I’m a builder. I’ve spent the last year building on top of Claude — wiring it into my workflows, my custom interfaces, my attempts to make AI usable on my terms. I love Claude as a model. I love Anthropic as a company. And I’m worried they’re starting to do the thing that Google did.
I worked at Google twice — once as an intern in 2014, when the culture was still mostly intact, and again as a developer relations engineer in Boulder from 2021 until the 2023 layoffs. The version of Google I returned to was already fragmenting. TGIF, the weekly all-hands where Larry and Sergey would actually tell us what was happening, had quietly turned into a monthly corporate readout. “Don’t be evil” — once a core guiding principle — had been demoted to a line buried in the code of conduct. And the product surface I worked on — Google Cloud — was fragmented in ways that mirrored the company itself: near-duplicate libraries, each owned by a team that didn’t talk to the others, each optimizing for its own metrics. The fragmentation in the product was downstream of the fragmentation in the communication.
What I came to see, and later wrote about in my master’s thesis Radical Reorganization, is that organizations succeed in changing times when they recognize themselves as living, interconnected wholes — and struggle when they let each part optimize in isolation. The decisions that lead to fragmentation are usually individually reasonable. The cumulative effect is what kills you. By the time the warning signs show up in the product or the org chart, the drift is usually well underway.
Anthropic is not Google. Not yet. But I’m watching them ship product decisions over the last six months that pattern-match closely to the early stages of what I saw at Google. And I think they can still course-correct. So I want to write this — to Anthropic, and to the community of builders who love Claude and have a stake in how this plays out.
Let me start with what I’m seeing on the ground, because the diagnosis matters before the prescription.
Claude is currently split across product surfaces in ways that make no sense to me as a user. Chat lives everywhere — web, desktop, mobile. Code lives in the terminal, in Claude Desktop, and now on the web. Cowork lives only on Claude Desktop — not web, not mobile, despite the fact that knowledge work happens across all those places. Design lives only on the web at claude.ai/design — not desktop, despite most design work happening alongside files on your computer. As a builder using these, I keep context-switching between products that are clearly running the same underlying model with different system prompts and slightly different tools surfaced. The proliferation isn’t capability — it’s an org chart leaking into my workflow. What I want is one Claude with surfaces that share my context, my memory, my tools, and pick up where I left off across whatever device I’m on. What I have is a matrix of products I have to mentally route between depending on what I’m doing and what device I’m using. The fragmentation has gotten bad enough that I’m building Parachute — an open, local-first memory layer that lets any AI read and write to the same knowledge graph — partly because nobody else is solving it.
Worse than the product fragmentation is the billing fragmentation, which is newer and more aggressive. Anthropic’s stated mission is to build AI that serves humanity’s long-term well-being — which I take to mean, among other things, that humans should be able to weave Claude deeply into how they think, work, and build. Until recently, you could. Pay $200/month for Claude Max and use Claude however you wanted — through chat, through the terminal, through the Agent SDK, through your own custom interfaces. That last category was the magic part. The Agent SDK was Anthropic opening the door to builders making their own things: Telegram bots, Slack agents, custom dashboards, home automation. Tools like OpenClaw took that promise and ran with it — an open-source harness that wrapped Claude into whatever surface you wanted, routing through your existing subscription. By early 2026, OpenClaw alone reportedly had over 100,000 active instances. This was personal computing for AI starting to take shape — not Claude as a destination you visit, but Claude as a capability you weave into wherever you already work.
Then on April 4, 2026, Anthropic pulled the rug. They had a real problem to solve: a $200/month subscriber piping requests through OpenClaw could easily consume thousands of dollars of tokens at API-equivalent prices, and the unit economics of flat-rate subscriptions don’t survive that pattern at scale. Their concern was legitimate. Their solution was a blunt instrument. Claude Code started silently detecting third-party harnesses by scanning git status and pattern-matching on strings. One user got hit with $200.98 in surprise API charges because the string “hermes.md” appeared in a recent git commit — they hadn’t even been using Hermes, but the detector fired and quietly switched their session to API billing while 86% of their subscription quota sat unused. After the case went viral, Anthropic refunded it.
Five weeks later they announced a partial walk-back — sort of. Starting June 15, 2026, subscribers get a separate Agent SDK credit pool ($20 for Pro, $100 for Max 5x, $200 for Max 20x) for use with claude -p, the Agent SDK, GitHub Actions, and third-party harnesses, billed at full API rates and non-rollover. In effect, your Claude subscription now comes with a tiny prepaid API account on the side. OpenClaw technically works again, but the personal computing it enabled — running Claude however and wherever you wanted, under your subscription — is functionally over.
