This week the AI industry served up one of its busiest Google I/O keynotes in years, a meaningful shift in business spending between the two top model vendors, and steady consolidation in how agents talk to tools and to each other. The center of gravity moved away from raw model launches and toward the plumbing around the models. That is good news for anyone trying to plan a real product roadmap rather than chase the next benchmark.
There was no flashy new frontier model release in the past seven days. What landed instead were platform changes that will shape how teams build, deploy, and pay for AI for the rest of 2026. Google rebuilt its agentic coding tool, shipped a stable command-line interface for Android development driven by AI agents, and proposed a new open web standard. Anthropic crossed a business adoption milestone that would have been hard to predict a year ago. And the protocol layer that lets agents reach into tools and into each other kept maturing in the background.
The thread connecting all of it is straightforward. The interesting work is happening at the seams between agents, between coding tools, between models and the real world. Below is what changed and why it matters.
AI Coding Tools: Google Bets the Stack on Antigravity 2.0
Google held its annual I/O developer keynote on May 19 in Mountain View. The headline announcement for anyone who writes code for a living was Antigravity 2.0, a complete rebuild of the agentic coding platform Google launched in November 2025.
The original Antigravity shipped as a VS Code fork tuned for AI-assisted development. It looked a lot like Cursor, just with Gemini in the driver seat. The new version goes much further. Google now treats Antigravity as a brand for an entire family of tools rather than a single editor. Version 2.0 includes a redesigned desktop app focused on running multiple agents in parallel, a brand-new command-line tool written in Go, an SDK for building custom agents, and a Managed Agents service inside the Gemini API for cloud execution. TechCrunch reported that the new desktop application lets developers orchestrate parallel subagents, design custom workflows, and schedule tasks that run in the background.
The most dramatic demo came during the keynote itself. Google claimed Antigravity 2.0 built the core framework of a working operating system in about 12 hours, as reported by Digit. According to the company, the platform spawned 93 separate subagents during the task, processed billions of tokens, and finished the project for under $1,000 in compute. To prove the result worked, Google ran the classic game Doom on the AI-built OS during the keynote. Doom failed to launch at first because keyboard drivers were missing, so Google asked Antigravity to write the drivers live on stage. The game then ran.
That story is half marketing theater and half a real data point. The marketing part is obvious. The data point is the cost figure. A working OS framework for under a thousand dollars in compute, even with caveats, is a different unit economics conversation than building software was two years ago.
Google also signaled where its older free tool is heading. The Next Web reported that consumer access to Gemini CLI and the Gemini Code Assist IDE extensions will end on June 18, 2026, for AI Pro, AI Ultra, and free-tier users. Enterprise customers on Gemini Code Assist Standard or Enterprise licenses keep access. The message to consumer developers is direct. Move to Antigravity.
For Android developers specifically, Google released Android CLI 1.0 at the same event. TechCrunch's coverage of Android CLI 1.0 explained that the toolset gives AI agents direct programmatic access to Android Studio's capabilities from the terminal. The first version supports Claude Code, OpenAI Codex, and Antigravity. Agents can run semantic code analysis, render Jetpack Compose previews, detect warnings, and run automated UI tests without ever opening the IDE. Google also bundled Android CLI support directly into Antigravity 2.0 so the platform can take a project from creation through deployment to a virtual Android device.
This is a quiet but important admission. Google built its own coding platform but openly designed the Android tooling to work with competitors. The Android team apparently looked at what developers actually use day-to-day and decided to meet them where they were. That is a different posture than the one major platforms typically take when they ship developer tools.
The rest of the I/O coding lineup filled out the picture. Google launched the Gemini 3.5 Flash model as the engine behind its managed agents in the Gemini API. The company shipped native Android app creation inside AI Studio, where you can describe an app in plain language and preview it inside an integrated emulator. Google also debuted Android Bench, a benchmarking leaderboard that ranks AI coding models on Android-specific development tasks, with support for open-weight models like Gemma 4.
