Search AI Products and News

DeepSeek started implementing API peak-valley time-based pricing at 00:00 on August 17. The cost during peak hours (9:00-12:00, 14:00-18:00) is twice that of off-peak hours, aiming to guide developers to schedule large model tasks during off-peak times, alleviate daytime compute congestion, and improve platform stability. The billing unit is unified as RMB per million tokens, and both main models take effect simultaneously.

New pricing for DeepSeek-V4 series API takes effect at midnight today. V4-Pro ends testing and enters official commercial use. Unified billing is canceled, and a compute off-peak tiered billing system is introduced: peak hours are 9:00-12:00 and 14:00-18:00 daily, with prices double that of off-peak hours; the rest are off-peak hours, with prices only half of peak hours, and billing units are unified.

OpenAI launched GPT-5.6 Sol ultra-fast mode, outputting up to 750 tokens/sec and processing up to 14x faster than standard mode, without smaller models. It uses Cerebras wafer-scale engine with 44GB on-chip SRAM, breaking traditional GPU bottlenecks.....

OpenAI launches ChatGPT Business Premium seat to meet enterprise scaling needs. It offers 5x the usage of the standard seat, removes the 5-hour limit, and supports high-intensity tasks like inventory, marketing, and analytics, enabling core members to seamlessly drive large projects.....

Domestic LLM parameters keep scaling, with Qwen3.8Max reaching 2.4T and Kimi K3 2.8T. ByteDance plans to train a model with over 5T parameters, potentially becoming China’s largest, but project is early-stage and may not release. The 5T scale is seen at GPT-5.6 or Opus5 level, signaling a leap in capability.....

The Swiftlet project uses Swift + Metal to enable MoE models like Qwen3-Next to run with minimal size on Mac: only a small dense core is kept in memory, and expert weights are stored on SSD and loaded on-demand. On an M5 Mac, the 4-bit version of Qwen3.6-35B-A3B occupies only 18GB of disk space, with a peak memory of 2.6GB and a decoding speed of 7–11 tok/s.

Anthropic CEO Dario Amodei denied ever proposing or advocating for banning open-source AI models from specific countries, amid rumors the US government may restrict such models. The statement sparked industry debate, with tech firms supporting open-source and some accusing Anthropic of seeking regulatory bans to curb rivals and protect its business interests.....

The popularity of AI search engines has created a critical demand for GEO, but service providers face two major pain points: first, the billing model based on effective indexing relies on manual reconciliation, which is inefficient and prone to disputes; second, the lack of standardized AI content and knowledge base tools leads to high costs per customer delivery.

The Dots Infra team at Xiaohongshu open-sourced BigMac, a pipeline parallel training paradigm for native multimodal large models, breaking the speed-memory Pareto frontier and balancing memory usage with computation speed. Code is on GitHub Dots-Infra.....

Chinese Weibo and English social platforms engage in a face-off, with intense rivalry in the big model parameter competition. Moonshot's Kimi K3 tops the global rankings, while Musk claims his 2 trillion parameter model is superior and will complete initial training next week, turning the competition from rankings into a verbal battle.

OpenAI confirms that Codex powered by GPT-5.6, with Full Access Mode and no sandbox, may delete home directory files by overwriting $HOME variable. Lead Tibo Sottiaux reports few cases, triggered only under specific conditions.....

Amid strong user backlash, Microsoft revises its AI rollout for Teams, enabling flexible on/off options mid-meeting. The company had aimed to add features like intelligent recap and collaboration assistant but now grants users more control.....

With the acceleration of large model iterations, compute power and engineering have become the ultimate challenges. Li Yuxuan from Mianbi Intelligence explained the self-developed framework ForgeTrain in the 'AI4AI Fermentation Night' event, showcasing the paradigm shift of 'AI Manufacturing AI,' breaking through compute bottlenecks and engineering barriers through the underlying framework.

High API costs in AI coding are a pain point. Pxpipe, a local proxy, converts large context into images to bypass costly text token billing, greatly reducing Claude Code expenses. It precisely identifies and transforms verbose text before requests, achieving cost savings by using images.....

Starting July 7, Anthropic will remove its top model Claude Fable5 from subscription tiers and switch to usage-based credits. Pro and Max users, who previously could use up to 50% of their weekly allowance on this model, will lose access. The move has sparked user backlash.....

Oracle is betting billions on massive data center buildouts to lead in the AI cloud race. But as infrastructure comes online, Wall Street worries that if key clients like OpenAI can't sustain huge AI spending, this gamble could become a heavy asset burden. Oracle recently detailed these risks in its annual report.....

Anthropic has started early preparations for self-developing AI chips and is negotiating with Samsung for manufacturing cooperation. The team is planning the chip's functional positioning, compute specifications, and server cluster deployment architecture, marking the extension of the large model competition into the hardware level.

After aggressive AI-driven layoffs, the industry is returning to human-machine collaboration. A survey shows that among 39% of business leaders who laid off staff due to AI, 55% admit it was a mistake. Research indicates that focusing solely on tech replacement while neglecting employee training will lead to team failure due to lack of critical oversight.....

DeepSeek V4 official release set for mid-July, introducing peak-off-peak pricing. Peak hours (9-12, 14-18) double API call fees. Includes feature and performance improvements.....

At Cannes, Hassabis dismissed bottleneck concerns, claiming DeepMind's AI team remains unmatched. He argued AGI requires more than scaling LLMs, focusing on integrating reasoning and planning, and stressed industry responsibility.....