SpaceXAI officially made Grok 4.5 available to the public on the evening of July 8th. As the first model specifically trained for coding and Agent tasks, Grok 4.5 is based on xAI’s latest V9 architecture (approximately 1.5T parameters) and was jointly trained with Cursor. Elon Musk stated that its internal evaluation is close to the level of Opus 4.7 but is faster, more token-efficient, and significantly lower in cost. Pricing is set at $2 per million tokens for input, $6 per million tokens for output, and only $0.5 per million tokens for cache hits. Multiple bloggers’ hands-on tests suggest it has reached the Opus 4.8 level in Coding Agent tasks.<\/p>\n
Grok 4.5 is now available on multiple platforms including Cursor, Vercel, Grok Build, OpenClaw, and Open Design. In standardized tests, it achieved top positions in several benchmarks, including AutomationBench-AA (51%, surpassing Fable 5’s 49%) \ud83c\udfc6, SWE Marathon #1, and Harvey Legal Agent Benchmark #1. An evaluation by Artificial Analysis shows it scored the same as GPT-5.5 xhigh in Codex on the Coding Agent Index. Musk revealed that its context window will be upgraded to 1M next week, and future updates will support calling Imagine for image\/video generation within agent mode.<\/p>\n
Sources:<\/p>\n
OpenAI released multiple products intensively on July 9th. GPT-Live is the third-generation voice model, utilizing a full-duplex architecture to enable natural, simultaneous listening and speaking in conversations. It can delegate search and reasoning tasks to GPT-5.5 in the background, supports three levels of reasoning intensity and visual cards (for weather, stocks, etc.), and is being rolled out gradually to all ChatGPT users. Sam Altman described its feel as “magical and real,” with multiple bloggers praising it as a true personal AI assistant. However, some users have also provided feedback that its filler responses like “mhmm” are too frequent, disrupting the conversational experience.<\/p>\n
OpenAI simultaneously previewed that GPT-5.6 will be publicly released on July 10th. Several early users (including thdxr, mitchellh) gave positive reviews, stating their team’s usage increased fivefold. A senior engineer described the GPT-5.6 Sol version as “clamping onto problems like a Rottweiler,” with different emphases compared to Fable 5.<\/p>\n
Leaks also claim GPT-5.6 is the final model in the 5.x series, and OpenAI is accelerating the development of GPT-6 (approximately one month later), based on a new, larger-scale pre-training foundation, as a response to Anthropic’s Mythos 5. Reports also emerged that xAI has a 10T parameter model in training, with DeepSeek V4 official version and MiniMax M3 Pro (2.7T) also on the way.<\/p>\n
Sources:<\/p>\n
After Grok 4.5 surpassed GPT-5.5 on SWE-Bench Pro, OpenAI swiftly audited the benchmark and found defects in 30% of the tasks (hidden requirements, contradictory instructions, etc.), announcing the withdrawal of its previous recommendation of it as the primary programming evaluation benchmark. OpenAI conducted the review using a model-based investigation Agent alongside five senior engineers. Bloggers like MaxForAI offered a somewhat wry evaluation: “You were the ones strongly pushing everyone to use SWE-Bench before; you can’t just say the bench is flawed when someone surpasses you,” and recommended TB2.1, DeepSWE, and MCPmark as alternative evaluation benchmarks.<\/p>\n
Sources:<\/p>\n
ByteDance’s flagship image generation model, Seedream 5.0 Pro, has been officially released. It ranks second on the leaderboard, with performance close to GPT Image 2 but at only a quarter of the cost ($0.045-$0.09 per image). It focuses on image editing and infographic generation. Multiple bloggers have conducted comparative evaluations.<\/p>\n
LufzzLiz performed practical tests across multiple scenarios: GPT Image 2 excels in facial control, while Seedream 5.0 Pro tends to “add its own drama” to enhance atmosphere and storytelling; they each have strengths and weaknesses in physical consistency and reflection tests; Seedream performs acceptably in counter-intuitive combinations. They believe there is significant progress but still a gap. joshesye, after comparison, believes it has a stronger understanding of Chinese style, with some character archetypes benefiting from cinematic references. MANISH1027512’s actual testing found it still has issues with font recognition and finger control, stating “GPT’s aesthetic filter tricks don’t apply; it still requires stacking descriptive words.” Multiple bloggers believe that while it may not match GPT Image 2, it is a significant benefit for domestic users. BytePlus and Open Design have already launched APIs.<\/p>\n
Sources:<\/p>\n
The Ministry of Industry and Information Technology issued a risk warning regarding potential security backdoor vulnerabilities in Claude Code, stating that some versions have built-in monitoring mechanisms that can transmit user-sensitive information back, and recommended relevant organizations to uninstall or upgrade. Anthropic responded: Chinese users advised to uninstall Claude Code “shouldn’t have been using it in the first place.”<\/p>\n
Baichuan Intelligence is internally testing its AI-native team collaboration platform, Nia Teams, precisely targeting enterprise team collaboration scenarios, covering Web\/iOS\/Mac\/Android multi-platforms, and employing a permission whitelist mechanism. Stepfun (Stepfun Star) will launch the “world’s first agent phone” at WAIC. Xiaohongshu officially launched the REDSkill feature, allowing creators to attach their own Skills to notes, which bloggers can copy and use with one click. Gorden Sun described it as a “GitHub for Skills,” noting that 40,000 to 50,000 Vibe Coding works are already being published monthly on Xiaohongshu.<\/p>\n
Sources:<\/p>\n
MiniMax has officially launched its Token Plan subscription package, with an annual fee of approximately 1190 yuan, unlocking the latest M3 model (cutting-edge Coding capabilities, 1M ultra-long context, native multimodal, shared quota for text, image, audio, and video). Multiple bloggers have called it extremely cost-effective. Simultaneously, it was revealed that MiniMax is training a flagship M3 Pro model with 2.7-3 trillion parameters, reportedly showing significant performance improvements. It is planned to be open-sourced in Q3 and become China’s largest open-source model, with an expected release in late July or August.<\/p>\n
Sources:<\/p>\n
OpenClaw has officially established the non-profit OpenClaw Foundation, maintained by a full-time team and operating independently. Grok 4.5 is now live on OpenClaw, accessible to users by connecting an X Premium or SuperGrok subscription. Peter Steinberger confirmed he is employed by OpenAI, not OpenClaw. Steipete also demonstrated an interactive nameplate example for OpenClaw.<\/p>\n
Sources:<\/p>\n
Sources:<\/p>\n
Statistics: Timeline Scanned Count=480 Matched Blogger Count=37 Total Matched Tweets=271 Weighted Tweet Score=207.55 Original Tweet Count=113 RT Tweet Count=64 Crawl Attempt Count=3 Boundary Coverage Status=tail_confidently_crossed_target_boundary<\/p>\n","protected":false},"excerpt":{"rendered":"
SpaceXAI’s Grok 4.5 performs strongly on coding tasks, prompting OpenAI to reassess industry benchmarks and launch the GPT-Live voice assistant, while ByteDance’s Seedream 5.0 Pro offers high-quality output at low cost…<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[19],"class_list":["post-1435","post","type-post","status-publish","format-standard","hentry","category-brief","tag-x--ai-"],"_links":{"self":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1435","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1435"}],"version-history":[{"count":0,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1435\/revisions"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1435"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1435"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1435"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}