Google commits $40B to Anthropic, DeepSeek V4 undercuts frontier pricing by 30x, and agents start negotiating real deals with real money.
> The capital arms race hit a new gear — Google pledged $40B to Anthropic, Meta locked in millions of Amazon CPUs, and Cohere swallowed Aleph Alpha. Meanwhile, Anthropic quietly ran an experiment where AI agents negotiated real transactions with each other. The money is getting bigger. The agents are getting busier.
Google committed up to $40 billion to Anthropic this week — $10 billion in immediate cash at a $350 billion valuation, with another $30 billion contingent on performance targets. The deal includes 5 gigawatts of TPU capacity over five years.
This landed days after Amazon announced its own $5 billion Anthropic investment. Combined with a CoreWeave data center deal earlier this month and a Google-Broadcom partnership for 3.5 gigawatts of additional TPU access starting 2027, Anthropic has secured more compute commitments in April than most countries have planned for the decade.
The timing isn't accidental. Anthropic is reportedly eyeing an October 2026 IPO at a valuation investors peg above $800 billion. This is as much a pre-IPO positioning play as a technology bet.
For engineers, the practical consequence: Anthropic won't be compute-constrained anytime soon. Expect larger context windows, faster inference, and more aggressive Claude API pricing in Q3. When your compute partner commits 5 gigawatts — roughly the electricity demand of a mid-sized city — you don't build incrementally. You build the thing that needs all of it.
The paradox remains — Google is simultaneously competing with Anthropic on models and bankrolling Anthropic's ability to compete back. That tension is sustainable exactly as long as Google Cloud needs Anthropic as a customer anchor. The moment it doesn't, this relationship gets interesting.
Meanwhile, the rest of the field isn't standing still. OpenAI shipped GPT-5.5 the same day, billing it as "the smartest and most intuitive model" yet and pushing further toward a unified "super app" combining ChatGPT, Codex, and an AI browser. Greg Brockman called it "a real step forward towards the kind of computing that we expect in the future." DeepSeek countered with V4 — the largest open-weight model ever at 1.6 trillion parameters — at a price point that makes frontier capabilities accessible to anyone with an API key. Three major model events in 48 hours. The pace is accelerating.
Mira Murati's Thinking Machines Lab landed Soumith Chintala — co-creator of PyTorch and 11-year Meta veteran — as CTO this week. Piotr Dollar, co-author of Segment Anything (also 11 years at Meta), joined alongside him. TML has ~140 employees, a $12 billion valuation, a multibillion-dollar Google Cloud deal, and access to Nvidia GB300 chips.
This is a different kind of talent acquisition than we've seen before. Murati isn't hiring researchers to write papers — she's assembling the specific people who've built the infrastructure layers that current AI runs on. PyTorch powers most training pipelines. Segment Anything redefined vision foundation models. When the people who built the plumbing leave to start fresh, they're not just taking expertise — they're taking the institutional knowledge of what went wrong the first time.
Meta FAIR has been bleeding senior researchers for two years now. At some point, the compounding loss of institutional knowledge becomes the story, not any single departure. For hiring managers at frontier labs: if your retention strategy is "we have the best GPUs," that's no longer differentiating when every well-funded startup has GB300 access.
Anthropic ran an internal experiment called "Project Deal" this week and published the results. The setup: 69 employees each got $100 in gift cards. AI agents represented both buyers and sellers on a classified marketplace. Every deal was real — honored after the experiment ended.
Results: 186 completed transactions totaling over $4,000 across four separate marketplaces using different AI models.
Here's what makes this architecturally interesting for anyone building agent systems.
Finding 1: Model quality creates invisible advantage. Agents backed by Anthropic's most capable model consistently secured better prices than those using weaker models. The gap was material — better models negotiated lower buy prices and higher sell prices. But the humans represented by weaker agents couldn't tell. They rated their experience similarly. This is the "agent quality gap": your agent might be leaving money on the table, and you'd never know.
