Get tomorrow's brief in your inbox
Today: OpenAI released GPT-5.4 mini and nano, making AI dramatically cheaper for high-volume tasks. Chinese AI companies surged after Nvidia's CEO called OpenClaw "the next ChatGPT." Google expanded Personal Intelligence to all US users, letting Gemini read your Gmail and photos to personalize responses.
OpenAI just made AI significantly cheaper with GPT-5.4 mini and nano, two new models optimized for speed and cost efficiency. GPT-5.4 mini runs twice as fast as the previous GPT-5 mini while improving performance across coding, reasoning, and tool use. GPT-5.4 nano is the smallest and cheapest option, priced at just $0.025 per million input tokens (about $0.0000025 per 100 tokens). That means you could generate descriptions for 76,000 photos for around $52. The new nano model even outperforms the previous GPT-5 mini at maximum reasoning effort, according to OpenAI's benchmarks.
For small businesses, this pricing shift matters. High-volume automation tasks like customer data extraction, document classification, invoice processing, and content moderation just got 60-80% cheaper. If you've been avoiding AI tools because of API costs, these models change the math. Tasks that cost $200/month to automate might now run for $40.
Both models are reasoning models, meaning they can think through multi-step problems instead of just pattern-matching. They work best for repeatable tasks where you need consistent logic at scale: tagging support tickets, extracting structured data from emails, generating product descriptions, or running automated QA checks on code. The speed boost means your automations complete faster, which matters when you're processing hundreds or thousands of requests per day.
If you're already using GPT-4 for high-volume tasks, switch to nano or mini and run a cost comparison for one week. You'll likely save 70% or more while getting better performance. (OpenAI announcement)
Zapier adds GPT-5.4 mini/nano and Opus 4.6. Zapier now supports OpenAI's new GPT-5.4 mini and nano models, plus Anthropic's Claude Opus 4.6. Opus 4.6 has a 1-million-token context window (in beta), which means it can process entire codebases or long document sets in a single pass. It's particularly strong at matching records across different systems, like recognizing "Vineet D." in your ticketing system and "Vineet Das" in your CRM are the same person. When data is ambiguous, Opus 4.6 stops and asks for clarification instead of guessing, making it a good fit for high-stakes workflows like financial reconciliation or compliance checks. If you're running multi-step Zaps that need to coordinate across Gmail, Salesforce, and Google Calendar, Opus 4.6 can handle those without constantly losing context. (Zapier blog)
Alibaba launches Wukong, an agentic AI platform for enterprises. Alibaba's new Wukong platform (named after the Monkey King) lets businesses manage multiple AI agents through a single interface. It handles tasks like document editing, approvals, meeting transcription, and research. Wukong is available as a desktop app or through DingTalk (Alibaba's Slack alternative), and the company plans integrations with Slack, Microsoft Teams, and WeChat. It's currently in invite-only testing. If you're already using DingTalk and have 20+ employees doing repetitive document or approval workflows, request access. For US-based businesses, wait until Slack/Teams integration ships before evaluating. (CNBC report)
Manus (Meta-owned) launches desktop app with local file access. Manus, the AI agent startup Meta acquired last year, released a desktop app that lets its agent work directly with files and applications on your laptop. The "My Computer" feature can read, analyze, and edit local files, organize thousands of images, and create apps within minutes (according to the company). Unlike free OpenClaw, Manus is a paid subscription service. The desktop app requires explicit approval before executing tasks, with "Allow Once" or "Always Allow" options. If you're evaluating AI agents that need local file access, compare Manus to OpenClaw and Claude Code before committing to a subscription. Pricing wasn't disclosed in the announcement. (CNBC report)
Gamma Imagine: AI-generated marketing assets and charts. Gamma, the AI-powered presentation platform, launched Gamma Imagine for creating brand-specific marketing assets. You can use text prompts to generate interactive charts, social graphics, infographics, and marketing collateral. The platform has 100+ templates and integrates with ChatGPT, Claude, Make, Zapier, and Superhuman Go. Gamma now has 100 million users (up from 70 million in November 2025) and $100 million in ARR. If you're tired of waiting on design resources for internal decks or social posts, Gamma sits between full design tools (Adobe, Figma) and legacy presentation software (PowerPoint). Free tier available; paid plans start at $8/month. (TechCrunch coverage)