Now zoom out from Google for a moment. The competitive picture for AI coding tools looks different than it did six months ago. The New Stack reported that Cursor, Claude Code, and OpenAI Codex are converging into a layered stack rather than picking a single winner. A survey by the Pragmatic Engineer in February 2026 found that Claude Code had the highest "most loved" rating among 906 software engineers, with 46% naming it their favorite. SemiAnalysis estimated that Claude Code accounted for roughly 4% of all public GitHub commits as of March 2026, with projections suggesting that number could hit 20% by year-end.
Cursor still leads on the inline editing experience inside the IDE. Claude Code dominates the terminal and the agentic surface. Codex is gaining traction for asynchronous, long-running tasks that run in a cloud sandbox without developer attention. Codex passed 3 million weekly active users in March 2026, up from 2 million a month earlier. JetBrains research published in April 2026 showed that 90% of developers used at least one AI tool at work as of January 2026, with 74% using specialized AI development tools beyond chat. GitHub Copilot remained the most adopted at 29% workplace usage, with Cursor and Claude Code tied at 18%.
The takeaway for engineering leaders is practical. There is no single AI coding tool to standardize on. Teams are running two or three in different contexts. A typical setup might pair Claude Code in the terminal for agentic work, Copilot or Cursor inside the IDE for autocomplete and inline edits, and a chat interface for thinking through architecture. That setup did not exist as a recommendation last year. It is the default now.
AI Processing: TPUs, Custom Silicon, and the Cost Question Behind Antigravity's Doom Demo
The hardware story this week was less about new chips and more about the math behind the demos. When Google said it built an operating system for under $1,000 in compute, the implicit follow-up question is what was running underneath.
That answer comes back to Google's eighth-generation Tensor Processing Unit family. TechCrunch reported at Google Cloud Next in April that Google announced TPU 8t and TPU 8i, with the 8t built for training and the 8i for inference. The company claimed up to 3x faster AI model training versus previous generations, 80% better performance per dollar for inference, and the ability to link more than 1 million TPUs into a single cluster.
According to Google's own AI infrastructure announcement, TPU 8i triples on-chip SRAM to 384 MB and increases high-bandwidth memory to 288 GB per chip. The chip doubles inter-chip interconnect bandwidth to 19.2 Tb/s and includes a dedicated Collectives Acceleration Engine that reduces on-chip latency by up to 5x. Those specs matter for one reason. They are tuned for the workload Antigravity 2.0 demonstrated on stage. Many agents running in parallel, each holding a large context window, calling tools, and emitting tokens fast enough to feel interactive.
The shift toward custom silicon is now the defining story in AI hardware. According to TrendForce data summarized by AIMultiple, custom ASIC shipments from cloud providers are projected to grow 44.6% in 2026, while GPU shipments are expected to grow 16.1%. The same report tracks the projection that ASICs will make up 40% of the AI inference market in 2026, up from 15% in 2024. Hyperscalers are no longer just buying NVIDIA. They are designing their own chips for their own workloads and using NVIDIA in parallel.
Anthropic illustrates the multi-platform pattern more clearly than any other model vendor. Anthropic announced plans to expand its use of Google Cloud technologies up to one million TPUs, an expansion worth tens of billions of dollars that should bring over a gigawatt of capacity online during 2026. Anthropic also remains committed to Amazon Trainium as its primary training partner and continues to use NVIDIA GPUs. The company is one of the few customers running serious workloads on all three platforms simultaneously.
There was a quiet hardware story tucked inside the Anthropic news this week as well. Digitimes reported on May 20 that Anthropic hired a high-profile AI researcher with a background at OpenAI and Tesla to accelerate Claude pretraining research. The same article noted a fresh compute rental agreement to support the work. Tying that to the May 6 Anthropic announcement of a compute deal with SpaceX and the earlier $200 million Gates Foundation partnership announced on May 14, the pattern is clear. Anthropic is locking in compute capacity across multiple vendors and multiple physical sites as fast as it can sign deals.
