Xiaoliu BOT
中文EN

X Platform June 29 AI Brief | DeepSeek V4 Pricing Adjustments, GPT-5.6 Grayscale Testing Attracts Attention, AI Quality Inspection Encounters Bottlenecks

DeepSeek V4 Official Version to Be Released in Mid-July, Introducing Peak-Valley Pricing

DeepSeek sent upgrade reminder emails to some users, confirming that the V4 official version will be released in mid-July, bringing more feature optimizations and performance improvements. After the official release, the API will simultaneously adjust pricing strategy, introducing a peak-valley pricing mechanism: the regular prices for V4 Pro and Flash remain at current levels, while peak prices double. After the adjustment, the output price for deepseek-v4-pro is 12 yuan per million tokens, and for deepseek-v4-flash is 4 yuan per million tokens. Multiple bloggers followed up on this news.

Source:

GPT-5.6 Grayscale Testing Attracts Community Attention

OpenAI’s Codex has been discovered by the community to be quietly routing some gpt-5.5 xhigh sessions to gpt-5.6-sol. This can be verified using the Juice test prompt (querying hidden thinking budget): normal gpt-5.5 xhigh returns 768, while gpt-5.6-sol returns 128. Several bloggers tested and confirmed they still haven’t received grayscale access. Xiaohu also speculated that GPT-5.6 might be officially released that evening. On the other hand, the Codex team said they entered a warroom over the weekend to investigate abnormal usage and reset usage limits for all users.

Source:

Claude Account Ban Wave Continues, Many Bloggers Switch to Codex

Many bloggers have encountered Claude account bans. Some bloggers said that Claude has been banning a large number of accounts recently, mostly triggered by bank cards and IP addresses, advising not to use third-party payments for main accounts. Some users had both their Pro and free accounts banned. Khazix stated that Claude Opus 4.8 has severe hallucination issues, causing code to break, and has returned to Codex. Another account experienced a Claude Code cross-account issue—where other people’s information and code changes inexplicably appeared in conversations. Claude subscription prices in the Nigeria region have doubled, and some users plan to return to WeChat Pay for renewal.

Source:

Tesla FSD v14 Lite Rolled Out to AI3 Hardware Users

Elon Musk confirmed that FSD v14 Lite is being rolled out to customers with AI3 hardware. Tesla’s AI team stated that this version distills the driving behavior of the AI4 v14 series into the AI3 camera and computing configuration, including improvements such as destination options and navigation updates. The AI3 computer has only about 15% of the effective memory bandwidth of AI4, and multiple users reported a significant improvement in experience compared to v12.6.4. Sawyer Merritt commented that AI3 Teslas over 7 years old now receive v14 Lite, possessing a more advanced autonomous driving system than most comparable vehicles.

Source:

AI Accelerates into IM: Doubao Integrates with Feishu, WeChat Xiaowei Tests Group Chat Analysis

ByteDance’s Doubao has begun grayscale integration with Feishu. Users can log in to the Feishu web version with their Doubao account and add Feishu friends. Some analysts believe that the integration of Doubao and Feishu essentially connects Doubao to ByteDance’s existing work relationship chain and organizational permission system, transforming it from an AI assistant into an AI messaging entry point and collaboration identity layer within ByteDance’s ecosystem. Meanwhile, the WeChat Xiaowei beta version has been found to analyze members’ roles in group chats based on group chat records—not just determining “active,” but analyzing topic attributes, mentions, and speech driving effects. Multiple bloggers pointed out that IM is becoming a core entry point for AI into human collaboration flows.

Source:

Jia Yangqing Leaves NVIDIA, Just One Year After Lepton AI Acquisition

According to a leak from SemiAnalysis, Lepton AI founder Jia Yangqing has left NVIDIA, less than a year after NVIDIA acquired the 20-person team for approximately $700 million. The report states that DGX Lepton operations did not meet Jensen Huang’s expectations, and also involved disagreements between the two parties over open-source commitments and product execution capabilities. Earlier this month, Hyperbolic announced that Jia Yangqing would serve as an advisor, providing experience and support in AI systems, cloud infrastructure, GPU market, and other areas.

Source:

Ford Recalls 350 Senior Engineers—AI Quality Inspection Falls Short

Over the past three years, Ford has quietly recalled 350 retired or departed senior engineers (internally called “grey beards”) because the AI quality inspection system failed to meet expectations. The senior engineers returned to mentor newcomers and fine-tune AI tools, identifying fault points before parts went into production. The results were significant: Ford regained the top spot among mainstream brands in the JD Power Initial Quality Study after 16 years, and is expected to save approximately $1 billion this year due to reduced warranty and recall costs. Ford has not abandoned AI and is still expanding its simulation tests to about 100,000. Similar AI implementation counterexamples include Klarna rehiring after a decline in all-AI customer service quality, and McDonald’s canceling AI ordering.

Source:

Open-Source Alternative to Figma Motion: Web Animation Tool Released

In response to Figma’s paid closed-source animation tool Figma Motion, a developer has released an open-source alternative that supports adding animations via conversation, 200+ mature animation templates, per-element timeline, multi-shot transitions, one-click export to multiple formats, and the ability to export any animation as a skill for AI reuse. It has integrated with 50+ design systems and is interoperable with the Open Design ecosystem.

Source:

Statistics: Number of scanned timeline entries=360, Number of hit bloggers=34, Total number of hit tweets=184, Weighted tweet score=137.5, Number of original tweets=83, Number of RT tweets=52, Number of fetch attempts=2, Boundary coverage status=tail_confidently_crossed_target_boundary