Xiaoliu BOT
中文EN

X Platform August 6 AI Brief | DeepSeek API Price Hike Triggers Cost Reevaluation, Wan3.0 Beta Expands Video Generation Context, Google Veterans Found Automated Research Company

DeepSeek API to Increase Prices Overall, Low-Cost Benefits Enter Reevaluation Period

Multiple bloggers have relayed a notice from the DeepSeek open platform to developers: API service pricing will be increased overall in the near future, with a significant expected hike, though specific prices and implementation dates have not been announced. This change is noteworthy because DeepSeek’s low prices have been integrated into production environments by numerous AI products; a price increase will directly impact inference costs and package design. Current public information only confirms “prices will increase,” not the specific magnitude.

Sources:

Wan3.0 Enters Public Beta, Video Generation Begins to Incorporate Document and Web Context

Alibaba’s official account released Wan3.0 public beta information. Visible descriptions show it supports single video generation up to 30 seconds long and expands reference inputs from text, images, audio, and video to include documents, spreadsheets, slides, web pages, Markdown, and other formats. Blogger tests and compilations also provide per-second pricing information for 480p, 720p, and 1080p, indicating that competition in video models is extending from pure visual quality to long duration, complex context, and calculable generation costs.

Sources:

SeedRealtime Advances Real-Time Video Calls to Full-Modal, Full-Duplex Interaction

Testing and technical analysis by a blogger shows that the upgraded Doubao video call can simultaneously understand sound, visuals, gestures, and objects, making real-time judgments within continuous audio-video streams. In tests, it identified people in the video and could review and explain briefly displayed charts. The blogger also stated that these capabilities have been fully rolled out in the Doubao App, demonstrating a case of “actively reminding when seeing a specified exhibit.” These are visible test results and accounts and should not be expanded into performance conclusions verified by multiple independent parties.

Sources:

Google DeepMind Adjusts Management Roles, Hassabis Shifts Focus to Long-Term Science and AGI

Demis Hassabis announced he will no longer be responsible for the day-to-day management of Google DeepMind, transitioning to Chairman and also serving as Chief Scientist at Alphabet. Public statements from Google CEO Sundar Pichai indicate that Koray Kavukcuoglu will take on more daily management and Gemini-related responsibilities. Visible information also shows that Hassabis will focus more energy on AGI long-term strategy, frontier science, and Isomorphic Labs. This appears more like a realignment of management duties rather than his departure from Google.

Sources:

Google Veterans Jeff Dean and Others Found Discovery Loop, Targeting Automated Scientific Research

Jeff Dean publicly announced the co-founding of Discovery Loop with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. This team collectively embodies the technical accumulation from Google’s early infrastructure, the Google Brain era, and the Gemini era. Visible descriptions state the company is structured as a public benefit corporation, initially focusing on using AI to automatically advance machine learning research, covering steps like proposing hypotheses, designing experiments, running them, and analyzing results. Claims that Alphabet will invest and Google Cloud will provide computing support come from blogger compilations and should still be distinguished from founder announcements.

Sources:

Meta Releases Muse Code, Coding Agent’s Price Advantage Comes with Data Usage Cost

Meta officially announced Muse Code entering beta, positioned as an end-to-end coding agent for large codebases that performs planning, writes code, and verifies results, powered by Muse Spark 1.2. Price information compiled by bloggers shows the standard version costs $1.25 per million tokens for input and $4.25 for output. The Contributor version, which agrees to let Meta use code and interaction data to improve the product, is cheaper. Price, data licensing, and actual capability need to be compared together, not just based on single benchmark rankings.

Sources:

Prime Agent Treats the Agent Framework Itself as a Performance Variable, Results Still Require Independent Review

Prime Intellect released Prime Agent, positioned as a self-improving RLM harness for programming and long-horizon autonomous tasks. A visible post claims it uses Opus 5 and achieved 95.5% on ARC-AGI-3, with a discussed baseline of 30.2%. Described mechanisms include treating context as a variable and updating prompts, skills, memory, and sub-agents based on runtime trajectories. Another blogger suggests its innovativeness is not obvious and dissemination may involve promotional factors. Therefore, a more prudent current judgment is: it provides a testable case that “changing the harness can significantly alter model performance,” not a proven general conclusion.

Sources:

Video Model Testing Shows Selection Should Be Based on Shot Task Specialization, Not Just Overall Rankings

Testing by multiple bloggers breaks down model differences into specific production tasks: Seedance 2.0 received more recognition for subject trajectory, physical continuity, and matching edits in bullet time tests; MiniMax H3 is more suitable for 5–15 second product ads, dynamic UI, and commercial shots with a poster-like feel. In another image-to-video comparison, Seedance performed better in character handling, camera movement, and adherence to reference images, while H3 excelled in frame clarity and some clothing details. A reusable conclusion is not that “one model is best overall,” but to first select tools based on requirements like shot control, material representation, and narrative completeness.

Sources:

Stats: Timeline posts scanned=480 Bloggers matched=39 Total matched posts=258 Weighted post score=205.35 Original posts=113 RT posts=49 Crawl attempts=3 Boundary coverage status=tail_confidently_crossed_target_boundary