Finding 2: Instructions barely mattered. Participants customized their agent's negotiation strategy — aggressive, conservative, detail-oriented. It didn't meaningfully change outcomes. The model's baseline capability dominated individual instructions. Real implications for prompt engineering in transactional contexts: you might spend hours tuning system prompts that have negligible effect compared to using a more capable model.
Finding 3: The marketplace worked. 186 deals among 69 people means roughly 2.7 transactions per participant. Agents handled discovery, negotiation, pricing, and completion autonomously. The friction of buying a coworker's old keyboard dropped from "post on Slack and haggle for three days" to "tell my agent I want a keyboard under $40."
The architectural pattern matters beyond commerce:
The scary implication: we're heading toward agents transacting on behalf of humans across procurement, hiring, scheduling, and financial services. If the model powering your agent is two generations behind, you'll get systematically worse outcomes — and you won't have the information to detect it. This is an asymmetric information problem that classical economics would recognize instantly, but we have zero regulatory framework for.
Anthropic published this as research, not a product. But the 186-deal dataset is the first empirical evidence that agent-mediated commerce works at a basic level.
Consider the timeline. Six months ago, agent demos involved booking a restaurant reservation and failing half the time. Now we have agents autonomously negotiating prices across multiple rounds, with measurable outcomes that favor more capable models. The gap between "agent that can use tools" and "agent that can transact" just collapsed from theoretical to demonstrated in a 69-person study.
The next question is whether it scales — and who captures the margin when it does. If I were building a procurement platform right now, I'd be running experiments with agent-mediated vendor negotiations before Q4. The window where this feels speculative is closing fast.
NousResearch/hermes-agent — "The agent that grows with you." Gained 18,223 stars this week alone, now at 116K total. An open-source agent framework with self-evolution capabilities — the agent builds new skills from experience. The growth rate signals that developers want agent infrastructure they can extend, not just consume.
mnfst/manifest — Smart model routing for personal AI agents, claiming up to 70% cost reduction by dynamically selecting the cheapest model that meets quality thresholds per request. 5,666 stars. With DeepSeek V4 at $0.14/M tokens and GPT-5.5 at 30x that price, intelligent routing between them is no longer optimization — it's table stakes.
obscura — Headless browser purpose-built for AI agents and web scraping. 4,746 stars this week. The gap between "agent that can browse" and "agent that reliably interacts with real websites" is a browser automation problem. Obscura targets that gap with agent-native APIs instead of wrapping Playwright.
CubeSandbox — Lightweight sandboxing for AI agent code execution from Tencent Cloud. 4,083 stars. Agent safety isn't just model alignment — it's what happens when your agent runs rm -rf in a real environment. CubeSandbox provides instant, concurrent isolation with minimal overhead.
Three numbers from this week tell the whole story: $40 billion (Google's Anthropic check), $0.14 (DeepSeek V4 Flash per million input tokens), and 186 (real deals closed by AI agents). Capital concentration accelerates at the top while the cost of running these models collapses at the bottom. That compression is the story of 2026.
Anthropic's agent marketplace experiment isn't a product launch — it's a preview of the economic layer that sits on top of all this infrastructure. The companies deploying billions in compute are building the rails. The agents are about to start using them to transact autonomously. If you're building anything that touches commerce, procurement, or negotiation, the window to understand agent-mediated economics just opened. It won't stay theoretical for long.
— Aaron, from the terminal. Back next Friday.
Compare three approaches to AI agent browser automation. Browser Use, Stagehand, and Playwright MCP tested with code examples, benchmarks, and architecture trade-offs.
AI EngineeringHow OpenClaw routes messages across Discord, Telegram, and Slack with an 8-tier priority cascade, then isolates agent execution in pluggable Docker/SSH sandboxes.
AI EngineeringSide-by-side comparison of how OpenClaw and Hermes Agent build system prompts, manage token budgets, and compress long conversations without losing critical context.
AI Engineering