Quanta Bits - May 16, 2026

Hi everyone,

It feels like Anthropic vs. OpenAI is entering a MySpace vs. Facebook moment. Except that OpenAI was the first one with the big splash and Anthropic was the one coming hot, gaining ground by focusing on enterprise customers.

Anthropic for a while seemed unable to do anything wrong. They focused on Enterprise customers and catered to developers and churned product features quickly. Then they got into a rut and recently it seems like they can't get out of their own way. Capacity shortages and confusing announcements about pricing changes and quality concerns seem to be alienating the very customers that drove their growth in recent months. A lot of folks, including myself a big Anthropic fan, are giving Codex GPT5.5 a serious try.

I still want Anthropic to stay strong. We need real competition here. But this shift made something clear: once AI becomes part of how work gets done, the product is not just the model. It is availability, cost, speed, integration, and whether the workflow still works next week.

Cheers.
Reza

Agent adoption is becoming an operating model problem. The repeated pattern this week was ownership: who approves agents, who gives them access, who measures results, and who cleans up when workflows cross systems. CIO, VentureBeat, and Glean all pointed the same way.

The labor story is splitting, not simplifying. The Economist found sharper employment declines in AI-exposed graduate fields, while Jensen Huang pointed toward demand around data centers, chips, power, and advanced manufacturing in his Carnegie Mellon commencement. Companies need to redesign first-pass analysis and apprenticeship around what AI now does cheaply and what still needs judgment.

I felt this shift before I understood the pricing mechanics.

For a while, my operating layer was Spock running through OpenClaw and Claude. Research, notes, content planning, task cleanup, and small automation chores moved through an agent layer instead of sitting in the "I'll get to it later" pile.

Then Anthropic changed how it handled third-party access, and overnight, the setup felt more expensive and less dependable. I struggled to find an alternative.

So I started using Codex more heavily. It felt faster, more predictable, and easier to steer. That is not really a product review. It is the larger operating question in miniature: when AI becomes part of work, the product is speed, availability, cost, integration, and whether the workflow still works next week.

Anthropic made that boundary visible this week. Starting June 15, 2026, Agent SDK usage and the claude -p command move into a separate monthly credit pool: $20 for Pro, $100 for Max 5x, and $200 for Max 20x. Interactive Claude Code, Claude Cowork, and normal Claude conversations stay in the regular subscription pool. Shared production automation belongs on the Claude Developer Platform with an API key. (Anthropic Help Center)

Confusing? You betcha.

The business distinction is simple: a person using Claude Code interactively is assisted work. A workflow calling Claude in the background is infrastructure.

A sales rep asking AI to research one account is assisted work. An agent that scans opportunities each morning, checks CRM gaps, drafts follow-ups, and flags forecast risk is infrastructure. It needs permissions, logs, exception handling, human review, a cost model, and someone who owns the result.

That is where the cheap builder story changes. The prototype can feel almost free inside a subscription product. The recurring workflow does not stay free when it becomes a programmatic workload across CRM, Slack, email, file storage, finance systems, and custom apps.

I do not read Anthropic's change as a one-company pricing story. I read it as a sign that the true cost of AI work is still settling.

Anthropic is adding capacity as fast as it can, including a SpaceX compute deal with more than 300 megawatts of capacity and more than 220,000 NVIDIA GPUs. (Anthropic) The Economist reported the broader squeeze: OpenRouter's weekly token volume quadrupled between January and March, and Anthropic throttled access at busy times.

McKinsey describes the same bottleneck from the industrial side: global data center spending could reach $7 trillion by 2030, while long-lead equipment such as switchgear and transformers is already stretched. (McKinsey)

The operating point is not "AI is too expensive." It is that capacity and costs are not stable enough to assume every labor-replacement business case will hold.

The Financial Times reported on May 12 that Amazon employees used an internal agent tool on non-essential tasks to inflate token consumption on leaderboards. Once tokens become the metric, people produce tokens.

BCG's March 2026 advice is more useful: agentic AI fits best where work is complex enough to justify it, risk is manageable, and the business case has metrics, timelines, and P&L impact. (BCG)

Do not ask whether people are using AI more. Ask which workflow improved, what it cost, what review it required, and whether the result was worth the complexity.

Not every task deserves the most expensive model. Some work should run on a frontier model. Some should run on a cheaper model. Some may be better handled by a smaller, specialized model.

VentureBeat reported that IBM built a model gateway so enterprises can switch between models while preserving observability and governance. Microsoft Foundry's model leaderboard points the same way: teams compare models by quality, safety, estimated cost, and throughput.

That is the practical answer. Use the expensive model for judgment-heavy synthesis, complex coding, and ambiguous research. Use cheaper or specialized models for classification, extraction, first-pass routing, and repetitive internal chores.

The point is the costs keep changing and we as users are not in control. Not yet. We need competition. But we also need some stabilization in capacity and market to gauge better the future ROI of our investments in AI.