The cost question that follows from all of this is the one every engineering organization is now wrestling with. AI workloads have lost their flat-rate pricing tier. InfoWorld reported on May 14 that Anthropic will separate programmatic Claude usage from standard chat subscription limits beginning June 15. The new policy creates a dedicated monthly credit pool for tools like the Agent SDK, GitHub Actions, and third-party frameworks such as OpenClaw. Pro users get $20 in programmatic credits, Max 5x users get $100, and Max 20x users get $200, billed at API-style rates.
Greyhound Research Chief Analyst Sanchit Vir Gogia, quoted in the InfoWorld piece, called this a broader industry transition rather than an Anthropic-specific move. GitHub is moving Copilot toward a token and credit system. OpenAI has always charged usage-based pricing for API access. The era of unlimited agentic AI on a flat $20 subscription is ending. That has real implications for how teams budget. A coding agent that runs tests continuously, browses the web, and calls models recursively can burn through token budgets in ways that human prompting never did. Two large companies, ServiceNow and Uber, reportedly burned through their AI token budgets for the entire year before the year was even half over.
Standards and Protocols: WebMCP, A2A v1.2, and the Layer Cake
The most consequential announcement at I/O for the long-term shape of the web may have been the one that got the least attention. Google introduced a proposed open standard called WebMCP, short for Web Model Context Protocol.
WebMCP was originally announced as a W3C Draft Community Group Report on February 10, 2026, co-developed by Google and Microsoft. Discovered Labs explained that WebMCP is a browser-native standard that lets AI agents interact with websites through structured tools rather than screen scraping. Until now, AI agents had two ways to interact with a website. They could take screenshots and read the text visually, which is slow and brittle. Or they could call backend APIs through the original Model Context Protocol, which requires server-side cooperation that most websites do not provide.
WebMCP adds a third path. Websites publish a structured list of actions, called a Tool Contract, that the page can perform. An AI agent reads the list and calls the actions directly. There is no guessing, no clicking around blindly. A travel booking site might expose a searchFlights tool. A real-estate site might expose a filterListings tool. The protocol is exposed through navigator.modelContext in the browser, with both a declarative HTML-attribute version for static content and an imperative JavaScript version for dynamic interactions.
The reason this matters more than it might look on the surface is what it implies about how AI traffic will flow on the web. Search Engine Land's 2026 SEO predictions, quoted in the same Discovered Labs analysis, put it plainly. "In 2026, SEO becomes two jobs: driving clicks from humans and supplying clean, trusted inputs for AI agents that may never visit your site." That changes the calculus for every business that depends on web traffic. If your site cannot be parsed by an AI agent, your pricing, features, and availability become invisible to buyers researching vendors through Gemini, Claude, or ChatGPT.
Chrome 146 Canary is the only browser supporting WebMCP behind a flag as of writing. Stable Chrome support is expected around March 2026 according to industry reporting, though Google has not officially confirmed the timeline. Microsoft Edge support is expected. Firefox and Safari have not announced plans.
The other half of the agent protocol story moved forward this week as well. The Agent2Agent protocol, the open standard for getting agents from different vendors to talk to each other, kept gaining traction. According to The Next Web's coverage of Google Cloud Next 2026 in April, A2A has reached production deployment at 150 organizations, including Microsoft, AWS, Salesforce, SAP, and ServiceNow. The protocol is now governed by the Linux Foundation's Agentic AI Foundation and is at version 1.2, with cryptographically signed agent cards for domain verification.
A2A and MCP are complementary, not competing. MCP handles the connection between an agent and the tools or data sources it needs. A2A handles the conversation between two agents that might be running on completely different platforms. A Salesforce agent built on Agentforce can hand off a task to a Google Vertex AI agent, which can query a ServiceNow agent for IT asset data, all through A2A. None of the three systems needs to understand the others' internal architecture. They just need to speak the same wire protocol.
Native A2A support is now built into Google's Agent Development Kit, LangGraph, CrewAI, LlamaIndex Agents, Semantic Kernel, and AutoGen. The Agent Development Kit hit stable 1.0 across Python, Go, and Java, with TypeScript also available. The active GitHub repositories tell the story. The Java SDK, JavaScript SDK, and Python SDK all had commits as recently as May 18-19 of this week.
