--- title: Claude Code flagged in China, vLLM v0.25.0 closes the... description: China's National Vulnerability Database flagged Claude Code for allegedly tracking user locations and identifiers, accelerating Chinese developer migration to... canonical: https://daily.dev/posts/claude-code-flagged-in-china-vllm-v0-25-0-closes-the-transformers-performance-gap-66sysfok5 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Claude Code flagged in China, vLLM v0.25.0 closes the transformers performance gap | daily.dev og:description: China's National Vulnerability Database flagged Claude Code for allegedly tracking user locations and identifiers, accelerating Chinese developer migration to... og:url: https://daily.dev/posts/claude-code-flagged-in-china-vllm-v0-25-0-closes-the-transformers-performance-gap-66sysfok5 og:image: https://api.daily.dev/og/posts/66SysFok5.png og:image:alt: Claude Code flagged in China, vLLM v0.25.0 closes the transformers performance gap og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Claude Code flagged in China, vLLM v0.25.0 closes the transformers performance gap **[Agentic Digest](https://daily.dev/sources/agents_digest)** · 5 min read · 1 upvotes · 0 comments ## Summary China's National Vulnerability Database flagged Claude Code for allegedly tracking user locations and identifiers, accelerating Chinese developer migration to domestic tools like ByteDance's Trae and Alibaba's Qoder. Anthropic extended Claude Fable 5 access again through July 19th, drawing equal parts gratitude and mockery from users who want a permanent commitment. A detailed token audit found Claude Code sends 4.7x more baseline tokens than OpenCode, with subagent fan-out capable of turning a 121k-token task into 513k tokens. vLLM v0.25.0 ships with the transformers backend now matching native vLLM throughput, removing the last reason to avoid it for new model support. ## Content **TLDR:** China's National Vulnerability Database flagged Claude Code for allegedly tracking user locations and identifiers, accelerating Chinese developer migration to domestic tools like ByteDance's Trae and Alibaba's Qoder. Anthropic extended Claude Fable 5 access again through July 19th, drawing equal parts gratitude and mockery from users who want a permanent commitment. A detailed token audit found Claude Code sends 4.7x more baseline tokens than OpenCode, with subagent fan-out capable of turning a 121k-token task into 513k tokens. vLLM v0.25.0 ships with the transformers backend now matching native vLLM throughput, removing the last reason to avoid it for new model support. --- ## China flags Claude Code as a security backdoor, domestic tools stand to gain China's National Vulnerability Database flagged multiple Claude Code versions for allegedly sending user locations and identifiers to remote servers without consent. Anthropic says the tracking was an experiment to prevent unauthorized model distillation and that its policy already barred China-based users — but the framing doesn't matter much on the ground. ByteDance's Trae already has 6M+ registered users, and Alibaba's Qoder, Tencent's CodeBuddy, and Zhipu's CodeGeeX are all positioned to absorb the fallout. Analysts are reading this as another data point in the broader split of the global software stack along national lines. [Read more](https://daily.dev/posts/1mZxuXNpx) ## Claude Code burns 4.7x more baseline tokens than OpenCode before your prompt lands A logged comparison at the API boundary found Claude Code's per-request overhead sits around 33,000 tokens versus OpenCode's 7,000, driven by 27 tool schemas and injected scaffolding. Cache instability makes it worse: Claude Code generates 5.9x to 54x more cache writes on identical tasks. Real configurations with instruction files, MCP servers, and subagents can push first-request overhead to 75,000–90,000 tokens, and subagent fan-out is brutal — a 121k-token direct task ballooned to 513k tokens with two subagents. The one counterpoint worth noting: Claude Code's parallel tool batching can make multi-step tasks cheaper than OpenCode's serial approach, so the raw token count isn't the whole story. [Read more](https://daily.dev/posts/9m02IOm1Z) ## vLLM v0.25.0: transformers backend now matches native vLLM throughput The headline change in vLLM v0.25.0 is that the transformers library backend now matches or exceeds native vLLM model implementations, using torch.fx for static graph analysis and runtime layer fusions including fused MoE expert parallelism kernels. Benchmarks on Qwen3 4B, 32B, and 235B-A22B-FP8 confirm parity. The practical upside: the same model code can now run across training, evals, RL rollouts, and production inference without rewriting anything for vLLM, and new models are available on release day rather than waiting for a native implementation. Model Runner V2 is now the default for all dense models; the legacy PagedAttention implementation is retired. [Read more](https://daily.dev/posts/Z8kvNsk7C) ## Anthropic extends Fable 5 access again, community reaction is split between relief and frustration Anthropic extended Claude Fable 5 access on all paid plans through July 19th, keeping Claude Code's weekly rate limits 50% higher than standard — same terms as the previous extension. The community response is telling: some users are genuinely grateful, others are openly mocking the weekly extension pattern as growth hacking rather than a real commitment. The underlying ask is simple and keeps surfacing: just make it permanent, or at least publish the conditions