Last Week in AI: October 14, 2025

OpenAI research & releases

ChatGPT gets apps

OpenAI rolled out a major update to ChatGPT: apps inside the chat window. Instead of sending you off to third‑party sites, these apps live directly inside ChatGPT and can display interactive maps, playlists, slide decks and booking interfaces. Pilot partners include Booking.com, Expedia, Spotify, Figma, Coursera, Zillow and Canva oai_citation:0‡techcrunch.com. Users can call an app by typing its name, ChatGPT will suggest relevant apps, and you can connect your existing accounts to complete tasks. Developers build apps using the Model Context Protocol (MCP), an open standard for tools to speak with models oai_citation:1‡openai.com. Apps are available on Free, Go, Plus and Pro plans outside the EU, and OpenAI promises a directory and monetisation details later oai_citation:2‡openai.com.

Sources: OpenAI Apps & Apps SDK

AgentKit turns anyone into an agent builder

At DevDay, OpenAI unveiled AgentKit, a suite of tools for building AI agents without writing much code. The toolkit features an Agent Builder that lets you chain multiple tool calls on a visual canvas, a Connector Registry to manage data sources across OpenAI products, and ChatKit, which embeds a chat interface directly into your product oai_citation:3‡openai.com oai_citation:4‡openai.com. New evaluation capabilities include datasets and trace graders to test and improve your agent oai_citation:5‡openai.com. There’s even reinforcement fine‑tuning: you can supply custom tool calls and grading functions to train agents in your own environment oai_citation:6‡openai.com. By packaging deployment, evaluation, and fine‑tuning, AgentKit threatens to upend low‑code automation platforms and prompted headlines like “OpenAI just killed n8n and Zapier”.

Sources: OpenAI AgentKit · AI Agents Unleashed Playbook

Codex goes General Availability

OpenAI’s Codex product—an AI agent that delegates tasks across the web—hit general availability. New features include a Slack integration that allows colleagues to request tasks inside chat, a Codex SDK for embedding the agent in your own workflows, and admin tools for managing usage oai_citation:7‡openai.com. Adoption is high: usage has grown 10×, serving over 40 trillion tokens oai_citation:8‡openai.com. Companies like Cisco and Instacart use Codex to triage user feedback and automate order issues oai_citation:9‡openai.com. Codex is available to Free, Go, Plus and Pro users outside the EU oai_citation:10‡openai.com. Sources: GPT‑5 Pro Documentation

Building the future of compute

OpenAI isn’t just writing code; it’s building hardware partnerships. Last week the company announced:

1. A six‑gigawatt GPU deal with AMD. OpenAI will deploy AMD’s MI450 GPUs starting in the second half of 2026. The partnership includes a warrant for up to 160 million AMD shares, and executives expect tens of billions of dollars in revenue oai_citation:11‡openai.com oai_citation:12‡openai.com.
2. A ten‑gigawatt custom accelerator collaboration with Broadcom. OpenAI will design specialized AI accelerators while Broadcom provides Ethernet connectivity. The plan is to begin deploying racks in late 2026 and ramp through 2029 oai_citation:13‡openai.com. Sam Altman says this infrastructure will “unlock the potential of AI,” and Broadcom’s Hock Tan calls it pivotal for the industry oai_citation:14‡openai.com.
3. Stargate Argentina. Reuters reports that OpenAI signed a letter of intent with Sur Energy to build a 500‑megawatt data centre in Argentina, with an estimated $25 billion investment oai_citation:15‡reuters.com. The project would be one of the country’s largest tech–energy initiatives and is the first Stargate site in Latin America oai_citation:16‡reuters.com.

Sources: AI Agents Unleashed Playbook

Sora 2 and Sora 2 Pro

OpenAI’s Sora 2 video and audio model makes physically accurate, multi‑shot videos. It simulates complex physics—basketballs rebound off backboards, characters fall believably—and adds realistic soundscapes oai_citation:17‡openai.com. Sora 2 can follow multi‑shot prompts and perform controlled camera moves oai_citation:18‡openai.com. An experimental Sora 2 Pro model offers even higher quality; ChatGPT Pro subscribers can access it via sora.com oai_citation:19‡openai.com.