I still believe in the build-it-yourself moment. My Newsletter Sifter (see below) exists because AI made it possible to build a small local tool around my own reading friction.

But building your own thing with AI is not the same as owning the economics underneath it. In a recent arXiv paper on buy-versus-build decisions, David Klotz argues that the "make" option still depends on external AI infrastructure. You may own the app. You may not own the cost curve. (arXiv)

If a person is using AI to think, draft, research, code, or clean up an annoying task, encourage it.

If the workflow is recurring, shared, background, or writing back into business systems, treat it like infrastructure. Before it scales, answer:

The build-it-yourself era is not ending. It is growing up. That is not a reason to stop building. It is a reason to build with clearer eyes.

Ramp's May 2026 AI Index has Anthropic passing OpenAI in paid business adoption for the first time: Anthropic at 34.4%, OpenAI at 32.3%, and overall AI adoption at 50.6%. Enterprise AI spend is moving faster than normal software vendor gravity should allow.

OpenAI puts Codex in your pocket. The Verge reports that OpenAI is rolling Codex into the ChatGPT mobile app. The point is cross-device supervision: review outputs, approve commands, inspect terminal output, and unblock a desktop agent from your phone.

Princeton brings back proctors after 133 years. The Daily Princetonian reports that Princeton will require proctors at in-person exams starting July 1. When AI makes convincing work easy to generate and hard for peers to detect, trust-based systems need new verification layers.

Legal agents move into the workflow, not beside it. Microsoft is launching Legal Agent inside Word, while Anthropic launched Claude for the Legal Industry with Microsoft 365 integration. The signal is AI packaged inside the systems where high-trust work already happens.

Taylor Swift turns identity into an IP strategy. The Los Angeles Times reports that Swift filed trademark applications for clips of her voice and an Eras Tour image. AI impersonation is becoming an IP and brand-protection problem.

Claude Security moves into the remediation loop. Anthropic moved Claude Security into public beta for Claude Enterprise customers, with GitHub scanning, finding validation, patch instructions, and Jira and Slack handoffs. Security AI is moving into workflow.

Human brain cells learned to play Doom. The Economist reports that Cortical Labs taught 200,000 human brain cells on a silicon chip to play Doom. The point is energy: a human brain runs on about 20 watts. What if the most efficient compute architecture already exists in biology?

HBR - "Why You Shouldn't Treat AI Agents Like Employees". In a study of HR and finance managers, people caught fewer errors and felt less accountable when AI was framed as an employee rather than a tool. Useful reminder: the metaphor changes the behavior.

Andrej Karpathy on moving from vibe coding to agentic engineering. Useful distinction: vibe coding raises the floor; agentic engineering preserves quality.

Ravi Mehta on context engineering for AI prototyping. Practical context engineering: functional context, visual context, and data context.

The Rundown AI interview with UiPath CMO Michael Atalla. Pilots fail when isolated from governed workflows.

Teaching Claude why. Models improved when training examples taught the reasoning behind the boundary, not just the approved action.

GPT-5.5 vs Claude vs Gemini. Not gospel, but close to what I am feeling: GPT-5.5 for fast execution, Opus for deeper judgment.

Foster by Claire Keegan, 2010, 101 pages. ★★★★★
The Quiet Girl / An Cailín Ciúin, directed by Colm Bairéad, 2022, 95 minutes. ★★★★☆

I watched the movie first and then read the book, and both are exceptional. Foster is the story of a young girl from a neglectful, overcrowded family who is sent to stay with distant relatives, the Kinsellas, for the summer. In their quiet farmhouse in rural Ireland, she discovers something she has never known: attention, warmth, the feeling of being wanted.

What unfolds is not dramatic in the conventional sense. There are no villains, no twists. Just the slow realization that kindness can reshape a child, and that having to leave it behind is its own kind of loss.

Keegan's prose is stripped down to the bone. At 101 pages, there is not a wasted word. The movie, filmed entirely in Irish, is equally restrained: long silences, a child actor who communicates more with her eyes than most actors do with dialogue, and a final scene that left me wrecked.

The book gives you the girl's inner voice. The movie gives you her silence. Together, they are devastatingly good.

I built a local Newsletter Sifter because reading 10+ AI and business newsletters one email at a time had become its own chore. It started as a Claude Code prototype, then moved into Codex/GPT-5.5 as a working local app. Claude Design helped reshape the reading experience.

The app syncs newsletters, breaks them into clips, and sorts items into Read, Skim, and Skip based on relevance, duplicates, past feedback, and my interests.

Small thing, but useful pattern: AI is very good at turning recurring friction into a small operating system for attention.

Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.

I collaborate with Spock, my AI agent. He researches extensively: scanning, filtering, and surfacing what's relevant across my business. I read, listen, and watch what resonates, and decide what matters. I provide direction, we draft together. The editorial judgment is mine. He'd tell you the same. Most logical. 🖖