Legacy Code

Anthropic researchers have developed a technique that maps an internal "reasoning space" within Claude, revealing how the model thinks through problems before generating responses. This hidden layer—distinct from the final output—shows Claude actually puzzles over concepts, constraints, and trade-offs in ways not visible to users.

Anthropic is moving Claude subscribers to usage-based pricing, requiring consumers to pay per token consumed when accessing Claude Fable 5 (the company's most capable model). While lower-tier models remain part of the base subscription, access to Anthropic's best consumer model will incur incremental charges based on actual usage.

Estonia's government has deployed an AI system to identify wording errors and logical inconsistencies in legislation before it becomes law. The program was inspired by a single typo that cost the government approximately $28 million in unanticipated expenses. The AI system scans legislative documents for ambiguous phrasing, cross-referencing errors, and unintended legal consequences.

Google has expanded its AI content labeling policy to require advertisers to disclose whether ads are created or edited with generative AI tools. The new requirement applies to ads on Google Search, Google Discover, and YouTube, visible through Google's My Ad Center under a "created or edited with AI" label.

OpenAI and Microsoft have deployed GPT-5.6 as the primary model powering Microsoft 365 Copilot, the AI assistant embedded in Word, Excel, PowerPoint, and Teams. The upgrade improves performance across document generation, spreadsheet analysis, presentation design, and meeting summarization.

Meta has launched Muse Spark 1.1, an AI coding assistant designed specifically for enterprise-scale tasks like bug fixing, code refactoring, and managing large system migrations. The tool targets the growing demand from engineering teams seeking AI that goes beyond single-function code completion to handle complex, multi-file transformations.

Microsoft announced that it is using AI to identify potential security vulnerabilities earlier in its development cycle, which will result in a higher volume of security updates in each monthly Patch Tuesday release. The company's AI systems scan code and systems to surface issues that might otherwise be missed or discovered later.

The New York Times has filed a motion alleging that OpenAI withheld tools and datasets that could identify copyrighted journalism in ChatGPT outputs, escalating their ongoing copyright lawsuit. The Times claims OpenAI hid evidence related to how the model reproduces or references journalistic content.

OpenAI has announced the discontinuation of ChatGPT Atlas, its experimental browser application that could autonomously execute tasks and navigate websites on behalf of users. The tool was introduced in October 2025 but is being "sunset" as part of a restructuring around ChatGPT Work.

OpenAI has released ChatGPT Work, a new agent interface that moves beyond conversation to actively execute tasks on behalf of users. The agent can take independent action across connected applications and files, maintain context on long-running projects over multiple hours, and deliver completed work rather than just suggestions or drafts.