After Anthropic released its new model Fable 5, the developer community quickly reported a dramatic increase in token consumption. A member of an MTS team consumed the equivalent of $1,500 in 10 hours and hit usage caps three times; half of the team reached quota limits during engineering work. SemiAnalysis tests showed that the $200\/month Claude Max subscription plan can actually consume API credits worth about $8,000, far exceeding the previously expected $2,000 cap. Meanwhile, Fable 5’s refusal mechanism for high-risk tasks has drawn criticism\u2014users reported the model actively refusing to work on topics like cybersecurity, bioscience, model distillation, and jailbreaking. Anthropic later apologized and announced adjustments: flagged requests will fall back to Opus 4.8 and be visible to users, with the API returning refusal reasons, but the refusal policy itself will not be removed. Many bloggers see this as “apologizing without changing.” Fable 5’s capabilities have also spawned new workflow practices: one user demonstrated using Fable to write code and command-line tools to complete a full video editing pipeline, including Whisper transcription, AI shot selection, FFmpeg rough cut, code color grading, Remotion animation, and Figma MCP collaboration.<\/p>\n
Sources:<\/p>\n
Alibaba saw multiple major changes on June 11. DingTalk CEO Chen Hang stepped down, replaced by Chen Yusen, a tech-savvy serial entrepreneur born in 1992 who founded Chaitin Technology (acquired by Alibaba Cloud) and internally developed the AI Agent product MuleRun in 2025. The day before, Alibaba’s partnership committee posted an internal memo titled “Loyalty, Integrity, and Growth: That’s Alibaba Culture,” sharply criticizing DingTalk’s management style as “not what Alibaba culture should be,” emphasizing that AI-era innovation relies on passion and creativity rather than high-pressure execution. Meanwhile, multiple bloggers shared layoff rumors, claiming Alibaba is cutting 35% of its workforce, with some tech departments seeing 50% cuts. On the AI organizational front, on June 8, Alibaba merged the Tongyi Large Model Business Unit and the Future Life Lab to form the TokenFoundry Business Unit, directly overseen by CEO Wu Yongming, with Zhou Jingren transitioning to Chief Scientist leading the AI Future Research Institute. This marks Alibaba’s third AI restructuring in three months, aiming to integrate models, products, and token consumption scenarios into one streamlined line.<\/p>\n
Sources:<\/p>\n
ChatGPT’s half-price discount for the Turkey region has been canceled, and the previously available Plus plan at about 80 RMB\/month is no longer available. One blogger lamented “how genius programmers fall,” linking it to the price hike to 999 lira in Turkey. On a broader level, SemiAnalysis tests reveal a pricing dilemma: the $200\/month Claude Max 20x plan can consume about $8,000 worth of API credits, while ChatGPT Pro 20x membership can reach $14,000, far exceeding subscription revenue. SemiAnalysis believes that as intelligence costs rapidly decline, model vendors are more likely to remove new features and models from subscription plans rather than directly weaken existing subscription features to avoid public backlash. Notably, the upcoming Mythos may ultimately offer only API access without a subscription plan.<\/p>\n
Sources:<\/p>\n
Google has open-sourced DiffusionGemma, a 26B-parameter sparse mixture-of-experts (MoE) model that activates only 3.8B parameters during inference. Its core innovation lies in the generation method: traditional LLMs output tokens one by one, while DiffusionGemma simultaneously generates text blocks of 256 tokens, refining them through iterative denoising, shifting the bottleneck from memory bandwidth to computation. It achieves 1000+ tokens\/second on H100 and 700+ tokens\/second on RTX 5090, delivering up to 4x inference speedup. After quantization, it can run on consumer GPUs with 18GB VRAM. Google acknowledges that overall output quality is lower than standard Gemma 4, positioning it as an experimental model for speed-sensitive, local low-concurrency scenarios like code completion and structured generation. Some analysts believe the combination of Diffusion and JEPA could be a bigger direction than pure diffusion language models\u2014using JEPA to understand world states, Diffusion to sample multiple possible futures in latent space, and a cost model to select the optimal path.<\/p>\n
Sources:<\/p>\n
Anthropic CEO Dario Amodei published a lengthy essay titled “Policy on the AI Exponential,” systematically outlining AI’s impact on regulation, economy, military, and geopolitics. The article proposes five key points: First, regulation should upgrade from transparency disclosure to access review, with frontier models undergoing pre-release third-party testing like aircraft and pharmaceuticals. Second, AI may simultaneously bring extremely rapid productivity growth and long-term cognitive labor substitution; policy should focus on how to let more people share in growth dividends, including employment monitoring, wage insurance, UBI, and universal capital accounts. Third, downstream industry regulation needs acceleration; the FDA\/EMA’s 7-8 year drug approval cycle will be choked by AI-driven surges in new drug candidates. Fourth, AI could become a tool for authoritarianism; autonomous weapons require constitutional and court oversight, and data broker loopholes should be closed. Fifth, build a democratic AI-sharing alliance, sharing chips and semiconductor equipment, and restricting adversaries’ access to critical supply chains. Some analysts believe Anthropic is transitioning from a model company to an institutional designer of AI order, with this essay layering regulatory narrative, industrial moat narrative, and geopolitical narrative.<\/p>\n
Sources:<\/p>\n
DeepSeek posted a job opening for an Agent Harness Researcher, believed to be the first time globally that “Harness Researcher” appears as a job title. The team is positioned as “Model + Harness = Agent,” responsible for context management, long-term memory, sub-agents, multi-agent systems, and self-evolving agents. Xiaomi simultaneously open-sourced MIMO Code, forked from OpenCode, featuring cross-session memory, context reconstruction, sub-agents, goal determination, and workflow orchestration, using tree-structured task IDs and judge model verification. Apodex-1 was released as a multi-agent deep research framework, replacing single-agent loops with an orchestrator-plus-sub-agents-plus-global-validator architecture, coordinating up to 150 sub-agents for 15,000 steps per task, surpassing GPT-5.5-pro on benchmarks like BrowseComp. Baichuan released the M4 medical agent, comprising three core components: a harness layer, a reasoning model, and a clinical tool layer. These developments collectively point to a trend: AI competition is shifting from single model capability to collaborative optimization between models and runtime frameworks.<\/p>\n
Sources:<\/p>\n
U.S. real estate tech company OpenDoor announced the closure of its India operations, laying off its entire offshore team of over 200 people, and is replacing them with smaller, AI-native teams based in the U.S. This case is seen as a signal of AI’s direct impact on the outsourcing industry model. Commentators note that when AI tools can replace a large amount of offshore development and operations manpower, companies may accelerate the contraction of overseas teams back home, using fewer people with stronger AI tools to accomplish the same or more work. This contrasts with the same day’s discussion on Fable 5 token consumption: on one hand, AI is replacing human labor; on the other, the cost of using AI itself is rising rapidly, and companies need to find a new balance between the two.<\/p>\n
Sources:<\/p>\n
Blogger @xiaohu0x shared their first overseas SEO experience: using a practice mini-game as a project, they reached 1 million unique visitors and 2.9 million page views in 30 days. Although the tweet itself is an Article type requiring a click-through for details, this data has drawn attention in the Chinese developer community, with many bloggers reposting and bookmarking it. The case provides a concrete signal: mini-games as SEO carriers still have significant traffic acquisition potential in overseas markets, especially for independent developers looking to validate overseas SEO methodologies as a low-cost trial path.<\/p>\n
Sources:<\/p>\n