Get tomorrow's brief in your inbox
Today: Google ships Gemini 3.5 with agentic coding and proactive assistants while companies scramble to control runaway AI costs. Anthropic filed for IPO as Alphabet raised $85 billion to fund the AI race. OpenAI launched Lockdown Mode to prevent prompt injection data theft.
The token bill is coming due, and it's uglier than anyone expected. Companies that gorged on AI subscriptions in 2025 are now facing budget disasters. Uber burned through its entire 2026 AI coding budget by April. Microsoft revoked Claude Code licenses. One company reportedly hit a $500 million Claude bill after failing to set usage limits. The shift from "move fast" to "how do we control this?" happened almost overnight.
The problem isn't per-token pricing (that's falling). It's agentic features multiplying consumption faster than anyone planned for. Per-developer token usage jumped 18.6x in nine months according to Jellyfish research. Even worse, the ROI is murky. Engineers using 10x more tokens are only about 2x more productive, and they're shipping more bugs and rewrites alongside the code.
If you're running AI tools in your business, this is your wake-up call. Set usage limits now. Track who's spending what. The Linux Foundation just launched the Tokenomics Foundation to bring FinOps-style discipline to AI spending, because the industry realized that without guardrails, budgets evaporate faster than value appears. Read more
Google I/O 2026 delivered the agentic era. Gemini 3.5 brings frontier intelligence for agents and coding, while Gemini Omni handles multimodal creation (video generation from any input). The Gemini app is becoming a proactive 24/7 assistant with personalized daily briefs and Gemini Spark, managing your inbox and scheduling in the background. Search gets information agents that monitor topics for you and Antigravity for building custom dashboards and mini apps inside search results. Android Halo gives you a unified space to track your agents' progress. If you've been waiting for AI that does more than answer questions, Google is betting this is it. Read more
OpenAI just shipped Lockdown Mode, and you should turn it on. This new security feature blocks the final step of prompt injection attacks by limiting outbound network requests that could exfiltrate your data. It's rolling out to Free, Go, Plus, Pro, and self-serve ChatGPT Business accounts. Lockdown Mode doesn't stop prompt injections from appearing in content (cached web pages, uploaded files), but it cuts off the attacker's ability to steal data and send it back. If you're feeding ChatGPT sensitive business information, this is the setting that protects you from the Lethal Trifecta (access to private data + exposure to untrusted content + exfiltration vectors). The fact that OpenAI built this implies default ChatGPT doesn't protect against determined data theft. Read more
Simon Willison released micropython-wasm, a Python code sandbox that actually works. If you've ever wanted to safely run user-submitted Python code without risking your server, this alpha package is worth watching. It uses MicroPython compiled to WebAssembly to create a controlled execution environment with memory limits, CPU limits, and zero network access unless you explicitly allow it. It's designed for plugin systems, data transformations, and scheduled automation tasks where you can't trust the code. It installs cleanly from PyPI with binary wheels. If you're building tools where users need to run custom code (enrichment pipelines, data apps, automation platforms), this solves the "how do I not get hacked" problem. Read more
Integration platforms are the air traffic control for your workflow automation. If you're managing data between SaaS apps, on-prem databases, and internal microservices, n8n just published a detailed comparison of eight cloud integration platforms (iPaaS). Key insight: picking an integration layer is a commitment that's hard to reverse. Choose wrong and you're locked into mounting costs and technical debt. Look for deployment models that align with your security requirements and give you visibility to troubleshoot when things break. n8n (open source, self-hostable) is built for technical teams who need HTTP Request nodes and JavaScript flexibility. MuleSoft Anypoint is the enterprise standard for bridging legacy systems with modern SaaS. SnapLogic uses AI-driven "Snaps" for high-volume data flows. The right choice depends on whether you're dealing with sensitive data and whether you need to self-host for compliance. Read more
LinkedIn pipeline attribution is a data problem, not a decision problem. Marketing teams are moving slowly on budget shifts because they don't trust the underlying data. Zapier's new playbook walks through how to make confident spend decisions by tracking pipeline-to-spend (not just cost per lead), comparing campaigns on conversion quality (not lead volume), and flowing closed-won data back to LinkedIn so you're optimizing toward real outcomes instead of form fills. If you're running LinkedIn ads and debating which campaigns deserve more budget, this is the framework to get out of analysis paralysis. Read more
Microsoft Power Automate pricing is more complicated than it looks. The "free" tier bundled with Microsoft 365 only covers standard connectors (Outlook, SharePoint, Teams, OneDrive). The moment you need premium connectors (Salesforce, SAP, SQL Server), desktop flows (RPA), or AI Builder, you're into paid territory. Premium plan costs $15 per user per month. There's also a $5,000/month add-on that isn't a typo. If your org is already on Microsoft 365, some automation is baked in, but the ceiling is low. For teams working across a wider app stack beyond Microsoft tools, Zapier connects to 9,000+ apps and works regardless of ecosystem. Read more
