Get tomorrow's brief in your inbox
April 3, 2026
Today: Google added Flex and Priority API tiers, cutting costs 50% for background tasks while guaranteeing reliability for critical workflows. OpenAI bought tech talk show TBPN for $120M to shape AI conversation. Microsoft shipped three new foundation models focused on transcription, voice, and image generation at lower cost than OpenAI or Google.
Google's new Flex and Priority tiers solve the agent economics problem
If you're building with AI agents, you've hit this wall: background tasks like research, data enrichment, or batch processing burn through API costs at standard rates, while user-facing features like chatbots need rock-solid uptime. Until now, that meant splitting your architecture between synchronous APIs and complicated batch systems.
Google just fixed it. The new Flex tier cuts API costs by 50% for latency-tolerant workloads while keeping the synchronous interface you already know. No job polling, no file uploads, just standard API calls at half price. Priority tier does the opposite: highest reliability for critical apps, with automatic graceful downgrade to Standard tier if you hit limits instead of failing completely.
For small businesses, this is a build-versus-buy moment. If you're running agents that do background CRM updates, content generation, or research loops, Flex immediately cuts your bill in half. If you're running customer-facing bots where downtime costs real money, Priority keeps you online even during peak platform load. Both tiers are available now for paid accounts across GenerateContent and Interactions APIs.
OpenClaw vs. Zapier: Which AI automation platform fits your business?
OpenClaw went from side project to global phenomenon in weeks. It's an open-source AI agent that runs on your machine and lets you control it through WhatsApp, Telegram, Discord, Slack, or Signal. You get full system access, persistent memory, and zero ongoing costs beyond API keys. The catch: you host it, secure it, and govern it yourself.
Zapier Agents is the managed alternative. You get 8,000+ app integrations, no-code setup, built-in governance (scoped permissions, human-in-the-loop approvals, activity dashboards), and model flexibility across Anthropic and OpenAI. Free plan available, paid plans start at $33.33/month.
Which one? If you're technical and want total control over a personal assistant, OpenClaw is free and powerful. If you're a team deploying AI automation at work with the guardrails to do it responsibly, Zapier Agents is built for that.
OpenClaw vs Zapier | Building safe AI agents
Google Vids adds AI avatar directing, Veo 3.1 video generation, free for everyone
Google Vids now lets you direct AI avatars through text prompts. Tell an avatar to interact with a product, demo equipment, or act out a scene, and Vids maintains character consistency across the video. You can customize appearance, change outfits, and create new backgrounds, all through prompts.
Every Google account gets 10 free Veo 3.1 video generations per month (8-second clips). Google AI Pro and Ultra subscribers can generate custom music tracks (30 seconds to 3 minutes), export directly to YouTube, and record screens via Chrome extension. Ultra accounts get 1,000 Veo videos per month.
If you've been paying for Synthesia or HeyGen for product demos or explainer videos, this is worth testing. The free tier is generous enough for small business marketing and internal training content.
Microsoft ships MAI-Transcribe-1, undercutting OpenAI and Google on price
Microsoft released three foundation models: MAI-Transcribe-1 (speech-to-text in 25 languages), MAI-Voice-1 (60 seconds of audio generated in 1 second), and MAI-Image-2 (image generation). All three are now available on Microsoft Foundry and AI Playground.
Pricing undercuts the competition: MAI-Transcribe-1 starts at $0.36/hour (2.5x faster than Azure Fast), MAI-Voice-1 at $22 per 1 million characters, MAI-Image-2 at $5 per 1 million tokens for text input. Microsoft says it's "half the GPU cost of other state-of-the-art models."
If you're transcribing meetings, customer calls, or podcasts at scale, this pricing makes it worth comparing against Rev, Otter, or Descript's API offerings.
Source | Microsoft's transcription model announcement
Google Gemma 4: Open-source models with vision and audio
Google released four new Apache 2.0 licensed models: Gemma 4 E2B, E4B, 31B, and 26B-A4B (mixture-of-experts). All four support vision, and the two smaller models (E2B, E4B) support native audio input for speech recognition.
The "E" in E2B stands for "effective parameter count." Google used Per-Layer Embeddings to maximize efficiency, meaning the models run fast on-device despite their capabilities. The 2B model is 4.41GB, the 4B is 6.33GB, and the 26B-A4B is 17.99GB.
If you're building on-device AI apps (mobile, edge devices, offline tools), these models give you vision and audio input without requiring cloud API calls. Free, open-source, and commercially licensed.
Granola security warning: Your meeting notes are public by default
If you use Granola (AI-powered meeting notes), check your settings. The app makes notes "viewable to anyone with a link" by default, and uses your data for AI training unless you opt out.
To fix: Open Granola, go to Settings, change "Default link sharing" from "Anyone with the link" to "Private." To disable AI training, toggle off "Use my data to improve models for everyone."
If you've shared Granola links in Slack, email, or project management tools, those notes are visible to anyone with the URL, even without a Granola account. If you're discussing sensitive business strategy, client details, or internal decisions, audit your shared links now.
How Zapier's YouTube manager built an influencer ROI dashboard with Claude Code
Zapier tracks over 100 influencer partnerships across YouTube, LinkedIn, and X. Each creator's content is logged in a Zapier Table: views, Bitly clicks, conversions, sponsorship cost, even LLM citation data (when a video appears in AI search results).
The problem: staring at 100+ rows of raw data doesn't show trends. The solution: a Claude Code-built dashboard with time-series charts, sortable creator tables, and filters for date range, creator, and metric type. Total build time: 25 hours (30 minutes for initial setup, the rest for prompt refinement).