Google expands Personal Intelligence to all US users. Google's Personal Intelligence feature, which connects Gemini to your Gmail, Google Photos, and other Google apps, is now available to free-tier users in the US (previously paid-only). With it enabled, Gemini can suggest shopping options based on your purchase history, troubleshoot devices using receipts from your inbox, or build travel itineraries based on hotel bookings in Gmail and past trip photos. It's opt-in and you can disconnect apps anytime. Google says it doesn't train directly on your Gmail or Photos library, but it does train on your prompts and Gemini's responses. If you're comfortable with Google reading your email to make AI responses more useful, turn it on. If you're handling client data or confidential business information, leave it off. Personal Google accounts only (not Workspace). (Google blog)
World (Sam Altman's company) launches AgentKit for AI shopping agents. World released AgentKit, a verification tool for websites that lets them confirm a real human is behind an AI agent's purchasing decisions. It uses World ID (derived from an iris scan via World's Orb device) and integrates with the x402 payment protocol developed by Coinbase and Cloudflare. Consumers register their AI agents with their World ID, which then communicates to websites that a verified human approves the purchases. It's currently in beta. If you run an e-commerce site and want to block bot purchases while allowing legitimate AI agents, this is worth tracking. But adoption depends on how many consumers actually get Orb scans (limited availability). (TechCrunch report)
Y Combinator CEO Garry Tan shares his Claude Code setup (gstack). Garry Tan released his Claude Code skill setup on GitHub under an open source license, calling it "gstack." It includes 13 reusable skills (prompts stored in .md files) that instruct Claude to act as a CEO evaluating ideas, an engineer writing features, a code reviewer checking for bugs, a designer, a documentation writer, and more. Tan describes using multiple skills in sequence: get Claude's opinion as CEO, write the feature as an engineer, review the code for security issues, add design, etc. The repo has 20,000 GitHub stars and went viral on Product Hunt. But it also drew criticism after Tan tweeted that a CTO friend said gstack instantly found a security flaw in his company's code. Critics questioned whether the CTO should have known about the flaw already. Tan said he's "sleeping four hours a night" because he's so excited working with AI agents. If you're using Claude Code regularly, Tan's skills are worth reviewing as examples, but customize them for your workflow instead of copying verbatim. (TechCrunch coverage)
Mistral launches Forge: build custom AI models on your own data. Mistral, the French AI startup, announced Mistral Forge at Nvidia GTC. Forge lets enterprises train custom models from scratch using their own data, rather than fine-tuning existing models or using retrieval-augmented generation (RAG). Training from scratch gives companies more control over model behavior, better handling of non-English or domain-specific data, and reduces dependence on third-party model providers. Forge includes tooling for synthetic data pipelines and evaluation, plus forward-deployed engineers who embed with customers. Mistral is on track to surpass $1 billion in ARR this year. If your business has years of internal documentation, workflows, and institutional knowledge that generic models don't understand, Forge is worth exploring. But expect a significant upfront investment in time and engineering resources. (TechCrunch report)
GPT-5.4 nano pricing: $0.025 per million input tokens, $0.10 per million output tokens. That's cheaper than Google's Gemini 3.1 Flash-Lite and dramatically cheaper than previous GPT models. For reference, describing a single image costs about 2,751 input tokens and 112 output tokens, totaling 0.069 cents (less than one-tenth of a cent). At that rate, processing 76,000 images would cost $52.44. If you're running high-volume image analysis, document processing, or data extraction workflows, nano is now the cheapest frontier model option. Switch your existing workflows and measure the cost difference. (Simon Willison's blog)
Google Personal Intelligence now free in the US. Previously available only to paid Google AI Pro and AI Ultra subscribers, Personal Intelligence is now free for all US users with personal Google accounts. If you've been paying $20/month for Google AI Pro just to get personalized Gemini responses, you can now downgrade and keep the feature. But remember: free means Google is training on your prompts and responses to improve the feature. (TechCrunch coverage)