Meta is testing facial recognition in smart glasses code. The company quietly added biometric facial recognition features to its app that would identify people via smart glasses. The exploratory feature would match faces against biometric databases. This hasn't launched yet, but it signals where Meta is heading with wearable AI. If you're running a business where customer privacy is a selling point, this is a trend to watch and potentially contrast against in your messaging. Read more
Alphabet is raising $85 billion in fresh capital to fund its AI buildout, including a $10 billion investment from Berkshire Hathaway. The company already raised over $55 billion in debt since November and now free cash flow is turning negative as capex ramps to as high as $190 billion this year (expected to "significantly increase" again in 2027). Why the cash grab now? SpaceX is IPO-ing next week aiming for $75 billion, Anthropic filed confidentially for IPO, and OpenAI is expected to follow. Capital isn't infinite, and Alphabet is securing its balance sheet before investor attention shifts to the new AI offerings. Google Cloud revenue is up 63% year over year, so the spending is showing up in revenue. But this is the first time Google has needed public equity markets to fund growth. Read more
SpaceX locked in a $920 million per month compute deal with Google. Under the terms, Google will pay SpaceX from October 2026 through June 2029 for access to approximately 110,000 NVIDIA GPUs and related components. This is roughly half the compute that Anthropic secured from SpaceX's Colossus 1 data center in late May. Google framed this as "bridge capacity to meet surging customer demand for our agent platform, Gemini Enterprise." Both deals include a cancellation clause after December 31, 2026 with 90 days' notice. SpaceX announced this one week before its Nasdaq IPO expected to raise $75 billion at a $1.75 trillion valuation (the largest IPO in history). Read more
Trump administration and OpenAI are discussing a possible government stake in the company. The talks have been ongoing for more than a year, with Sam Altman first pitching the idea in 2025. OpenAI could donate equity to seed a "Public Wealth Fund" that would invest in long-term assets and enable citizens to participate in AI growth upside. No official terms have been decided. Trump signed an executive order in February calling for a sovereign wealth fund. The administration has already taken stakes in Intel, IBM, and other quantum and critical mineral companies. OpenAI is valued at over $850 billion and gearing up for IPO as soon as this year. Read more
Anthropic called for a global slowdown in AI development. The company flagged the risk of models "self-improving" and wants a coordinated plan to stop them. The timing is awfully convenient given Anthropic's confidential IPO filing and the backdrop of Alphabet's $85 billion raise. Skeptics note this is classic AI lab behavior: race to build frontier models, then call for regulation once you're far enough ahead. Read more
New York passed a one-year moratorium on new large data centers (20+ megawatts). If Governor Kathy Hochul signs it, this will be the first statewide ban of its kind. The bill directs the state's environmental agency to assess electricity, water, land use, and pollution impacts, and requires companies to hold and fund public hearings at least three months before gaining approval for projects. The state is currently reviewing 24 data center proposals totaling over 9,000 megawatts. Surveys show most Americans oppose data centers in their communities, and this is becoming a galvanizing issue across the political spectrum. Maine tried a similar ban earlier this year but the governor vetoed it. Industry groups argue the moratorium will damage the economy by blocking helpful projects. Read more
Stop AI hallucinations before they reach production. n8n's new guide breaks down the five types of hallucinations (factual fabrication, citation hallucination, inconsistent labeling, refusal failures, prompt constraint violations) and how to catch them. Quick fix: lower the temperature and set seeds for extraction, classification, and structured generation tasks. For anything that needs exact accuracy, use retrieval over verified sources instead of relying on the model's parametric memory. Track quality metrics alongside execution metrics (success rate, latency, tool call count). If you're running AI agents in production and haven't instrumented outputs for hallucination detection, you're flying blind. Read more
Use Zapier Formatter to fix data format mismatches automatically. When you're syncing data between apps and names are together when they should be split, dates are formatted wrong, or text is in the wrong order, Formatter can handle it without manual editing. Split names into first and last, extract email addresses or links from text blocks, change date formats, and capitalize text. Formatter steps don't count toward your task usage on Pro plans and higher. If you're spending time manually reformatting exported data before importing it somewhere else, this tool does it for you. Read more
The AI industry is entering a new phase where costs are catching up to capabilities. Companies that treated AI subscriptions as unlimited are discovering hard budget limits, and a market is forming around tokenomics and cost control. Meanwhile, the biggest players (Alphabet, OpenAI, SpaceX) are raising historic amounts of capital to fund compute buildouts, signaling that the infrastructure race is accelerating even as businesses pull back on spending. Watch for more tools focused on AI cost management, usage limits, and ROI measurement. The "move fast and figure it out later" era is over.