Key insight from the dashboard: views and clicks move in tandem, conversions follow the same pattern at lower amplitude. One creator's video (Kevin Stratvert's "Zapier AI for beginners" tutorial) outperformed everything else on conversions because it targeted people actively looking to solve a problem, not passive viewers.
Takeaway: If you're managing creator partnerships, affiliate programs, or any multi-channel marketing, a custom dashboard built with Claude Code can surface patterns your spreadsheet hides. The iteration loops are fast, and you don't need to be technical.
Marketers are using coding agents (Claude Code, Cursor, Codex) to build workflows
Tools that engineers use to write code are now some of the most powerful tools in a marketer's stack. Claude Code, Cursor, and Codex can build workflows, connect to APIs, scrape data, generate content, and automate multi-step processes through plain English instructions.
Real example: Morning Brew co-founder Alex Lieberman built a "content machine" in Claude Code that turns one idea into dozens of platform-specific posts. The system scans Slack and Notion for "content spikes" (strong stories or opinions), runs AI interviews to draw out the best version, generates drafts for LinkedIn, X, newsletters, and long-form content, then runs drafts through an AI editor board for revision. Total build time: 25 hours. Result: content he posts directly without editing.
The unlock: you don't need to be technical. If you can describe what you want in English, coding agents can build it. Zapier's YouTube channel manager (not an engineer) is saving hours per week with automated workflows he built himself.
n8n's Production AI Playbook: How to build reliable AI systems
n8n published a playbook on building production AI systems. The core principle: use AI only where you need AI. Everything else should be handled by deterministic steps (formatting dates, validating emails, routing based on status fields).
Why? Every LLM call costs tokens and adds latency. A code node that validates an email address is instant and free. An LLM doing the same thing is slower, costs money, and might still get it wrong. Multiply that across thousands of workflow executions and the costs add up fast.
The reliability gap in AI workflows rarely comes from the model itself. It comes from messy inputs, missing validation, ambiguous routing, and lack of structure around what goes in and what comes out. The fix: wrap AI steps in deterministic logic that cleans data before the model sees it, validates output before it moves downstream, and routes decisions through explicit rules.
n8n's workflow architecture makes this pattern easy to implement. If you're building AI automations in production (customer-facing bots, data enrichment pipelines, content generation), this playbook is worth reading.
Introduction | Deterministic + AI steps
OpenAI adds pay-as-you-go pricing for Codex (Business and Enterprise)
Codex now includes pay-as-you-go pricing for ChatGPT Business and Enterprise customers, giving teams a more flexible option to start and scale adoption without committing to seat-based billing upfront.
If you've been holding off on Codex because seat-based pricing didn't fit your team size or usage patterns, this makes it easier to test and expand gradually.
OpenAI acquires TBPN tech talk show for reported $120M
OpenAI bought TBPN (Technology Business Programming Network), a daily live talk show that interviews AI executives and tech leaders. The show airs weekdays at 2PM PT for three hours on YouTube and X, averaging 70,000 viewers per episode. It generated $5M in ad revenue this year, with projections of $30M in 2026.
TBPN will retain editorial independence but also help with OpenAI's corporate comms and marketing. The team reports to Chris Lehane, OpenAI's VP of global policy (former Clinton aide, crypto super PAC founder, master of "political dark arts").
What this means for your business: OpenAI is fighting a public image battle ahead of its reported IPO this year. Buying a media platform lets them shape the AI conversation while claiming editorial independence. Watch for how TBPN covers OpenAI controversies (DoD deal backlash, Sora shutdown, revenue pressure) versus how they covered them before the acquisition.
TechCrunch | The Verge | OpenAI announcement
Kintsugi shuts down after failing to get FDA clearance for depression-detecting AI
Kintsugi spent seven years building AI that detects signs of depression and anxiety from speech patterns (how you talk, not what you say). After failing to secure FDA clearance in time, the company shut down and open-sourced most of its technology.
The problem: FDA's "De Novo" pathway is designed for traditional medical devices (hip implants, pacemakers) whose design stays fixed once approved. For AI systems that continuously improve and update, the framework doesn't fit. Add in government shutdowns that slowed the process, and Kintsugi ran out of funding waiting for final submission.
The company's founder said regulatory experts confirm "there's nothing that helps them do that except loud yelling from the top," despite Trump administration pressure to cut red tape for AI.
If you're building healthcare AI, this is a cautionary tale about FDA timelines and funding runway. Plan for years, not months.
Audit your AI tool privacy settings this week
The Granola incident is a reminder: AI tools often default to "convenient" settings that expose your data. If you use any AI meeting assistant, transcription service, or note-taking app, check:
Most tools let you disable AI training and lock down sharing, but you have to do it manually. Budget 15 minutes this week to audit your stack.
Test Google Vids for your next product demo
If you're paying for Synthesia, HeyGen, or Descript for product demos or explainer videos, Google Vids just gave you 10 free video generations per month. That's enough to test whether AI avatars and video generation work for your use case before committing to a paid plan elsewhere.
Three patterns across today's stories: pricing pressure, production readiness, and platform consolidation. Google and Microsoft are undercutting OpenAI on price (Flex tier at 50% off, MAI-Transcribe at half the GPU cost). Teams are moving from "AI works sometimes" to "AI works reliably in production" by wrapping models in deterministic logic. And OpenAI is buying media companies while Google and Microsoft ship their own foundation models, signaling the platform wars are heating up.
Watch for: More aggressive API pricing from Anthropic and OpenAI in response to Google's Flex tier. Expect Microsoft to expand MAI model lineup with reasoning and multimodal capabilities. And keep an eye on how TBPN's editorial independence holds up under OpenAI ownership.