Get tomorrow's brief in your inbox
Today: DeepSeek V4 just undercut frontier AI pricing by 90% while matching GPT-5.5 performance. OpenAI released prompting guidance showing GPT-5.5 needs a fresh start, not copy-paste upgrades. AI agents matured from hype to deployment-ready tools, but enterprises are choosing governed platforms over OpenClaw's security risks.
DeepSeek V4 just changed the economics of AI for small businesses. The Chinese AI lab released two preview models (V4 Pro and V4 Flash) that match OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6 on coding and reasoning benchmarks, but cost a fraction of the price. V4 Flash costs $0.14 per million input tokens and $0.28 per million output tokens, undercutting every major provider including GPT-5.4 Mini, Gemini Flash, and Claude Haiku. The larger V4 Pro costs $0.145/$3.48 per million tokens, still cheaper than GPT-5.5, Claude Opus, and Gemini Pro.
This matters because it breaks the assumption that frontier AI performance requires frontier AI budgets. If you've been holding back on building AI features because API costs scared you off, that calculus just changed. DeepSeek V4 is open-weight (you can download and run it yourself) with a 1 million token context window, large enough to process entire codebases or documentation sets in a single prompt. The Pro model has 1.6 trillion parameters (49 billion active), making it the biggest open-weight model available.
The catch: DeepSeek models are text-only (no image, audio, or video support), and they trail frontier models by 3-6 months on knowledge tests. But for coding, automation, document analysis, and agent tasks, the performance gap has essentially closed. More than 90% of experienced developers in DeepSeek's internal survey chose V4 Pro as a top model for coding work.
What to do: If you're building AI features, test DeepSeek V4 Flash for high-volume workflows where cost has been a blocker. If you're currently using GPT-4 or Claude Sonnet for tasks that don't require multimodal input, run a cost comparison. The API is live now at deepseek.com.
DeepSeek previews new AI model | Why DeepSeek V4 matters
GPT-5.5 is available in the OpenAI API, and it needs different prompts than GPT-5.4. OpenAI released a prompting guide that opens with a warning: treat GPT-5.5 as a new model family to tune for, not a drop-in replacement. Start with the smallest prompt that preserves the product contract, then tune reasoning effort, verbosity, and output format against examples. Don't carry over every instruction from older prompt stacks.
One useful trick for long-running tasks: send a short user-visible update before tool calls that acknowledges the request and states the first step. Keep it to one or two sentences. This makes the model feel responsive instead of crashed. OpenAI's Codex app already does this.
If you're migrating existing code, run $openai-docs migrate this project to gpt-5.5 in Codex. The upgrade guide includes light instructions on rewriting prompts to fit the model. Simon Willison's llm CLI tool now supports GPT-5.5 with verbosity controls (low, medium, high) and image detail settings.
What to do: If you're using GPT-5.4 or earlier in production, don't just swap the model ID. Re-baseline your prompts and test against representative examples. GPT-5.5 prompting guide | llm 0.31 release
Nothing launched Essential Voice, an AI-powered dictation tool that works system-wide. It turns speech into formatted text in any app, removes filler words, and supports 100+ languages with real-time translation. You can create custom voice shortcuts for repeated phrases, links, or templates (say "my address" to insert your full address). The feature is available now on the Phone (3), rolling out to Phone (4a) Pro later this month and Phone (4a) next month.
This is one of the first system-level AI dictation integrations on Android. Similar tools like Superwhisper, Wispr Flow, and Willow exist, but they're app-specific or require workarounds. Nothing's approach is closer to Google's offline dictation tool released recently. If you spend hours typing emails, messages, or notes, system-level dictation saves real time.
What to do: If you're on a supported Nothing phone, try Essential Voice for drafting emails or messages. For iPhone users, Superwhisper recently released action key mapping for dictation. Nothing introduces AI dictation
ComfyUI raised $30 million at a $500 million valuation for node-based AI media control. The tool started as an open-source project in 2023 to give creators granular control over diffusion models (Midjourney, DALL-E, Stable Diffusion). Instead of prompt roulette, where changing one detail overwrites everything, ComfyUI's node-based interface lets you link specific components of the generation process. You can change one element (lighting, clothing, background) without re-rolling the entire image.