Chinese AI stocks surge after Nvidia CEO touts OpenClaw. Nvidia CEO Jensen Huang called OpenClaw "definitely the next ChatGPT" during an interview on CNBC's Mad Money. Chinese AI companies MiniMax and Zhipu (Knowledge Atlas Technology) skyrocketed 22% and 14% respectively in Hong Kong trading. Both companies have been building agentic AI tools on OpenClaw. SenseTime gained 2.43% and UCloud Technology advanced 13%. OpenClaw is an open-source AI agent platform developed by Peter Steinberger, who has since joined OpenAI. Many Chinese tech companies are integrating OpenClaw or launching their own versions. For US businesses, this signals growing adoption of agentic AI workflows in Asia, which will likely accelerate competitive pressure to adopt similar automation. (CNBC report)
Microsoft reorganizes Copilot under new leadership. Microsoft is unifying Copilot across commercial and consumer products under Jacob Andreou, who will report directly to CEO Satya Nadella. Mustafa Suleyman (Microsoft AI CEO) will now focus on building Microsoft's own AI models instead of working on Copilot's assistant features. This reorganization addresses the longstanding problem that Copilot for consumers and businesses looked completely different and didn't share features. If you're using Microsoft 365 Copilot, expect a more consistent experience across products in the coming months. (The Verge coverage)
Pentagon developing alternatives to Anthropic after contract breakdown. The Pentagon is building tools to replace Anthropic's AI after their $200 million contract fell apart over restrictions on mass surveillance and autonomous weapons. OpenAI and xAI (Elon Musk's company) have both signed agreements with the Department of Defense. Defense Secretary Pete Hegseth designated Anthropic a supply-chain risk, which bars Pentagon contractors from working with Anthropic. For businesses that work with the DOD or defense contractors, using Anthropic's Claude may now create contract complications. (TechCrunch report)
OpenAI signs AWS deal to expand government footprint. OpenAI partnered with Amazon Web Services to sell AI products to the US government for classified and unclassified work. OpenAI models will be available through Amazon Bedrock in AWS GovCloud and classified regions. This is notable because Anthropic uses AWS as its primary cloud provider and is deeply integrated with Amazon Bedrock. OpenAI is essentially moving onto Anthropic's home turf. The deal positions OpenAI to serve multiple government agencies beyond the Pentagon contract. For enterprises, government contracts signal trust and reliability, which often unlocks more private-sector sales. (TechCrunch coverage)
Google pledges $12.5 million to open source security (with Amazon, Anthropic, Microsoft, OpenAI). Google, as a founding member of the Linux Foundation's Alpha-Omega Project, announced $12.5 million in collective funding with Amazon, Anthropic, Microsoft/GitHub, and OpenAI to secure open source software against AI-driven threats. The funding will help maintainers deploy fixes faster and access advanced security tools. Google is extending its Big Sleep and CodeMender tools (AI-powered vulnerability detection and fixing) to open source projects. For businesses that rely on open source libraries, this funding means fewer unpatched vulnerabilities in your dependencies. (Google blog)
Try GPT-5.4 nano for your next high-volume automation. If you're running any workflow that processes hundreds or thousands of items per day (email classification, data extraction, content moderation), swap in GPT-5.4 nano and track costs for one week. You'll likely cut API expenses by 70% or more while getting better performance. Start with one workflow, measure savings, then expand to others.
Audit your AI agent's local file access. If you're using tools like Manus, OpenClaw, or Claude Code that access your local files, review which directories they can read and write to. Create a separate folder for AI agent work and restrict access to sensitive client data, financial records, or credentials. Always use "Allow Once" instead of "Always Allow" until you're confident the agent's behavior is predictable.
The AI cost wars are accelerating. OpenAI's GPT-5.4 nano is dramatically cheaper than previous models, which will force Google, Anthropic, and others to match on pricing or differentiate on capabilities. For small businesses, this means automation that was too expensive six months ago is now viable. Watch for more model releases focused on speed and cost efficiency rather than frontier capabilities.
Agentic AI is moving from hype to deployment. Alibaba's Wukong, Meta's Manus desktop app, and Nvidia's endorsement of OpenClaw all signal that enterprises are moving from chatbots to agents that take autonomous actions. If you're still evaluating whether to adopt AI tools, the question is shifting from "should we?" to "how do we deploy agents safely?" Start with low-risk, high-volume tasks (like organizing files or processing routine documents) before moving agents into customer-facing or revenue-critical workflows.