Political bias evaluation

OpenAI researchers published a new method for measuring political bias in large language models. They created around 500 prompts spanning five ideological axes and found that models remain nearly objective on neutral queries but show moderate bias on emotionally charged questions oai_citation:20‡openai.com. Using the metric, they show GPT‑5’s “thinking” mode reduces bias by about 30 % oai_citation:21‡openai.com.

Google News

Opal: build AI apps without code

Google expanded Opal, its no‑code AI mini‑app builder, to 15 new countries, including Canada, India, Japan, South Korea, Vietnam, Indonesia, Brazil and several Latin American nations oai_citation:30‡blog.google. Opal lets users design apps using natural language, with advanced debugging, localized error messages and parallel execution to speed up complex workflows oai_citation:31‡blog.google. The update signals Google’s push to make AI app creation accessible worldwide.

Nano Banana image editing

The viral Nano Banana update (also called Gemini 2.5 Flash Image) supercharges Google’s AI image editing. Users can combine two images seamlessly, remove or add elements while preserving the rest, pick colours from an image and apply textures such as flowers onto boots oai_citation:32‡timesofai.com. The tool adds visible and invisible watermarks through SynthID to discourage misuse oai_citation:33‡timesofai.com. Both paid and free users can try Nano Banana inside the Gemini app, which explains why the model topped the LM Arena image‑editing leaderboard oai_citation:34‡timesofai.com.

CodeMender: autonomous code security

Google DeepMind announced CodeMender, an autonomous agent that scans open‑source code for vulnerabilities and rewrites fixes. It uses Gemini models to debug complex bugs, and is both reactive—patching new vulnerabilities as they arise—and proactive, rewriting code to remove entire classes of security flaws oai_citation:35‡vice.com. Over six months, CodeMender upstreamed 72 security fixes into open‑source projects, some as large as 4.5 million lines oai_citation:36‡vice.com. Humans still review patches before release.

Sources: Google DeepMind CodeMender

Other releases

Lumina‑Image 2.0

Shanghai AI Laboratory released Lumina‑Image 2.0, a text‑to‑image generator built on a Unified Next‑DiT architecture that treats text and image tokens as a single sequence. It employs a Unified Captioner (UniCap) to generate high‑quality captions and uses progressive training to improve efficiency oai_citation:37‡arxiv.org. With only 2.6 billion parameters, Lumina‑Image 2.0 matches or beats larger models on several benchmarks oai_citation:38‡arxiv.org.

Lumina‑DiMOO

Lumina‑DiMOO is an omni diffusion foundation model for multimodal generation and understanding. It uses a unified discrete diffusion architecture to handle text, image and audio inputs and outputs; offers versatile capabilities including text‑to‑image, image‑to‑image editing and image understanding; and improves sampling efficiency with a bespoke caching method, resulting in twice the sampling speed compared with earlier models oai_citation:39‡github.com. The model leads on multiple multimodal benchmarks oai_citation:40‡github.com.

Sources: Lumina‑DiMOO

n8n funding and the low‑code race

n8n, a no‑code automation platform, raised $180 million at a $2.5 billion valuation days after AgentKit’s launch oai_citation:41‡analyticsindiamag.com. The timing sparked debate about whether OpenAI’s tool would kill n8n and similar platforms. Users noted that AgentKit still requires technical configuration, while n8n focuses on reliability and data privacy. Vodafone used n8n for threat‑intelligence workflows, saving £2.2 million and thousands of person‑days oai_citation:42‡analyticsindiamag.com. The funding round shows investors still see room for specialised automation platforms, even as OpenAI pushes into the space.

Sources: Paper2Video · Grok Imagine · NVIDIA ChronoEdit

Wrap up

That’s it for this episode of Last Week in AI. From OpenAI’s app ecosystem and agent tools to tiny models and giant data centres, the AI landscape continues to evolve at breakneck speed. Developers have more power than ever, and businesses are racing to harness it. If you found this update useful, share it with your peers and subscribe so you don’t miss next week’s roundup. Until next time, keep building and exploring.