DeepSeek is building its own AI chips, and OpenAI gives ChatGPT a real-time voice

The week on the beat

1. DeepSeek is designing its own AI chip

The week's most-covered story by a wide margin (41 outlets). Facing tighter US export controls, the Chinese lab is reportedly building custom silicon to run its models, aiming to cut its dependence on foreign hardware and its training costs. For builders, this is the long game behind why Chinese models keep undercutting Western ones on price.

Read the story →

2. OpenAI ships GPT-Live: ChatGPT can now listen and talk at once

New voice models cut the latency and let the assistant hold a genuinely real-time conversation instead of the old walkie-talkie turns. If you build anything voice-facing, this resets the latency bar.

Read the story →

3. Anthropic puts Claude Cowork on mobile and web

The agent now runs tasks in the background after you close your laptop and pings your phone when it needs a decision. Claude Pro subscribers get it first, with a wider rollout coming. This is the "agent keeps working while you walk away" pattern getting real.

Read the story →

4. Mistral enters robotics with Robostral Navigate

An 8B model that steers robots through unfamiliar spaces using a single ordinary camera, trained in simulation. It is Mistral's first step into physical AI, aimed at industrial partners in Europe.

Read the story →

5. China forces its biggest platforms to pull AI companions

ByteDance and Alibaba removed humanlike chatbot personas ahead of new Beijing rules tightening how AI can imitate people. A reminder that the regulatory ground under Chinese models can shift fast, which matters if one of them is in your stack.

Read the story →

Model moves

Six models landed in the tracker this week. In rough order of who they are for:

• New: Grok 4.5 (xAI). xAI's new flagship, tuned for coding, STEM, and knowledge work, at $2.00/$6.00 per 1M and a 500K context. Model page →
• New: Hy3 (Tencent). A 295B mixture-of-experts model (21B active) built for reasoning and agentic production work, priced to undercut most Western models at $0.14/$0.58 with 262K context. Model page →
• New: Laguna XS 2.1 (Poolside). A small, dedicated coding-agent model cheap enough to run in a loop, $0.06/$0.12, 262K context. Model page →
• New: Nex-N2-Mini (Nex AGI). An open-weight agentic mixture-of-experts model and the cheapest thing to ship all week, $0.025/$0.10, 262K context. Model page →
• New: Aion-3.0 (AionLabs). A roleplay and storytelling system built on the GLM family, aimed at character apps rather than dev work, $3.00/$6.00, 131K context. Model page →
• New: Aion-3.0 Mini (AionLabs). The smaller, cheaper Aion, built on DeepSeek models instead, $0.70/$1.40, 131K context. Model page →
• Price watch: GLM-5.2 climbed to $0.90/$3.08 per 1M (in/out), up from roughly $0.57/$1.80 at the start of the week, after its OpenRouter providers repriced several times. Cheap, but not yet stable enough to bank on without a fallback. Model page →

Personal take

The thread running through this whole week is China. DeepSeek is building its own chips, Beijing is forcing companion bots offline, and the cheap Chinese models everyone is eyeing on OpenRouter kept lurching around on price, GLM-5.2 nearly doubled its output cost in seven days without a single announcement. The cost advantage is real and it is why these models are worth watching, but "cheap" is not the same as "stable," and a price that moves 50 percent in a week with regulators in the mix is not something you build a margin on without a fallback wired in. Meanwhile the US labs spent the week on ergonomics: OpenAI on talking faster, Anthropic on working while you are away from the keyboard. Pick your models for where they will be in six months, not just today's price per token.

Until next Thursday, Anmol