ComfyUI claims 4 million users, including visual effects studios, ad agencies, and industrial design teams. Job boards now list "ComfyUI artist or engineer" as a role. The company raised $19 million in Series A in late 2024, and this new round (led by Craft Ventures) brings total funding to $49 million. CEO Yoland Yan says even as foundational models improve, they still get 60-80% of the way there, and the last 20% requires precision control that prompting can't deliver.
What to do: If you're creating marketing visuals, social media content, or product mockups with AI, try ComfyUI. The open-source version is free. The startup offers managed services for teams. ComfyUI funding announcement
Gemini gets spring cleaning features: image-based clutter audits, fridge inventory scanning, and inbox decluttering. Google's latest Gemini tips focus on organizing physical spaces and digital life. You can upload a photo of a cluttered drawer and ask for spatial suggestions. Use Gemini Live to pan your camera across your fridge and get recipe ideas from leftovers. Ask Maps to find donation centers or eco-friendly cleaning supplies along your route. Nano Banana (Gemini's image editing model) can show you how a room would look with a new paint color or furniture arrangement without leaving your phone.
Ultra subscribers in the U.S. get Agent Mode for inbox management. Ask Gemini to "help organize my inbox" and it creates a plan with smart actions like archiving clutter or creating tasks from reminders. You approve with a click. AI Inbox (also for Ultra subscribers) shows to-dos and topics that matter most in one place.
What to do: If you're a Google Workspace user, try the fridge inventory feature to reduce food waste. If you're drowning in email, Ultra's Agent Mode might be worth the subscription cost. Gemini spring cleaning tips
The best AI agents for enterprises aren't the ones with the most hype. Zapier's guide to AI agents (based on hours of testing and real user reviews) ranks deployment-ready agents by completeness, security, and ease of rollout. Top picks: Zapier Agents (enterprise automation across 9,000+ apps), Claude Code and Cowork (agentic desktop and coding work), ChatGPT Workspace Agents (research and task completion), and Lindy (personal assistant in iMessage and your inbox).
What separates enterprise-ready agents from personal hobby tools: managed credentials with scoped permissions, audit logging, human-in-the-loop controls, integration coverage, safety and compliance checks, and predictable cost with support. Agents that run on your laptop with full system access and community-contributed skills are governance risks, no matter how good the underlying model is.
Zapier competes directly with Workato, which targets IT-led, centralized automation for large enterprises wiring heavy systems like SAP, NetSuite, and Salesforce. Workato fits teams where integration engineers control what runs in production. Zapier fits teams that want bottom-up innovation where non-technical people can build and own automations while admins maintain visibility and governance. Best AI agents for enterprises | Workato vs. Zapier comparison
OpenClaw's security problems are a cautionary tale for enterprises. The open-source AI agent became a breakout developer tool in early 2026, but security researchers found tens of thousands of exposed instances on public IPs with weak credentials. Malicious skills disguised with professional documentation have been found on ClawHub, the public marketplace. Once installed, they can run code on your system, potentially installing keyloggers or malware.
OpenClaw runs with system-level permissions and can execute code from external sources. When you connect it to Slack or Google Workspace, the agent can reach messages, emails, documents, and authentication tokens. There's no built-in governance, no centralized audit trails, and no compliance controls by default. For regulated companies, that's a disclosable incident waiting to happen.
The alternatives: Zapier (end-to-end AI orchestration), Claude Code and Cowork (agentic desktop work), n8n (self-hosted automation for teams that need data residency), Cursor (AI-assisted coding inside an IDE), AWS Bedrock Agents (custom agents in enterprise cloud), and Microsoft Power Automate (RPA and process automation for Microsoft shops). OpenClaw alternatives for enterprise
A Hollywood production startup is using AI to cut costs and time on film production. Innovative Dreams (backed by AWS and Luma) combines cameras, LED walls, virtual production, motion capture, and AI tools (Luma, Google's Nano Banana, ByteDance's SeeDream) to shoot shows without leaving a soundstage. The first project, "The Old Stories: Moses," shot in a single week with actors in 40 locations, compared to five or six weeks for traditional production. The show debuts this spring.
