We anticipated a domino effect of last week’s major AI releases from OpenAI and SpaceXAI, and sure enough, the world of AI didn’t disappoint. It was funny to see how OpenAI wanted to hog headlines all week and how Uncle Dario’s Anthropic dropped another free week for all the Claude Fable lovers, neutralising Brother Sam’s ‘news efforts.’
But that wasn’t all.
The most shocking update came at the end of the week when China’s Moonshot AI released Kimi K3 as an open-weight AI model offering Claude Fable kind of performance at one-fourth the price. This was all while Sam thought that scribbling the ChatGPT logo on a basketball would keep the world mesmerised about OpenAI. If that wasn’t enough, he decided to charge $70 for it!
Not so easy, pal, said Tim Apple, as he managed to jump over massive regulatory hurdles to bring Apple Intelligence to the world’s biggest smartphone market – China. And yes, Perplexity returned to the news headlines too.
A lot of gossip, so let’s round up everything of prime importance from the world of AI this week.
Moonshot AI unleashes Kimi K3 to challenge the giants
Moonshot AI completely upended the open-source community by launching Kimi K3. This absolute monster of a model is a sparse Mixture-of-Experts (MoE) architecture boasting a staggering 2.8 trillion total parameters (activating 16 of 896 experts per token) and a massive 1-million-token context window. It features native vision capabilities and heavy reasoning.
In public evaluations like the Artificial Analysis Coding Index, K3 placed a razor-thin 0.3 points behind Anthropic’s closed Fable 5. It captured the #1 spot on the Arena AI Frontend Code Leaderboard and topped benchmarks for editorial writing and instruction-following.
The model is live now across the Kimi ecosystem (Kimi Work, Kimi Code, and API) for a highly competitive $3 per million input tokens and $15 per million output tokens. Moonshot says it will release the full weights for free under a modified MIT license, making it the largest open-weights model in history.
Did China just overtake the world in generative AI?
Mira Murati has an ‘Inkling’ that says otherwise
While Moonshot was flexing its parameter count, Mira Murati played her first big cards from the side stands. Marking a public debut for the former OpenAI CTO’s highly anticipated startup, Thinking Machine Labs launched the Inkling model family as an open-weight, multimodal architecture packing 975 billion total parameters (41 billion active) and a 1-million-token context window.
Trained natively on custom, powerhouse Nvidia infrastructure following a massive multi-billion-dollar partnership, the model weights have been released under an Apache 2.0 license on Hugging Face. The startup has heavily optimised this release for local enterprise customisation and advanced audio processing, intensifying the pressure on traditional closed-API token monetisation.
OpenAI lures with cool new keyboard, and GPT-5.6 Work
Brother Sam is flexing his corporate muscles again, and this week, OpenAI decided to enter your physical desktop setup. OpenAI partnered with custom hardware manufacturer Work Louder to release the Codex Micro for $230 as the company’s very first branded physical hardware device.
Built specifically for developers utilising the Codex AI coding assistant, it features dedicated “Agent Keys” that glow different colours based on what the AI is doing. The keyboard also includes a physical joystick to navigate AI workflows and a rotary encoder dial to adjust the AI’s “reasoning level” on the fly. Also, there’s the $70 basketball that’s just a basketball with a ChatGPT sticker on it.
On the software front, developers spent the week stress-testing GPT-5.6 and the newly launched ChatGPT Work workspace ecosystem. Built to act as a hyper-capable desktop agent, the workspace allows users to execute complex data manipulation and coding workflows natively on local systems. The superapp ambitions also led to the end of the Atlas web browser.
Meanwhile, Anthropic chose to keep users hooked by extending free access to its top-tier Claude Fable 5 model for premium subscribers, keeping the model price war aggressively alive. Cheeky Uncle Dario!
Tim Apple’s parting gift to China
Tim Apple has finally overcome a massive, year-long regulatory ordeal to bring Apple Intelligence to one of the world’s largest consumer markets. Apple has officially registered its AI suite with China’s Cyberspace Administration. But it wasn’t an easy task.
To get the green light, Apple had to adapt to local compliance laws by striking a major partnership with Chinese tech giants Alibaba and Baidu. Alibaba confirmed that its Qwen AI model will be natively integrated to power text and image generation features across iOS 27, macOS 27 Golden Gate, iPadOS 27, and visionOS for Chinese users.
Globally, Apple officially pushed out the first public beta builds for iOS 27 and macOS 27 Golden Gate, giving regular non-developer users their first global taste of next-generation system-wide multimodal writing assistants and upgraded Siri AI features.
Perplexity’s ‘SPACE’ for autonomous agents
To support its autonomous web-browsing and coding tools, Perplexity AI officially launched a secure, next-gen runtime platform called SPACE (Sandbox Platform for Agentic Computing Environments).
Most existing sandbox environments are designed for quick, disposable tasks. AI agents, however, need to work for hours, but leaving sandboxes open indefinitely exposes sensitive user credentials and data. Perplexity’s SPACE spins up ephemeral sandboxes using AWS’s Firecracker microVMs and takes snapshots of the active memory every single minute.
Sounds complex, but this means you can walk away mid-task, come back a week later, and your AI agent will pick up precisely where it paused.
What else?
Regulations around AI and infrastructure are increasingly witnessing a lot of activity from world governments. This week, there were some major implications:
India’s MeitY halts OpenAI/Anthropic deployments: A department under India’s Ministry of Electronics and Information Technology (MeitY) officially directed all central ministries to freeze deployments of foreign OpenAI and Anthropic models for operational and cybersecurity tasks. The freeze focuses heavily on data sovereignty, explicitly rejecting a proposal from the Ministry of Finance to use OpenAI’s upcoming GPT-5.5 architecture for automated vulnerability discovery.
New York blocks data centres: In a major regulatory pushback against the structural demands of AI, New York State has moved to halt the construction of new large-scale data centres for up to one year. The freeze is intended to prevent power-hungry AI training workloads from completely overwhelming the state’s energy grid.
Satya Nadella’s AI warning: Microsoft’s Satya Nadella warned enterprises that they risk “paying for intelligence twice” — first financially for the software, and second by losing their unique institutional knowledge to foundation model providers through user prompts and agent traces. He urged organisations to build strict private trust boundaries to protect corporate IP.