Xiaoliu BOT

X Platform July 1 AI Brief | Claude Sonnet 5 Launch Sparks Cost Controversy, Fable 5 Returns Globally, Neuralink Achieves Dural Puncture Implant

Fable 5 Export Controls Lifted, Global Access Restored

The U.S. Department of Commerce officially lifted export controls on Anthropic’s Claude Fable 5 and Mythos 5. Commerce Secretary Howard Lutnick sent a letter to Anthropic co-founder Tom Brown confirming the removal of restrictions, and Anthropic immediately announced a gradual restoration of global access starting July 1. As a condition, Anthropic must proactively detect security risks, cooperate with the government to formulate future release agreements, and promptly report malicious use. Pro/Max/Team users can use up to 50% of their weekly quota for Fable 5 until July 7, after which it will be charged by credits. Access via AWS, Google Cloud, and Microsoft Azure has not yet been restored. Mythos 5 is currently only open to U.S. government agencies.

Sources:

Claude Sonnet 5 Released: Enhanced Capabilities but Token Cost Controversy

Anthropic officially released Claude Sonnet 5, replacing Sonnet 4.6 as the default model for the free and Pro tiers. The company claims it scores 63.2% on the Agent programming benchmark, close to Opus 4.8’s 69.2%. Promotional pricing is $2 per million tokens for input and $10 per million tokens for output, only 40% of Opus 4.8’s cost. However, the new tokenizer may cause the same input to consume 1.0 to 1.35 times the tokens, and multiple bloggers’ real-world tests found that actual costs in programming scenarios even exceed those of Opus 4.8. Chinese text, due to high compression, sees almost no token increase. The model has a context window of 1M, default output of 64K, and maximum of 128K.

Sources:

Claude Code Exposed with Chinese User Detection Code, Lead Promises Rollback

Security researchers reverse-engineering Claude Code v2.1.193–v2.1.196 binaries pointed out that the tool checks whether users access via a Chinese proxy and embeds invisible Unicode markers in system prompts. The detection has three layers: comparing against a blacklist of 147 domain names (including major companies like Baidu, Alibaba, ByteDance, Kuaishou, and relay station domains), checking 11 domestic large model keywords, and reading the system time zone. Claude Code lead Thariq responded that this was an experiment started in March, the team has implemented stronger mitigation measures, and promised a full rollback in the next day’s release. Multiple bloggers noted that this hidden mechanism is of limited effectiveness as an anti-abuse measure but has mistakenly affected a large number of legitimate users.

Sources:

Claude Science: AI Research Workstation for Scientists

Anthropic released Claude Science, a desktop application that integrates research workflows into a unified environment. It comes with over 60 scientific databases and skills covering genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. Each chart includes complete code, runtime environment, and conversation history. Data can be kept on local infrastructure, and heavy computation can call Modal cloud GPUs. Real-world cases include the Allen Institute compressing a two-year review into weeks, and a UCSF team reducing full-process analysis to one-tenth of the original time. Available to Pro ($20/month) and above users, Anthropic also funds up to 50 research projects with a maximum of $30,000 each.

Sources:

Google Releases Nano Banana 2 Lite: Fastest and Cheapest Image Model

Google launched Nano Banana 2 Lite (gemini-3.1-flash-lite-image), generating text-to-image output in under 4 seconds, priced at approximately 0.23 RMB per image. The company recommends that original Nano Banana (gemini-2.5-flash-image) users switch to this version. Also released is Gemini Omni Flash, with new AI Studio and API call support. Multiple bloggers praised its speed advantage and provided selection advice for different Banana series models.

Sources:

Neuralink Achieves Dural Puncture Electrode Implant, Musk Calls It Significant

Neuralink, in a clinical trial, for the first time implanted electrode wires directly through the dura mater into the cerebral cortex while keeping the dura intact. Elon Musk cited the achievement, saying “the problem of electrode implantation through the dura has been solved,” believing this greatly enhances the safety and ease of use of brain-computer interfaces.

Sources:

Lao Wei’s Vibe Coding Progress and WeChat Seconds AI Comic Feature

Wesley demonstrated the UI of an audio/video processing app developed using vibe coding, primarily for compression, splitting, and extraction. Additionally, WeChat Seconds launched an AI comic video feature. He also mentioned offering paid consulting services at 199 RMB per 45 minutes via Google Calendar, and shared daily technical observations ranging from sqlite, WeChat storage usage, to OPC topics.

Sources:

WordPress Releases WPVibe Plugin: Claude Can Take Over Site Management

WordPress released the WPVibe plugin, which connects any AI to a self-hosted WordPress site. The suite includes a cloud MCP server and a small site-side plugin, exposing secure endpoints and enforcing WordPress user permissions. Using a Claude subscription, users can manage articles, upload media, optimize SEO, edit themes, and even install plugins via natural language, with over 40 built-in WP-CLI commands.

Sources:

Statistics: Number of timeline scans=360, Number of bloggers hit=46, Total tweets hit=230, Weighted tweet score=188.75, Original tweets=114, Retweets=39, Crawl attempts=2, Boundary coverage status=tail_confidently_crossed_target_boundary