CEO Jon Erwin says the approach merges traditional filmmaking (actors, cameras, lens choice) into a digital world rather than replacing cameras with prompts. AWS provides cloud and AI infrastructure for real-time hybrid production tools. Luma (valued over $4 billion) invested and is using Innovative Dreams for product feedback.
The risk: Los Angeles County has lost over 40,000 entertainment industry jobs since 2022, with production activity at the lowest level since 1995. The writers and actors guild strikes in 2023 were driven partly by AI concerns about job losses. AI tools that cut production time also cut production jobs. Amazon-backed Hollywood AI startup
Meta signed a deal to use millions of Amazon Graviton chips for AI workloads. The three-year deal makes Meta a top-five customer of AWS Graviton, which are ARM-based CPUs designed for "always-on reasoning workloads" like real-time AI agents, code generation, search, and multi-step task coordination. This is separate from GPUs used for training large models. Once models are trained, agents create compute-intensive workloads that CPUs like Graviton handle more cost-effectively than GPUs.
Amazon CEO Andy Jassy said the company's annual revenue run rate for its chip business (Graviton, Trainium, Nitro) is over $20 billion, growing triple-digit percentages year-over-year. The Meta deal brings more of Meta's cash back to AWS instead of Google Cloud, where Meta signed a six-year, $10 billion deal last August. AWS timed the announcement right as Google Cloud Next wrapped up, a virtual smirk at its cloud rival. Meta signs Amazon Graviton deal
Google plans to invest up to $40 billion in Anthropic. The deal includes $10 billion now (at a $350 billion valuation) with another $30 billion if Anthropic hits performance targets. This follows Anthropic's recent infrastructure scramble: a $5 billion Amazon investment earlier this month, a deal with CoreWeave for data center capacity, and a partnership with Google and Broadcom for 3.5 gigawatts of TPU-based computing capacity starting in 2027. Google Cloud is now providing 5 gigawatts of capacity over the next five years with room to scale further.
Anthropic's valuation stood at $350 billion in February; investors have since been eager to back the company at $800 billion or more. The company is reportedly considering an IPO as soon as October. This matters because it shows the AI race is increasingly defined by access to compute, not just model quality. OpenAI has moved aggressively to secure capacity through multi-hundred-billion-dollar deals. Anthropic is following the same playbook. Google invests $40B in Anthropic
Mac minis are sold out and flooding eBay at markups due to AI demand. The $599 M4 Mac mini base model (16GB RAM, 256GB storage) is sold out on Apple's retail website with no delivery or in-store pickup options. Higher storage models (512GB+) ship in June. eBay listings show markups of $100-$300 over retail. The shortage is driven by AI developers using Mac minis to run on-device models like OpenClaw, ZeroClaw, Anthropic and OpenAI tools, Perplexity Computer, and local models. Mac minis run quietly and are more reliable for 24/7 use than laptops.
The shortage coincides with an industry-wide memory crunch and plans for a Mac mini refresh. Apple hasn't commented. Mac Studios are also sold out across several configurations. What this tells you: if you're building AI tools that run locally, your customers are willing to pay a premium for hardware that can handle always-on workloads. Mac mini shortages and eBay markups
Tim Cook is stepping down as Apple CEO in September, handing the reins to hardware chief John Ternus. Ternus inherits one of the most durable businesses in tech, but also a very different ecosystem than the one Cook spent decades shaping. The App Store's 30% cut is under pressure, the behind-the-scenes power Apple once held over developers is being challenged, and vibe-coded apps are changing what it means to build on Apple's platform. Cook's legacy includes AirPods, the Apple Watch, and the shift to Apple Silicon, but also the Touch Bar and years of criticism for incremental iPhone updates. Tim Cook stepping down
SpaceX has a $60 billion option to buy Cursor, the AI-powered code editor. The deal includes a $10 billion breakup fee, signaling Elon Musk's AI strategy post-xAI merger with SpaceX. xAI filed for an IPO earlier this year. The deal raises questions about whether Musk is building an AI strategy around coding tools or simply preventing competitors from acquiring them. Cursor competes with GitHub Copilot, Replit, and other AI coding assistants. SpaceX-Cursor $60B option
