GitHub previews Copilot SDK as Alibaba releases Qwen3-TTS

GitHub introduced the Copilot SDK as a technical preview on January 22, 2026. Developers can use it to embed the agent loop behind GitHub Copilot CLI in their own applications. The SDK provides an execution loop that lets an agent plan tasks, call tools, edit files, and run commands. Developers do not have to build that loop from scratch.¹

The SDK is intended for teams that want agent capabilities inside their software rather than only through Copilot CLI. GitHub also lists support for custom tools and MCP integration.¹² Its technical-preview status is a caveat for companies considering deployment: teams should assess whether it meets their requirements before using it in production.¹

Alibaba released Qwen3-TTS, a family of text-to-speech models, on January 22, 2026.¹ The publicly available models support Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian.² Alongside preset voices, the family can clone a voice from a recording or create a new one from a description.² It also supports streaming speech generation, allowing output to begin before synthesis is complete.³

Qwen3-TTS brings several speech capabilities together in an open model family rather than requiring separate tools.⁴² That could make it easier for companies to build their own voice assistants and multilingual applications. The published technical information does not, however, establish a fixed output latency across hardware and use cases.³

Anthropic published an updated “constitution” for Claude on January 21, 2026.¹ The document describes the behavior the company wants from Claude and the values intended to guide its responses.¹ It also explains the reasoning behind its directions rather than simply listing rules.¹² The constitution is publicly available, and Anthropic says it uses the document directly in training its models.¹³

The stated order of priorities puts safety and ethics ahead of compliance with company guidelines and helpfulness.¹ This develops Anthropic’s Constitutional AI approach, in which written principles help shape model behavior.¹² For businesses, the document provides a reference point for comparing their expectations of Claude with its developer’s stated goals. It does not, however, establish how reliably the model follows those priorities in specific tasks or edge cases.¹

Google DeepMind is bringing in Hume AI co-founder and CEO Alan Cowen and several engineers to work on Gemini’s voice features. On January 22, Hume AI said it had agreed to license certain technologies to Google on a non-exclusive basis and that Cowen would move to Google. The arrangement is a talent and licensing deal, not an announced acquisition of the company.¹²

Hume AI develops voice models and conversational interfaces designed to respond to emotional cues. The agreement could help Google improve spoken interactions with Gemini. Hume AI will remain independent under new leadership and can continue supplying its technology to other AI companies. Neither a launch date nor the form of any resulting Gemini voice upgrade has been announced.¹³²

World Labs introduced the public World API on January 21, 2026. Developers can use its Marble model to generate explorable 3D worlds from text, images, panoramas, multiple views, or video. Outputs can be rendered on the web, exported to other tools, or integrated into interactive systems and simulations. The API is available through the World Labs Platform.¹

The launch gives developers a way to incorporate generated environments into their own applications.¹ Separately, Bloomberg reported that World Labs was discussing a financing round worth hundreds of millions of dollars at a possible valuation of about $5 billion. That figure is not a confirmed valuation from a completed round.² The product announcement does not provide comparative performance results.¹

Google DeepMind introduced D4RT on January 22 as a research model for reconstructing and tracking dynamic scenes from video. It combines spatial information and changes over time in a 4D representation.¹ The model encodes video into a shared scene representation, then uses it to answer queries about specific locations and times.¹

Earlier approaches often treated depth estimation, camera pose, and point tracking as separate tasks. D4RT brings them into one model, which could simplify the analysis of moving scenes.¹ For robotics and augmented reality, the practical question is whether spatial relationships and motion can be captured reliably from video.² The announcement presents D4RT as research and does not specify a product launch, pricing, or public access.¹

Anthropic has replaced Todos in Claude Code with Tasks for longer coding projects.¹² Tasks are stored locally and remain available after a session ends.¹³ That lets an agent resume its work later; multiple sessions or subagents can also work with a task list.¹² The change is part of Claude Code rather than a standalone service.¹

Tasks can represent dependencies and blocked steps.²³ That matters for projects where several agents handle related pieces of work and need to coordinate their progress.¹² Local storage also has a limit: it does not, by itself, mean task lists automatically sync between machines.¹² Teams therefore need to decide how to make task lists available and review progress during longer agent runs. Task tracking can support coordination, but the resulting code still needs review.