GPT-5.6 will see a wider release this week: China’s government moves to control its companies’ AI products

In today’s edition of Data Points, you’ll learn about our top headlines, and more:

• Copilot will decrease reliance on GPT, Claude
• Claude Cowork gets web, mobile implementations
• Nvidia’s Audex blends audio and voice with text intelligence
• OpenAI mini voice model promises to keep costs low

But first:

GPT-5.6 will arrive (at least via the API) on Thursday

OpenAI will release its GPT-5.6 models on Thursday after a government-imposed delay that kept them restricted to select partners for weeks. The Department of Commerce approved the public launch following additional safety testing by the Center for AI Standards and Innovation, though OpenAI criticized the hold as keeping advanced tools away from developers. On TerminalBench 2.1, OpenAI’s GPT-5.6 Sol Ultra scored 91.9 percent versus Anthropic’s Claude Mythos 5 at 88 percent, while matching Mythos 5 on cybersecurity tasks while using only a third of the tokens. Pricing for GPT-5.6 varies by tier: Sol comes in at $5 per million input tokens and $30 per million output tokens, while Terra ($2.50/$15) and Luna ($1/$6) offer cheaper options—making the overall family roughly half of Anthropic’s rates at the comparable tier. The rollout highlights a broader tension: Trump’s latest AI executive order calls for binding safety standards before releasing frontier models, but no such standards formally exist yet. (The Decoder)

China may restrict global access to top AI models

Chinese authorities began talks with major tech companies about limiting foreign access to the country’s most advanced AI models, including systems still in development. Companies consulted include Alibaba, Bytedance, and Z.ai. The discussions, held over the past month, suggest Beijing is considering tighter controls over which AI capabilities leave the country, a move that could reshape how Chinese AI research reaches global markets and developers. The proposed scope covers both released and unreleased models, suggesting restrictions could apply even before products launch. This would negate some of the advantages Chinese companies have in the AI market, where open source and low costs have helped spur worldwide adoption. (Reuters)

Microsoft will curb use of third-party models in its own applications

Microsoft began replacing OpenAI and Anthropic models with its own internally built MAI models in widely used applications like Excel and Outlook, processing tens of thousands of AI prompts weekly. The shift reflects a broader effort to reduce dependency on expensive third-party AI services as the company’s discount partnership with OpenAI approaches expiration. At Microsoft Build in June, AI chief Mustafa Suleyman announced seven new models and explicitly stated the goal to “reduce and ultimately eliminate” spending on Anthropic, which the company acknowledged it had been paying “a lot of money” to use. Microsoft’s MAI models are already available in GitHub Copilot and will expand to Teams for transcription in coming months. While the internal models still account for a small portion of Microsoft’s total AI usage, the strategy signals the company is building viable alternatives to remain independent of pricing decisions by leading AI labs. (Bloomberg)

Claude Cowork moves to mobile devices, backed by the cloud

Anthropic expanded Claude Cowork beyond the desktop, launching web and mobile versions in beta for Max plan subscribers. The move lets users start tasks on one device, monitor progress from another, and let Cowork run autonomously in the cloud—useful for the “work around the work,” as Anthropic calls it: email sifting, spreadsheet consolidation, and other administrative tasks that eat up professional time but aren’t core job duties. Desktop remains the full experience with local file and browser access, but cloud-based tasks can now run on schedule without a machine powered on. The company analyzed 1.2 million Cowork sessions and found a notable finding: over 90 percent of usage had nothing to do with software development. Instead, 50 percent fell into business operations and content creation—33.4 percent on business process and operations alone, 16.4 percent on copywriting. The data suggests Cowork’s real value lies in helping professionals organize and structure information. (ZDNET)

Nvidia avoids the “text tax” on its latest voice/audio model

Nvidia released Audex, a 30-billion-parameter mixture-of-experts model that handles both audio and text input and output while maintaining the text reasoning capabilities of its backbone. Most multimodal models suffer a “text tax,” such that adding audio or vision capabilities degrades performance on text benchmarks, but Audex avoids this through a straightforward architectural design: audio inputs are encoded into the text embedding space and audio outputs are treated as quantized tokens, allowing the model to use standard LLM infrastructure like Megatron-LM and vLLM. The training pipeline uses multi-stage supervised fine-tuning followed by text-only reinforcement learning, which keeps text scores matched to the backbone while adding audio capabilities. On reasoning benchmarks like HMMT and IMO AnswerBench, Audex outperforms the comparably sized Qwen3-Omni model, and it leads comparable open audio-LLMs models on speech recognition with a 6.82 word error rate. The model also generates general audio beyond speech—a capability most competing open models lack—though it trails some specialized audio LLMs on certain audio understanding tasks. The weights ship under a noncommercial license. (MarkTechPost)

Smaller OpenAI voice model keeps the reasoning, cuts the price tag

OpenAI shipped gpt-realtime-2.1-mini, a cheaper voice model that reasons about multi-step problems while keeping prices low. The move fills a gap: Developers can now trade some raw capability for cost control while keeping the logic intact. Both new Realtime models (the standard gpt-realtime-2.1 and the mini tier) process audio and text over a live connection using a single-model architecture. This skips the usual speech-to-text-to-speech pipeline, cutting latency and preserving vocal nuance instead of degrading audio through multiple conversions. OpenAI also squeezed p95 latency down by at least 25 percent across all Realtime voice models through improved caching. Cached audio input on gpt-realtime-2.1-mini now costs $0.30 per million tokens versus $10 for fresh input, a 33x difference. (The larger gpt-realtime-2 model carries a steeper 80x gap: $0.40 cached versus $32 for fresh input.) This cost savings compounds during long sessions with repeated context. Developers can configure reasoning effort (low, medium, high), which lets the agent narrate what it’s doing before hitting a database. (MarkTechPost)

Want to know more about what matters in AI right now? 

Read the latest issue of The Batch for in-depth analysis of news and research. 

Last week, Andrew talked about the challenges of navigating the noisy AI landscape, our “learners first” philosophy at DeepLearning.AI, and why we partner with experts to provide high-quality, technically deep courses. 

“When we are exploring what to teach, we explore a wide range of topics, gathering from a network of technical experts in companies and academia, senior technology and business leaders, writings by technical experts, our team’s practical experience, and objective metrics of different tools’ traction and/or performance.”

Read Andrew’s letter here.

Other top AI news and research stories covered in depth:

• OpenAI previewed three GPT-5.6 Models (Sol, Terra, and Luna), with a wider release coming soon.
• Sakana debuted dedicated orchestrator models, Fugu and Fugu-Ultra, that spawn Claude, Gemini, and GPT agents task by task.
• Microsoft revealed MAI-Thinking-1, a Claude Sonnet 4.6-sized reasoning model developed without distillation.
• Inside RoboReward, a family of vision-language reward models that train robots to take action.

A special offer for our community

In case you missed it, DeepLearning.AI launched our first-ever subscription plan for our entire course catalog! As a Pro Member, you’ll immediately enjoy access to:

• Nearly 200 AI short and long courses from Andrew Ng and industry experts
• Labs and quizzes to test your knowledge
• Projects to share with employers
• Certificates to testify to your new skills
• A community to help you advance at the speed of AI

Enroll now to lock in a year of full access for $25 per month paid upfront, or opt for month-to-month payments at just $30 per month. Both payment options begin with a one-week free trial. Explore Pro’s benefits and start building today!

Try Pro Now

Data Points is produced by human editors with AI assistance.