AI Coding Assistants

Today's Stories

1. 1

   Anthropic Tightens Access Controls Against Chinese Engineers

   Anthropic is cracking down on workarounds that Chinese companies have used to access Claude AI, including corporate accounts routed through Singapore entities and personal subscriptions accessed via VPN. The company said it explicitly prohibits access from unsupported regions including China and uses detection systems to identify and ban violating accounts. Chinese engineers value Claude Code—which produces outputs usable for distillation, a technique to train smaller AI models—making unauthorized access a competitive concern for Anthropic. While the workarounds breach Anthropic's terms of service, they do not violate U.S. or Chinese law, leaving Anthropic to rely on its own enforcement.

   Anthropic continues to target 'transfer station' services that relay mainland requests through overseas accounts, though larger Chinese AI groups avoid these due to concerns that operators may store or resell prompts.

2. 2

   Microsoft replaces OpenAI, Anthropic models with cheaper in-house AI in Copilot

   Microsoft is replacing AI models from OpenAI and Anthropic with its own in-house models across Copilot products including Excel and Outlook. The in-house models are already processing tens of thousands of requests per week, though they still handle only a small fraction of total requests. Microsoft also unveiled seven new AI models at Build, including MAI-Thinking 1, its first reasoning model. Copilot and Office customers may end up paying the same price for weaker AI as Microsoft cuts costs. Microsoft's head of AI stated in June the goal is to "reduce and ultimately eliminate" spending on Anthropic. CEO Satya Nadella has also hinted that pricing could shift to usage-based models, potentially making cheaper in-house models the default with third-party models available as premium add-ons—passing costs to customers as a surcharge.

   Microsoft claims its in-house models are trained on clean, commercially licensed data, but the technical paper shows it used Common Crawl, a freely accessible web dataset whose use for AI training is not legally settled. On benchmarks, MAI-Thinking 1 trailed OpenAI and Anthropic models by a wide margin and landed roughly on par with Deepseek V3.2, despite Microsoft's claims it could match Sonnet 4.6 and Opus 4.6 in coding.

3. 3

   Visual Studio Code now shows GitHub Copilot token consumption in real time

   Microsoft has added a feature to Visual Studio Code that displays the number of tokens consumed by GitHub Copilot, along with alerts when usage approaches or reaches a preset spending threshold. GitHub Copilot switched to consumption-based pricing in June 2026, charging users based on tokens used for AI chat and agent features. The new display helps developers avoid unexpected costs by monitoring usage as they work.

   The feature became available on June 17 with Visual Studio Code 1.125.

4. 4

   Anthropic expands Claude Cowork to mobile and web

   Anthropic launched Claude Cowork on mobile and web platforms, beginning with a beta rollout to Max subscribers. The tool now allows work to start on one device, continue running in the background, and be reviewed on another — even after the user closes the app. Cowork was designed to bridge AI coding agents and a much larger market of knowledge workers who do not write code. Usage data from 1.2 million anonymized sessions across more than 600,000 organizations shows the overwhelming majority of Cowork users are not coding, suggesting strong demand beyond developers.

   The rollout begins in beta with Max subscribers before expanding to additional plans. The cross-device capability is a strategic shift: tasks can run autonomously in the background while users move between laptop, phone, and other devices.

5. 5

   AI coding assistant is quietly shipping your secrets

   AI coding assistant is quietly shipping your secrets

6. 6

   Anthropic shares Claude Fable 5 tactics: map your blind spots first

   Anthropic published a usage guide for Claude Fable 5 on July 6, outlining how engineers can improve their work with the AI model by identifying and reducing "unknown elements"—gaps between what they know and what the codebase actually requires. The company's engineer Tariq Shihipar described a four-category framework (known knowns, known unknowns, unknown knowns, unknown unknowns) to surface hidden constraints before and during implementation. When users give vague instructions to AI coding tools, the AI cannot see the real-world constraints embedded in existing code—much like a map cannot capture every geographic obstacle. By clarifying these blind spots upfront, developers can reduce costly revisions later and make more effective use of Fable 5 during both planning and execution phases.

   Shihipar recommends starting each project by asking Claude to help identify what you don't yet know, then using brainstorming, prototyping, and AI interviews with reference materials to fill gaps. He also suggests documenting exceptions, creating post-implementation explainers, and turning AI changes into quizzes to verify understanding of what was actually built.