The protocol stack for agents is starting to look stable enough to build against. MCP for tool access. A2A for inter-agent collaboration. WebMCP for browser-based agent action. Each handles a different layer of the problem, and each is governed by an open foundation rather than a single vendor. That last point is the one to pay attention to. The protocols that won past internet platform shifts were the ones that no single company controlled.
Anthropic Crosses Over: A Quiet Inflection in Enterprise AI
One of the most significant signals of the week was a number rather than a product launch. TechTimes reported on May 15 that more US businesses paid for Anthropic's Claude than for OpenAI's ChatGPT in April 2026, according to the May 2026 Ramp AI Index. It was the first time in the AI industry's short history that Anthropic held the top position by paid business subscriptions.
The Ramp data has methodological caveats worth noting. The sample skews toward tech-forward, venture-backed companies that lean toward Anthropic's developer-first products. The index measures paid subscriptions, not usage intensity, and likely undercounts free-tier and enterprise-contract use. An OpenAI spokesperson told Axios that the Ramp methodology may underweight invoiced six- and seven-figure enterprise contracts where customers do not pay with a credit card. OpenAI has said it expects to generate more revenue than Anthropic in 2026.
That said, two pieces of evidence corroborate the trend. OpenRouter's leaderboard, which samples a different population of users, last ranked OpenAI above Anthropic in December 2025. And Anthropic's own business numbers have been growing fast enough that the company projected 10x revenue growth over a recent period and instead saw annualized growth of roughly 80x in a single quarter, according to CEO Dario Amodei's remarks at the company's May 5 financial services briefing.
The engine behind the climb has a single name. Claude Code. The product accounts for a meaningful share of Anthropic's revenue and growth, with SemiAnalysis estimating $2.5 billion in annualized run rate as of February 2026 and business subscriptions quadrupling between January 1 and the spring.
OpenAI is not standing still. On May 11, the company launched the OpenAI Deployment Company, a standalone unit backed by more than $4 billion from 19 private equity firms, consultancies, and systems integrators including TPG, Bain Capital, Goldman Sachs, Capgemini, and McKinsey. The unit will place Forward Deployed Engineers directly inside client organizations to redesign workflows around AI, borrowing the model that made Palantir successful in government and applying it to frontier AI at unprecedented scale.
Anthropic also expanded its own enterprise distribution this week and last. The company announced its acquisition of Stainless on May 18, a tooling company that builds SDKs for API providers. On May 19, Anthropic announced a strategic alliance with KPMG that integrates Claude across KPMG's core business and workforce of more than 276,000 employees. The week before, the company announced PwC was deploying Claude to build technology, execute deals, and reinvent enterprise functions for clients.
The pattern across both vendors is the same. The center of AI revenue is moving from consumer subscriptions to enterprise contracts. Both companies are racing to lock in distribution through the largest consulting and integration firms in the world. The economics of consumer AI subscriptions are not strong enough to fund frontier model training. Enterprise contracts are.
One human-interest detail from the week is worth mentioning because of what it says about talent flows. Andrej Karpathy, the OpenAI co-founder who later founded the education startup Eureka Labs, joined Anthropic's pretraining team. Anthropic also hired cybersecurity expert Chris Rohlf to join its frontier red team, which tests advanced AI models for potential security threats. Researchers move between frontier labs constantly. The Karpathy hire is notable because it lands at exactly the moment Anthropic crossed Ramp's adoption threshold.
What to Watch Next Week
A few specific things to track in the coming days. The Claude Opus 4.7 Fast mode research preview, announced in late April, is rolling out to more developers. Fast mode supports significantly faster output token generation at premium pricing for the same task quality. If you have been waiting to deploy agentic workflows that need lower latency, the waitlist is open.
