Xiaoliu BOT
中文EN

X Platform August 13 AI Brief | DeepSeek V4 Pro Relaunched with Major Price Adjustments, Harness Open-Source Preview Reconstructs Agent Runtime, Claude Browser Agent Achieves Cross-Device Synchronization

DeepSeek V4 Pro 0813 Re-released After Rollback, Agent Capability Upgrades and Significant Price Adjustments Emerge Simultaneously

Yesterday’s visible timeline indicates that the release of DeepSeek V4 Pro 0813 went through stages of information being removed from the official website, a suspected rollback, and then a re-launch. What warrants attention is not just the model’s scores, but also deployment stability and usage costs. Official DeepSeek announcements were relayed by accounts as: V4-Pro/Flash adding low, high, and max reasoning tiers, with upgrades for Agent workflows; the re-released version was also noted to support the Responses API and Codex adaptation. Regarding pricing, a blogger compiled peak rates as: cached input at 0.30 RMB, non-cached input at 9 RMB, and output at 27 RMB. Compared to the previously circulated figures of 0.025, 3, and 6 RMB, the increases are concentrated in caching and output. It’s important to note that the reasons for the rollback, performance anomalies, and price comparisons primarily stem from blogger observations and relayed information, and cannot substitute for official announcements or independent evaluations.

Sources:

DeepSeek Harness v0.1 Open Source Preview, Deconstructs Agent Runtime into Composable Plugins

The core value of DeepSeek Harness lies in its attempt to provide a recomposable Agent runtime, not just another Coding Agent. Visible official relay confirms v0.1 has entered Developer Preview, uses an MIT license, and is powered by the Cordis meta-framework; related introductions treat the model, tools, Skill, Session, Sandbox, Loop, Orchestration, and even UI as replaceable plugins. Actual testers also identified four presets: Standard, PTC, Minimalist, and Creative. PTC uses TypeScript to orchestrate multi-step tool calls, while Creative mode allows inspecting the runtime, experimenting with plugins, and defining Agent presets. Its impact is that developers can reorganize permissions, context, tools, and workflows around the model, but the project is still in early preview, and its stability should not be judged as that of a mature product.

Sources:

Harness Quickly Enters Practical Testing Phase, Transparent Traces and Cache Costs Become Early Highlights

After release, discussion quickly shifted from “what is it” to “can it be used.” One tester recorded completing a high-difficulty task in about 35 minutes with a 99% cache hit rate, and stated that for the same task in their test, Harness cost about $2 while Codex cost about $7; this is a single-account practical test and cannot be directly extrapolated as a stable benchmark. Another user completed WSL2 installation and began configuring a DeepSeek key, with a tutorial author providing the shortest path via npm installation and web profile. Multiple accounts also reported the project reaching tens of thousands of Stars within hours of release, indicating rising developer interest, but Star screenshots and personal experiences remain visible social media signals, not equivalent to real active users or production reliability.

Sources:

Discussion on Grok 4.6 Entering the Top Tier, Advantages Concentrated in Speed, Cost, and Agent Programming

Visible official relay states that Grok 4.6 shows significant improvement over 4.5 while maintaining the same price; a blogger compiled API prices at $2 per million tokens for input and $6 for output, and mentioned the addition of xhigh reasoning, 500K context, and enhancements for Coding and long-range Agent tasks. Another account cited scores from CursorBench, DeepSWE, APEX-Agents, etc., suggesting it is competitive in some Agent tasks and performance-cost ratio. Practical feedback is not one-sided: one user stated Grok 4.6 is faster, more comprehensive, and successfully completed a fix in a task where DeepSeek V4 Pro failed to show results; these are user experiences and should not be written as a comprehensive victory.

Sources:

Claude Browser Agent Begins Cross-Device Session Sharing, While Bringing Web Page Injection Risks to the Forefront

Claude officially disclosed yesterday that Claude in Chrome sessions can be continued on desktop, web, and mobile clients, with skills and connectors saved with the account; the sidebar uses the same Claude Cowork session, meaning tasks are no longer tied to a single device. The official announcement also cautioned that browser Agents can be induced by hidden instructions within web pages, explaining that while safeguards are being built, users are still advised to maintain basic operational habits. The former enhances cross-device continuous workflow capabilities, while the latter indicates that the primary risk for browser Agents has shifted from “whether they can interact with web pages” to “how to determine if web page content is manipulating the Agent.”

Sources:

Codex and Enterprise Agent Usage Continue to Expand, with Linux, Skills, and Plugins Becoming Key Entry Points

Information visible yesterday extends the competition for Agents from model capabilities to distribution and usage habits: OpenAI officially stated that the top 10% of organizations by enterprise usage call plugins about twice as often as typical enterprises and use Skills about six times as frequently; another account relayed that the Codex desktop experience has entered preview on Ubuntu, Debian, and Fedora. Related Codex accounts also claimed the active user count has surpassed 15 million and previewed a new quota reset. The first two items are official product or data statements, while the 15 million figure and reset come from social media relay, best viewed as growth signals; taken together, the competitive focus of the Agent ecosystem is shifting to cross-platform entry points, reusable skills, and organizational workflow penetration.

Sources:

Effective Paths for AI Video Generation Remain Short Shots, Reference Images, and Revisable Workflows

High-quality practical feedback shows that the stability of tools like Seedance 2.5 still depends on shot breakdowns, not simply extending duration. One creator’s practical test suggests 15-second scenes are most stable, 30 seconds are more suitable for music videos or short stories, and generating directly for 60 seconds or more carries a higher “blind box” risk; a more reliable approach is to prepare character, scene, and storyboard reference images and regenerate problematic segments individually. Another author shared long prompt and reference image techniques for Seedance 2.5, indicating that consistency control already has reusable methods; others have packaged AI generation and editing into a Skill, preserving DaVinci projects for manual review and continuation.

Sources:

Cordis Paper Grounds the Challenge of Self-Evolving Agents in Revocable Components and Local Recovery

Interpretations of a Peking University collaborative paper that appeared the same day around DeepSeek Harness focus not on model scores, but on how an Agent can avoid breaking itself after continuously modifying its own runtime. The visible abstract breaks the problem into Temporal Composability and Spatial Composability: component side effects need traceable inverses, and components must be revocable upon unloading; when dependencies change, only affected components are reactivated, paired with lifecycle management, dependency ordering, fault recovery, and transactional Hot Module Replacement (HMR). The interpretation also mentions that under conditions like independence and acyclic dependencies, dynamic loading and replacement can ultimately converge to a state equivalent to directly assembling the final configuration; this is a single account’s summary of the paper, best viewed as a research direction signal, and should not be taken as a claim that Harness has already implemented the paper’s findings.

Sources:

Statistics: Scanned timeline entries=480 Matched blogger count=41 Total matched tweets=298 Weighted tweet score=243.4 Original tweet count=135 Retweet count=46 Crawl attempt count=3 Boundary coverage status=tail_confidently_crossed_target_boundary