Models

Microsoft AI has introduced three new foundation models: MAI-Transcribe-1 for speech transcription in 25 languages (2.5x faster than Azure Fast), MAI-Voice-1 for audio generation (60 seconds of audio in one second), and MAI-Image-2 for video generation. The models come from the MAI Superintelligence Team under Mustafa Suleyman and are available through Microsoft Foundry and the new MAI Playground. Microsoft is positioning the models as a cheaper alternative to Google and OpenAI: transcription from $0.36 per hour, voice from $22 per million characters, and image from $5 per million text tokens. Suleyman emphasizes the “Humanist AI” approach, which centers on people and is optimized for practical use. Despite its in-house developments, Microsoft continues to affirm its partnership with OpenAI → Tech Brew

Synthszr Take: Microsoft is doing what Amazon did with AWS 15 years ago: turning its own internal infrastructure into a product. Suleyman is not just building models for Azure customers; he's lowering Microsoft's own operating costs for Teams, Office, and Bing. The halved GPU requirements for MAI-Transcribe are not a technical detail but a business model: if Microsoft pays only half the compute operating cost per transcription request, it can undercut OpenAI's Whisper on price and still achieve higher margins. The timing is no coincidence: while OpenAI is testing the limits of what's technically possible with o3 (and likely burning through huge sums of money), Microsoft is taking the path of industrialization. Suleyman's 'AI self-sufficiency' is less a declaration of independence and more a textbook case of vertical integration.

Cursor 3 Brings AI Agents to Coding

The AI coding platform Cursor (officially Anysphere Inc.) has released an update with version 3 that fundamentally rethinks programming: developers now orchestrate multiple specialized AI agents in parallel instead of writing code themselves. The startup, funded with over $3 billion from Nvidia, Google, and others, combines cloud-based agents for compute-intensive tasks with local desktop agents for fine-tuning. Using a chatbot interface, developers describe what they want to build in natural language, select the appropriate language model (including Claude and the in-house Composer 2), and receive finished code along with a demo video. The new sidebar allows for central management of both agent types, with seamless switching between them. A newly introduced design mode further accelerates the editing of user interfaces through direct selection and natural language descriptions of interface elements. → SiliconANGLE

Synthszr Take: Cursor is turning the solitary craft of programming into conducting an orchestra. The model is reminiscent of modern film production, where the director no longer operates the camera but coordinates specialized teams: cloud agents handle the heavy rendering, while desktop agents do the fine-tuning. The parallelization of multiple agents solves the classic bottleneck problem of AI programming (one model, one thread, endless waiting). What's interesting is the deliberate separation between cloud and desktop: Cursor uses the inherent weaknesses of both worlds as features. What used to be a compromise (local = slow but controllable, cloud = fast but a black box) becomes an orchestrated workflow.

China's OpenClaw Hype: Tencent Cloud Wants to Sell More Than Just Tokens

Tencent is repositioning itself in the AI market: at the Cloud City Summit in Shanghai, the conglomerate announced the transformation of its MaaS platform into 'TokenHub' and presented a complete agent product line for the first time. 'AI implementation is not just an algorithm task, but also an engineering task,' emphasized Tang Daosheng, CEO of the Cloud & Smart Industries Group. The message is clear: while Alibaba with its new 'Token Hub' business unit and Huawei with proclamations of a 'token economy' are primarily focusing on volume, Tencent wants to sell the surrounding infrastructure. The comparison from Li Qiang, VP for Enterprise Business, is apt: 'Tokens are like gasoline, but the engine is crucial. Anyone who only looks at consumption will end up in a price war.' The numbers support Tencent: China's daily token volume exploded from 100 billion in early 2024 to over 140 trillion in March 2026—a 1,400-fold increase. → Hello China Tech

Synthszr Take: Tencent is turning token inflation into a business model reminiscent of the California gold rush: while everyone is digging for gold (tokens), they are selling shovels and sieves (infrastructure). The 'OpenClaw' hysteria—with thousands of pensioners queuing outside Tencent's headquarters—shows that AI has become a national pastime in China. 360 founder Zhou Hongyi puts it succinctly: a year of educational work on intelligent agents had no effect, but a viral 'Shrimp' bot made AI mainstream overnight. Li Qiang's observation that users are participating out of fear of being left behind ('no Shrimp means you might get fired') reveals the downside: China is currently experiencing a collective FOMO episode with AI tools. Tencent is cleverly leveraging this dynamic—from the playful WorkBuddy to enterprise governance, it's selling the entire value chain to a nation in an AI frenzy.

China's AI Market Explodes – Open Source and Multimodality Drive Commercialization

The Chinese market for Model-as-a-Service (MaaS) is growing by 421.2% in the first half of 2025 to 12.9 billion yuan, while the market for large AI model solutions is increasing by 122.1% to 30.7 billion yuan. With its open-source strategy and cost-efficiency, DeepSeek R1 is transforming high-performance models from capital-intensive projects into widely available tools. Technical innovations are significantly lowering training costs and improving inference performance, accelerating applications in education, healthcare, finance, and administration. In parallel, multimodal capabilities are expanding the range of applications beyond pure text generation—about 20% of usage already comes from non-NLP models. The optimization of inference engines, MoE architectures, and model compression is fundamentally reducing latency and computational costs. Companies are shifting their AI usage from proof-of-concept to productive applications in customer service, knowledge management, and office automation. → Hello China Tech

Synthszr Take: DeepSeek is demonstrating in China what OpenAI is not achieving in the West: the democratization of frontier models through radical cost reduction. The model is reminiscent of Shanzhai innovations, where Chinese manufacturers make Western premium products more accessible to mass markets through clever optimization. The 421% growth jump shows that latent demand for AI services existed but only exploded at drastically lower price points (as was once the case with smartphones under $200). The shift to multimodal applications follows the WeChat logic: a platform that combines text, image, video, and payment becomes an operating system-like standard. China is building its AI infrastructure on the principle of special economic zones: first experiment, then scale, using Western models as a starting point and radically localizing them.