Meta is hiring from Thinking Machines Lab, which is hiring from Meta. The AI talent war runs in both directions. TML hired Weiyao Wang (8 years at Meta, worked on SAM3D) and Kenneth Li (10 months at Meta). TML's CTO Soumith Chintala co-founded PyTorch at Meta (11 years). Business Insider reported last week that Meta has poached seven of TML's founding members. LinkedIn profiles show TML has been hiring more researchers from Meta than from any other single employer. TML's headcount is around 140. The startup is valued at $12 billion and just signed a multibillion-dollar cloud deal with Google for Nvidia GB300 chips. Meta vs. Thinking Machines talent grab
Elon Musk vs. Sam Altman trial starts April 27 in Oakland. Musk is suing OpenAI for fraud, claiming Altman and Greg Brockman deceived him. Several claims (including a laughable invocation of racketeering law) have been dismissed. High-profile witnesses include Microsoft CEO Satya Nadella, CTO Kevin Scott, former OpenAI executives Ilya Sutskever and Mira Murati, and former board members involved in Altman's 2023 ouster. The trial is a win for Musk even if he loses, because it damages OpenAI's reputation and produces embarrassing disclosures (Musk's "rhino ket" use, Greg Brockman's diary excerpts, Mark Zuckerberg's texts about moderating content for DOGE). Musk vs. Altman trial preview
Trump threatens "big tariff" on UK unless it drops digital services tax on U.S. tech firms. The 2% levy on revenues of search engines, social media, and online marketplaces affects Google, Meta, and Apple. The tax raised £800 million ($1.08 billion) in 2024-2025. Trump said, "If they don't drop the tax, we'll probably put a big tariff on the U.K." without providing a specific figure. The comments follow publicly critical remarks about UK Prime Minister Keir Starmer ahead of King Charles III's four-day U.S. state visit starting Monday. Trump tariff threat to UK
Use DeepSeek V4 Flash for high-volume summarization tasks. If you're summarizing customer feedback, support tickets, or research articles at scale, test DeepSeek V4 Flash at $0.14/$0.28 per million tokens. For 10,000 summaries of 500-word inputs with 100-word outputs, GPT-4o Mini costs about $1.50. DeepSeek V4 Flash costs about $0.16. Over a year at 10,000 summaries per week, that's $780 vs. $83. API docs at deepseek.com.
Set up a system-level dictation shortcut for repeated phrases. Whether you're using Nothing's Essential Voice, Superwhisper, or Google's offline dictation, create voice shortcuts for email signatures, addresses, standard responses, or meeting links. Say "meeting link" instead of typing your Zoom URL 50 times a day. Five minutes of setup saves hours over a month.
Test GPT-5.5's status update pattern for long-running tasks. If your app uses AI agents or multi-step workflows, send a short user-visible update before tool calls: "Got it, I'll search for pricing data and compare the top three options." One sentence. It makes the wait feel intentional instead of broken. Users trust the system more when they know what's happening.
Three patterns across today's stories: the frontier is fragmenting, agents are graduating from demos to deployment, and infrastructure deals are defining who wins.
DeepSeek V4 proves you don't need $10 billion in capital to compete at the frontier on reasoning and coding. Open-weight models are closing the gap on closed-source leaders, and pricing is collapsing faster than most expected. If you've been waiting for AI to become affordable enough to build on, the wait is over.
AI agents matured from 2024's promise to 2025's hype to 2026's deployment reality. The gap between a capable agent and a safely deployable agent is wider than the demos suggest, and enterprises are choosing governed platforms (Zapier, Claude Code, ChatGPT Workspace) over self-hosted tools with security risks (OpenClaw). If you're evaluating agents, prioritize audit logging, scoped permissions, and human-in-the-loop controls over raw capability.
The AI race is now an infrastructure race. Anthropic, OpenAI, Meta, and others are signing multi-hundred-billion-dollar deals for compute, chips, and energy. Google, Amazon, and Microsoft are competing to be the infrastructure suppliers. The companies that secure capacity today will train and deploy the next generation of models. For small businesses, this means cloud costs and API pricing will be battlegrounds, and competition is working in your favor.
Watch for: DeepSeek V4 API stability under load, GPT-5.5 adoption rates among developers migrating from GPT-5.4, and whether OpenClaw's security issues force a governance reckoning in the open-source agent ecosystem.