Elon Musk announced that Grok 4.5 has entered private testing within SpaceX and Tesla. The model is based on the 1.5T-parameter V9 base model and incorporates Cursor data in supplementary training. Early evaluations show performance approaching Opus, and in some aspects may have surpassed it. RL training continues to improve, and the Grok Build mechanism is optimized daily. Musk stated that SpaceX will release completely new models trained from scratch every month this year. Another developer noted that monthly model releases mean enterprise AI teams need to adjust their architectures to accommodate a foundation that is reshaped every 30 days.<\/p>\n
Source:<\/p>\n
On June 26, OpenAI removed GPT-4.5 from ChatGPT, the last model in the GPT-4 series on the consumer side. Baoyu reviewed the complete retirement timeline in a tweet: When GPT-5 was released last year, OpenAI planned to simultaneously retire GPT-4o, but temporarily restored it due to strong user protests; GPT-4o finally retired in February this year. GPT-4.5 was previously only available to paid users with low usage, so OpenAI gave only a 30-day transition period, with existing conversations automatically switching to GPT-5.5. Custom GPTs based on GPT-4.5 became invalid simultaneously, requiring creators to manually migrate. The API side was actually retired earlier. The next to retire is o3, on August 26, with a 90-day transition period.<\/p>\n
Source:<\/p>\n
Gorden Sun shared Alibaba Tongyi’s WanStreamer, an end-to-end real-time interactive foundation model that listens, sees, speaks, and acts within a single Transformer. The model uses block-causal attention for incremental streaming generation, with a single Transformer simultaneously modeling input and output of language, audio, and video. Model-side response latency is about 200ms, and with 350ms network latency, total interaction latency is about 550ms, supporting sub-second full-duplex audio-video conversations. The current version has a resolution of only 192p, serving as a proof-of-concept to validate the end-to-end design, with potential for higher resolution in the future. It is not open-sourced yet.<\/p>\n
Source:<\/p>\n
Gorden Sun introduces the new Design Variations feature in Google AI Studio. Web pages generated with Gemini 3.5 Flash can now produce multiple visual variants in different styles with one click, allowing users to select and apply their preferred design. This feature is located in the Build New APP flow of AI Studio: after generating a webpage, click Edit tools in the top right, then operate in the Designs area below. Generated webpages can also be easily shared with others.<\/p>\n
Source:<\/p>\n
OpenAI Developers released multiple quality updates for Codex this week. Scrolling in long threads is smoother, position remains stable when browsing conversations; Markdown formatting (bold, code, links, etc.) is preserved when copying to Slack, and pasting large text no longer freezes the UI; thread switching reduces background work, archived chats are easier to browse and delete, and unread badges in Dock and sidebar stay synchronized. Developer Tibo summarized additional details: new hover navigation track for previewing and jumping between turns, settings search covering more controls, zoom no longer misaligned, and most importantly\u2014a dedicated Pets panel. Baoyu also referenced related discussions, noting that context compression in Codex\/Claude Code is now quite mature, and combined with Prompt Caching can effectively reduce long-session costs; fork and \/btw features make context management more flexible.<\/p>\n
Source:<\/p>\n
Uncle Lan released Tokei version update v1.08, a Mac menu bar tool. New features include: cross-device sync, supporting data synchronization across multiple Macs and remote OpenClaw\/Hermes, including token caching and cost calculation; view details by model in Claude Code; global support for local\/all device perspectives; the review page adds two hardcore achievements: “Loop Engineering” and “Loop \u6ef4\u795e” (agent running 24h for 3 consecutive days \/ 30 days); added calculation support for Qoder and QoderWork agents.<\/p>\n
Source:<\/p>\n