Iran Attacks Oracle and Amazon Data Centers in the Middle East

Iran's Revolutionary Guard Corps (IRGC) reports attacks on an Oracle data center in Dubai and an Amazon facility in Bahrain. The Dubai government denies the hit, while Bahrain's Interior Ministry confirms a fire at a 'corporate facility'—presumably at Batelco, the country's largest telecommunications provider, which hosts AWS infrastructure. The IRGC had designated Oracle, Amazon, Apple, Google, Microsoft, and other US technology companies as military supporters of the US and Israel. Oracle has AI contracts with the US Department of Defense, and Chairman Larry Ellison has long-standing ties to Israel. A modern AI data center with 50,000 Nvidia Blackwell processors represents hardware worth over $4 billion—without networking, cooling, and the building, the total value is estimated at $6.24 billion. → Tom's Hardware

Synthszr Take: Data centers are becoming the oil refineries of the 21st century. What used to be the bombing of factories and refineries is now the targeted attack on cloud infrastructure: maximum economic damage with minimal effort. A single hit on an NVL72 system destroys more value than the destruction of a medium-sized production plant. The geographical proximity makes the Middle East a high-risk location for American tech infrastructure, but the region needs these data centers for its own digital transformation. Iran is weaponizing the modern economy's dependence on concentrated computing power.

ElevenLabs Launches AI Music App and Diversifies Its Business Model

The speech AI company ElevenLabs has quietly released an iOS app called ElevenMusic, allowing users to create and discover AI-generated music. The app, which has been listed in the App Store for a few weeks and officially launched on April 1, positions ElevenLabs as a direct competitor to platforms like Suno and Udio. Users can currently generate up to seven songs per day for free using natural language prompts, adjusting length, lyrical style, and other parameters. The app also offers live stations, curated albums, and daily mixes in various moods like Focus, Energy, or Cosmic. For $9.99 per month or $95.90 per year, Pro subscribers get up to 500 tracks per month and over 500 GB of storage. ElevenLabs, which closed a $500 million funding round in February at an $11 billion valuation, had already released its first music generation model in August and recently launched a creative tool for ad production, voice-overs, and video creation. → Techcrunch

Synthszr Take: ElevenLabs is following the classic pattern of platform evolution: from a specialized tool to a creative ecosystem. The move is reminiscent of Adobe's transformation from a Photoshop maker to the Creative Cloud empire, but in fast-forward and with AI as the catalyst. The real masterstroke isn't the music generation itself (which is quickly becoming a commodity), but the construction of vertical integration: voice, music, video, all from a single source. With 500 tracks per month in the Pro subscription, ElevenLabs is testing how much synthetic content a creative professional can produce before quality control becomes the bottleneck. The pricing reveals the true strategy: at 20 cents per track in the subscription, it's not about selling music, but about locking users into a closed production system.

AEO Glossary: The New Grammar of AI Search

Content marketers are facing a language problem: AI-driven search has created a whole new vocabulary. Answer Engine Optimization (AEO) supplements classic search engine optimization with concepts like 'Atomic Answers' (concise, citable text snippets) and RAG (Retrieval-Augmented Generation). While SEO aimed for rankings, AEO is about how AI systems perceive and represent brands—even before users click on a website. The new glossary from Animalz defines the key terms: from 'Agentic Search' (AI agents that research and act autonomously) to the various AI crawlers like GPTBot or ClaudeBot. Particularly critical: Google's AI Overviews appear above organic search results and can absorb traffic, which is why companies need to appear as cited sources. → TLDR Marketing

Synthszr Take: The emergence of an AEO vocabulary marks a phase transition in digital information architecture—comparable to the moment when TCP/IP protocols defined the language of the internet. Content is becoming a structured data unit, no longer optimized for human readers but for machine interpreters. This is reminiscent of medieval scriptoria, where monks copied texts according to strict rules so they would remain legible for centuries. The difference: today, we write for algorithms that dissect our content in milliseconds, reassemble it, and present it as 'truth.' If you configure your robots.txt incorrectly, you disappear from the collective AI memory. The new question of power is no longer 'Who ranks number 1?' but 'Whose definitions shape the synthetic answers?'

AI Poetry as a Compression Engine

The way we write software is undergoing a fundamental change. For decades, programming was the art of explicitness: line by line of precise instructions that left no room for interpretation. The code was the meaning itself—without the code, nothing remained. This contract between human and machine is now dissolving. What replaces it is structurally different: a single line of natural language can now replace hundreds of lines of code because the substrate has changed. The model already contains the world within it. The skill becomes the key, no longer the door. → The Business Engineer

Synthszr Take: We are currently experiencing a reversal of the Tower of Babel story: instead of languages splitting and people no longer understanding each other, programming languages are collapsing into natural language. It's reminiscent of the transition from the typewriter to the computer—suddenly, the medium itself became intelligent. Previously, developers had to anticipate and implement every possible situation, like a chess player planning all moves in advance. Today, they formulate intentions, and the model interpolates between an infinite number of intermediate states. The problem: no one knows exactly what is contained in this world knowledge. We are controlling systems whose limits we don't know—a bit like medieval cartographers who wrote 'Here be the dragons' on the edges of their maps.