Here’s the thing I want to name clearly, because I think it matters: Anthropic isn’t wrong that flat-rate subscriptions can’t absorb unlimited automation. They’re wrong about which axis to slice along. Interactive vs. programmatic isn’t where the abuse actually is — a developer running Claude Code interactively for twelve hours a day can produce more tokens than someone running an Agent SDK script for personal use. The hobbyist with their own Telegram interface gets penalized as if they were running a production workload, while heavy interactive users sail past. The distinction that matters isn’t how the tokens are invoked. It’s who’s at the end of the token stream. One human consuming the output is personal use. A product where many other humans consume the output is commercial use. The current architecture gets the axis wrong, which is why the implementation feels capricious even when the underlying concern is legitimate.
And there’s a deeper question underneath the axis problem: where the efficiency gains go. Anthropic, like every frontier lab, is getting real efficiency gains every quarter — better hardware, quantization, smarter serving. Right now those gains are being plowed almost entirely into next-generation capability. But they could also be spent on giving subscribers more room in the model they already have. The personal-vs-commercial framing doesn’t just give Anthropic a cleaner axis to slice on — it gives them a place to direct the efficiency dividend. Restriction isn’t the only response to “subscribers consume more than $200 of API-equivalent compute.” So is “let’s make $200 of compute go further.” Both can be true. Right now only one of them is being pursued.
The result is that builders are migrating. Not all of us, not yet — I haven’t gone, because I still prefer Claude as a model. But many builders I talk to are at least running Codex now, and many have switched primary loyalty. Claude Code is the better tool. Codex is the better posture. Builders are choosing posture.
Three things, all of them obvious once you start using it.
The harness is open source. Apache-2.0, 75,000+ stars, 428 contributors, a $1 million grant program from OpenAI for developers building on top of it. When something’s confusing, I can read the source. When something’s broken, I can patch it. When I want to extend, I can fork. Building on Claude Code, by contrast, I’m reverse-engineering from logs — the repo at github.com/anthropics/claude-code contains plugins and examples, but the orchestration code is proprietary, marked “All rights reserved.” Demand for an open version is so high that when the full source leaked on March 31, 2026, a community rewrite called claw-code hit 100,000 GitHub stars in 24 hours — the fastest-growing repo in GitHub history. Anthropic responded with DMCA takedowns.
It’s one tool, not five. Codex is a single interface where I write, I plan, I code, I generate images. Increasingly I see people using it as their general chat interface, replacing ChatGPT. The architectural insight is that this isn’t actually hard — most of what differentiates Anthropic’s separate products is system prompts and tool availability, not fundamentally different machinery. Codex has the discipline of letting that be true at the product layer too.
It works anywhere. I can call Codex from inside Claude Code when I need to generate images, and it draws from the same ChatGPT subscription I’m already paying for. OpenClaw now offers Codex as a backend option. The model goes where the builder is. Compare this to Claude, which increasingly forces builders into specific surfaces with specific billing rules and specific friction.
To be clear: I’m not arguing the open-weights question here. That’s a longer conversation with real considerations on both sides, and this isn’t the article for it. What I’m arguing is much narrower: the harness isn’t the model. The harness is TypeScript that calls an API. When it runs, Anthropic still sees every request and can enforce every policy. Open-sourcing the orchestration layer changes nothing about the safety properties of the model — it just lets builders inspect, fork, and extend the orchestration the way the rest of modern software works.
I want to name something that sits underneath all of the above. The entire AI field is currently operating under one shared assumption about what progress means: bigger models, more capabilities, more tokens per task, more reasoning per query, more everything per generation. The labs are spending nearly every available efficiency gain on the next training run.
I’m not arguing the labs should stop training new models. Sonnet 5, whenever it lands, will almost certainly be both more capable and more efficient than 4.6 — capability and efficiency tend to land together each generation. The capability frontier is going to keep moving forward regardless of whether Anthropic also shifts priority toward the surrounding ecosystem.
The point is that for the vast majority of what builders are actually trying to do, Sonnet 4.6 is already extraordinary — and the trajectory means the next generation will be even better-suited for everyday work, not just for pushing the frontier. There’s a related issue worth naming: many builders reflexively reach for the most capable model available for every task, and increasingly that isn’t necessary. Sonnet 4.6 handles the bulk of real agent work — and much of what people reflexively route through Opus is just as well served by Sonnet. The bottleneck for most of us isn’t model capability. It’s what’s around the model — the tooling, the harness, the surfaces, the flexibility to compose Claude into our workflows.