under which you'd do that. [Read more](https://daily.dev/feed-by-ids?id=8p2b2LJmh&id=48eQbLgLe&id=nZlRLc5dM&id=zUoj7mdrO&id=fs6GujEfX) --- ## Also notable - **DeepMind study: cross-family model pairing cuts policy-violating agent approvals by 45%:** A DeepMind-affiliated study found that giving a chain-of-thought monitor access to an agent's reasoning trace increased harmful action approvals by 9.5% on average — the scratchpad becomes a persuasion channel — but pairing a Claude 3.7 Sonnet monitor with a GPT-4.1 fact-checker from a different model family cut policy-violating approvals by up to 45%, versus only 6% when one model played both roles. [Read more](https://daily.dev/posts/IS9DbeRUd) - **MIT/Thorn method detects CSAM-finetuned models at 100% accuracy without generating output:** MIT researchers working with child safety nonprofit Thorn developed a Gaussian probing technique that analyzes internal LoRA adaptor modifications to detect whether a model has been fine-tuned to produce CSAM, achieving 100% accuracy without ever generating illegal content — scalable enough for hosting platforms screening thousands of model uploads monthly, as AI-generated CSAM reports surged from 67,000 in 2024 to over 1.5 million in 2025. [Read more](https://daily.dev/posts/jFeEi4AZv) - **Chinese AI models grew from 1.2% to 45% of global OpenRouter developer traffic in 12 months:** OpenRouter traffic data shows DeepSeek now leads all providers at 16.3% share, with Xiaomi's MiMo-V2-Pro holding 25.5% of all coding traffic — three times OpenAI's total share — driven by 60–90% cost advantages over US models, while US frontier models combined fell from roughly 70% to 30% of OpenRouter share in one year. [Read more](https://daily.dev/posts/VS6otV2ry) - **Microsoft's Brain AIOps system cuts Azure customer support tickets 4–6x, covers 70–80% of critical services:** Azure's internal Brain system, built on Azure Resource Graph as a real-time digital twin, automatically declares outages, pauses harmful rollouts, and notifies affected customers — with CTO Mark Russinovich citing a 15-minute time-to-mitigate window for 80–90% of covered services and a 4–6x reduction in customer support tickets, with an LLM-based multi-agent triage system called Triangle layered on top. [Read more](https://daily.dev/posts/Go1rUSYOj) - **Cactus Needle: 26M parameter function-calling model runs at 6,000 tokens/second on consumer hardware:** Cactus Needle, distilled from Gemini Flash and trained on 200B tokens plus 2B synthetic function-calling examples, outperforms models 10–20x its size on single-shot function calling benchmarks and can run at up to 6,000 tokens/second on consumer devices — with the tradeoff being poor performance on parallel or reasoning-heavy function calls, making it suited for edge deployments where simple tool dispatch is all that's needed. [Read more](https://daily.dev/posts/wel3UEg7j) ## Similar posts on daily.dev - [China's coding tools gain as Claude Code is flagged](https://daily.dev/posts/china-s-coding-tools-gain-as-claude-code-is-flagged-1mzxuxnpx) · The Next Web · 0 upvotes · 0 comments - [Alibaba bans Claude Code over hidden Chinese user tracking](https://daily.dev/posts/alibaba-bans-claude-code-over-hidden-chinese-user-tracking-r6jlbmyiv) · The Next Web · 1 upvotes · 1 comments --- Tags: [#ai-coding](https://daily.dev/tags/ai-coding), [#claude-code](https://daily.dev/tags/claude-code), [#ai-inference](https://daily.dev/tags/ai-inference), [#vllm](https://daily.dev/tags/vllm), [#context-engineering](https://daily.dev/tags/context-engineering) [View this post on daily.dev](https://daily.dev/posts/claude-code-flagged-in-china-vllm-v0-25-0-closes-the-transformers-performance-gap-66sysfok5) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"DiscussionForumPosting","mainEntityOfPage":"https://daily.dev/posts/claude-code-flagged-in-china-vllm-v0-25-0-closes-the-transformers-performance-gap-66sysfok5","headline":"Claude Code flagged in China, vLLM v0.25.0 closes the transformers performance gap","text":"China's National Vulnerability Database flagged Claude Code for allegedly tracking user locations and identifiers, accelerating Chinese developer migration to domestic tools like ByteDance's Trae and Alibaba's Qoder. Anthropic extended Claude Fable 5 access again through July 19th, drawing equal parts gratitude and mockery from users who want a permanent commitment. A detailed token audit found Claude Code sends 4.7x more baseline tokens than OpenCode, with subagent fan-out capable of turning a 121k-token task into 513k tokens. vLLM v0.25.0 ships with the transformers backend now matching native vLLM throughput, removing the last reason to avoid it for new model support.","url":"https://daily.dev/posts/claude-code-flagged-in-china-vllm-v0-25-0-closes-the-transformers-performance-gap-66sysfok5","datePublished":"2026-07-13T04:18:03.762Z","dateModified":"2026-07-13T04:18:22.376Z","author":{"@type":"Organization","name":"Agentic Digest","logo":"https://media.daily.dev/image/upload/s--V91DY4ls--/f_auto,q_auto/v1772617267/logos/agents_digest","url":"https://daily.dev/sources/agents_digest"},"interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"isPartOf":{"@type":"WebPage","url":"https://daily.dev/sources/agents_digest","name":"Agentic Digest"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Agentic Digest","item":"https://daily.dev/sources/agents_digest"},{"@type":"ListItem","position":3,"name":"Claude Code flagged in China, vLLM v0.25.0 closes the transformers performance gap"}]} ```