1｜Anthropic Releases Fable 5, Begins Graded Rollout of Cutting-Edge Capabilities
Anthropic has released Claude Fable 5, its first Mythos-class model open to the general public. According to reports from The Verge, Axios, and Business Insider, Fable 5 is Anthropic's most powerful publicly available model to date, targeting software engineering, knowledge work, and visual tasks. It shares the underlying architecture with Mythos 5 but incorporates routing and degradation mechanisms for high-risk requests related to cybersecurity, biology, chemistry, and model distillation.
This is not just a regular model upgrade but an experiment in distributing cutting-edge capabilities. Anthropic had previously warned about the risks of self-improvement and is now making a stronger model available to the masses while locking dangerous capabilities behind graded access. It aims to prove one thing: models can be made stronger, but capabilities cannot be unleashed indiscriminately all at once. The future of AI companies may lie not in selling a single model but in a set of access control mechanisms. Whoever can define which capabilities are given to whom is defining the societal boundaries of cutting-edge AI.
(Source: Anthropic / The Verge / Axios / Business Insider)
2｜Cohere Open-Sources Coding Agent, Small Models Continue to Vie for On-Premises Adoption
Cohere has launched North Mini Code, its first open-source agentic coding model aimed at developers. According to the company's blog, this model uses an MoE architecture with a total of 30B parameters, activating 3B per token, supporting 256K contexts, and capable of running on a single H100 card at FP8. VentureBeat sees this as Cohere's clear entry into the open-source coding agent market.
When viewed together with the Fable 5 announcement, it becomes clear that the AI model competition is splitting into two tracks. One track is exemplified by Anthropic and similar cutting-edge models, which release stronger capabilities through permissioned and secure routing. The other track is represented by Cohere and similar small open-source models, pushing for controllable deployment, low hardware requirements, and local agent workflows for developers. The real concern for enterprises is not which model is better at writing code but which set of models can be integrated into their endpoints, repositories, and on-premises workflows.
(Source: Cohere / VentureBeat / Hugging Face)
3｜US Strikes Iran, Ceasefire Shifts from Diplomatic Window to Military Risk
Al Jazeera, AP, Washington Post, and Guardian report that the US launched strikes on Iran after an Apache helicopter was downed. Trump stated that two pilots were rescued, but the US had to respond. CENTCOM later confirmed carrying out targeted strikes, including radars and air defense systems. US officials emphasized that this was a proportional response and not seeking all-out war.
This marks the next phase of the June 8 Iran-Israel conflict. Previously, the US was urging restraint on Israel, trying to preserve the negotiation window with Iran. Now, the US itself has entered the chain of strikes. The core of Middle East risk is no longer just a cycle of retaliation between Israel and Iran but how the US maintains boundaries between negotiation, deterrence, and energy pathway security. A helicopter near the Hormuz has pushed the ceasefire back into military logic.
(Source: Al Jazeera / AP / Washington Post / Guardian)
4｜Lovable's Annual Revenue Hits $500 Million, Vibe Coding Starts to Enter Revenue Books
TechCrunch reports that Lovable has announced an annual revenue exceeding $500 million, with 1 million new projects added weekly, totaling over 50 million projects. Its user base is mainly non-technical, but an increasing number of projects are being used for monetization or internal software replacement. On the same day, 36Kr wrote about another direction, where context-mode is attempting to optimize context management through MCP to reduce AI programming costs by 98%.
With both signals combined, the competition for vibe coding is shifting. The early question was whether natural language could be turned into an app, but now the question has become whether the app can be maintained, costs can be contained, and non-technical users can turn projects into real business. Lovable's revenue proves demand, while context-mode points to cost bottlenecks. AI programming is not just replacing programmers; it is shifting software production from professional teams to the business site.
(Source: TechCrunch / 36Kr / Yahoo Finance)
5｜AI Power Pressure Spills Over, Car Batteries Redefined as Grid Assets
The Verge reports that GM is advancing vehicle-to-grid technology, aiming to have electric cars play a buffering role in the power pressure brought by AI data centers. The company has also announced a commercial energy storage strategy, including sodium-ion batteries for industrial grid applications. On the other hand, the Nashville Zoo is opposing the conversion of a nearby 1.6-acre plot into a data center, stating that the land was originally planned for an educational and conservation center.
This is not the same project, but it points to the same issue. The power needs of AI data centers have already begun to rewrite the roles of other industries. Car companies are not just selling cars; they are starting to turn batteries into grid assets. Zoos are not just public spaces but are also being drawn into the competition for data center locations. The externalities of AI infrastructure are spreading. It needs electricity and also land. The result is that roles not originally in the AI industry chain are being forced to participate in the cost allocation of computing power expansion.
(Source: The Verge / Tom's Hardware / GM Energy)
Also worth noting ↓
SpaceX unveils 11 million square feet Gigasat factory, aiming to achieve 1GW of space AI computing power by the end of 2027. This announcement continues the track record data center narrative from the SpaceX IPO, indicating that the company is incorporating AI computing power into its manufacturing capacity showcase. (Source: Tom's Hardware)
SemiAnalysis tracks the performance evolution of the DeepSeek V4 1.6T model from Day 0 to Day 43. Open-source model competition is no longer just about the initial scores but about how subsequent engineering teams continue to optimize inference performance and hardware compatibility. (Source: SemiAnalysis)
Sunpoint Technology completes nearly a hundred million yuan financing round, betting on AI chip processor IP. 36Kr reports that the company hopes to lower the barrier to customized processor design using RISC-V+DSA and IP shelves. The verticalization of AI models is driving the verticalization of chip design in reverse. (Source: 36Kr)
Court approves Visa-Mastercard settlement with 12 million merchants over swipe fees. The card organizations' fee dispute is entering the settlement phase. The power structure of payment infrastructure still revolves around who bears the transaction costs. (Source: Payments Dive)
Klarna launches a savings account, further delving into banking services. BNPL companies are no longer satisfied with installment payment gateways but are migrating toward full-fledged financial accounts. The boundaries between payment companies and banks continue to blur. (Source: Payments Dive)
Google releases Gemini 3.5 Live Translate. Real-time voice translation is moving from demo scenarios to more devices and application entry points, and multimodal models are increasingly taking over everyday communication layers. (Source: Ars Technica)
Microsoft AI Lead Mustafa Suleyman criticizes Anthropic's suggestion that Claude is conscious. The two companies' diverging narratives on AI risk continue to widen. One emphasizes the boundary of capabilities, while the other emphasizes avoiding anthropomorphism. (Source: The Verge)
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