The June 15 transition of Claude programmatic usage to its new credit-based meter is the next big pricing inflection. Teams running CI pipelines, scheduled automations, or long-running coding agents on Claude subscriptions should benchmark their token consumption now while the current model is still in place. Forecasting becomes much harder once usage is tied directly to token consumption rather than subscription tier.
The Antigravity 2.0 enterprise rollout is the wildcard. Google priced its top tier at $100 per month, putting it in roughly the same range as Anthropic's Max plan and OpenAI's premium tiers. The question is whether enterprise developers actually adopt it. Google has historically struggled to displace incumbent developer tools even when its underlying technology was strong. The fact that the company is killing Gemini CLI rather than letting it coexist with Antigravity suggests Google believes the parallel-agent thesis enough to force the migration.
And on the standards side, watch for WebMCP adoption signals from publishers. Search Engine Land's framing turns out to be the right one. If 2026 is the year SEO splits into a human track and an agent track, the sites that publish Tool Contracts first will own a disproportionate share of the agent track. That is a lot like the early days of structured data and schema.org markup, except the upside is bigger and the time window is shorter.
The thread running through every single one of these stories is integration over isolation. The most valuable work in AI right now is not building a better model in a vacuum. It is connecting the model to the tools, to the other agents, to the websites, to the codebases, and to the people who actually do the work. That is a less glamorous story than the next billion-parameter benchmark, and it is the one that will define the year.
A Few More Things Worth Flagging
A handful of smaller items from the week deserve a mention.
Anthropic published a piece this week explaining why Claude will remain ad-free. The argument is that advertising incentives are structurally incompatible with a genuinely helpful AI assistant. The post lays out how the company plans to expand access without compromising user trust. That is a deliberate stance in a market where every other consumer software category eventually monetizes through ads. The company seems to be betting that enterprise revenue will be large enough to fund frontier research without consumer ad dollars. That bet ties directly back to the Ramp adoption data.
The Anthropic acquisition of Stainless deserves more than a passing mention as well. Stainless is the tooling company behind the SDK generators that several large API providers use. Bringing that capability in-house gives Anthropic tighter control over the developer experience for Claude APIs across every major programming language. SDK quality is one of those things developers rarely talk about directly but always notice. If the Python SDK feels rough, you reach for the competitor. Anthropic just bought a moat in that specific dimension.
On the open source side, the parallel-agent pattern that Antigravity 2.0 showed off is starting to show up in other tools as well. Cursor shipped its Agents Window in early April. The Pragmatic Engineer survey showed developers running two and three AI tools side by side. The shape of the workflow is converging toward something like an air traffic controller view of multiple agents, each working on a different piece of the same project. The interesting question is which tool ends up being the controller versus which ones are the planes. Right now most teams use a terminal-native agent like Claude Code as the controller and use Codex or Cursor for specific subtasks. That pattern is not stable. It will likely change again before the year is out.
One more thing worth tracking. The Ramp index covered April 2026. May will be the first full month after the Anthropic policy change separating programmatic usage. If business adoption holds or grows after the credit-meter change goes into effect, it will be a meaningful signal that the new pricing model does not break enterprise adoption. If it drops, it tells us the price elasticity of agentic AI usage is higher than vendors think. That number will land in early July when Ramp publishes the next index. It is one of the few data points that cuts through marketing claims on both sides.
Resources to Go Further
The AI landscape changes fast. Here are tools and resources to help you keep pace.
Try Dremio Free - Experience agentic analytics and an Apache Iceberg-powered lakehouse. Start your free trial
Learn Agentic AI with Data - Dremio's agentic analytics features let your AI agents query and act on live data. Explore Dremio Agentic AI
Join the Community - Connect with data engineers and AI practitioners building on open standards. Join the Dremio Developer Community
Book: The 2026 Guide to AI-Assisted Development - Covers prompt engineering, agent workflows, MCP, evaluation, security, and career paths. Get it on Amazon
Book: Using AI Agents for Data Engineering and Data Analysis - A practical guide to Claude Code, Google Antigravity, OpenAI Codex, and more. Get it on Amazon
Top comments (0)