Meanwhile, the efficiency gains every lab is getting are real and compounding — roughly 4× per year in cost-per-token for serving the same capability, from hardware generations, quantization, MoE sparsity, and better serving stacks. The field is spending almost all of that dividend on bigger, more capable next-generation models. You can spend it the other way too. Same Sonnet 4.6, but with an open harness the community can extend; with one unified product surface instead of five; with a token budget that doesn’t fight you. The capability is already there. The leverage is in building a better system around it. And this happens to be what Anthropic’s mission specifically calls for — “build AI that serves humanity’s long-term well-being” asks for the depth and breadth of how AI gets into people’s lives, not just the ceiling of what AI can do at the frontier.
Here’s the announcement I want to read from Anthropic: We’re going to keep training models. We’re going to keep pushing the capability frontier. But for the next eighteen months, we’re also going to invest seriously in the layer around the model — opening up the harness, unifying the product surfaces, expanding what subscribers can do with the capability we’ve already shipped. They have the room to do this. They have the brand permission, because they are the safety-and-mission lab. They’re uniquely positioned to lead the field on something other than vertical scaling alone.
Three changes. None of them are technically hard. All of them are decisions.
Merge the product surfaces into one Claude. Claude Code, Cowork, Design, and Chat should be one product with different system prompts and different tools surfaced based on what you’re doing. The user should pick the task, not the product. This is mostly a UX and packaging change; the underlying machinery is already shared. The fragmentation in the product portfolio isn’t capability — it’s an org chart leaking into the user experience.
Open-source the harness under a permissive license. Apache-2.0, the same license Codex uses. The model stays closed. The orchestration layer becomes inspectable, forkable, extensible. The community has already prototyped what this looks like — claw-code did it involuntarily and got 100K stars for the effort. The only thing the proprietary harness license is currently doing is preventing Anthropic from capturing the ecosystem energy that’s already organized itself.
Unify the token budget around personal vs commercial. One subscription, one budget, fungible across chat, terminal, Agent SDK, custom harnesses, third-party tools, your own Telegram bot. The line is drawn at redistribution — not at incorporation, not at monetization, not at scale. A founder building a pre-revenue startup with Claude Code is a personal user; the same founder shipping their app and routing customer queries through Claude is commercial, and those queries belong on the API. This actually addresses the unit economics problem that drove the April restriction: at-scale automation where Claude is being resold goes to the API where token-based billing makes sense. What it stops doing is punishing the hobbyist with a Telegram bot or the researcher with a Python script. This is how Figma, JetBrains, and most other software is already licensed.
And one structural thing alongside the three: publish the actual token allotments. Stop hiding what $20 and $100 and $200 buy behind “hours of usage” and “rate limits may vary.” The opacity isn’t protecting builders. It’s protecting Anthropic’s option to quietly adjust the ceiling without it reading as a price change. That option is worth less than the trust it’s costing.
This is the part I want to write directly to those of us who love Claude.
We have more leverage than we realize. When Anthropic quietly removed Claude Code from the Pro plan on April 21, they reversed it within 24 hours because the developer community made enough noise. The Hermes.md billing bug got refunded because the case went viral. Anthropic listens to its users in ways most companies don’t — but they can only act on what they actually hear. The question isn’t whether we can influence them. The question is whether we use that influence well.
Give specific feedback, especially when friction bites. “More tokens” isn’t actionable. “Publish exact token allotments per tier” is. “Eliminate the interactive/programmatic distinction and bill on personal vs commercial” is. “Open-source the harness under Apache-2.0” is. When the Agent SDK credit pool runs out on something that’s clearly personal use, say so publicly. When a third-party harness gets blocked, ask why specifically. When a billing classification feels wrong, write about it. The goal isn’t to complain. It’s to make the friction visible to a company whose people genuinely want to do right by their users but are operating with imperfect information about how their rules actually land.
Support open community efforts as a signal, not as defection. Star claw-code. Contribute to OpenClaude. Run open-weight models on inference providers and write about what works. None of this is about leaving Claude. It’s about showing Anthropic that demand for an open ecosystem is real and the community is willing to put energy behind it. The clearer the signal, the easier it is for the people inside Anthropic who want to make the case for openness to make it.
Be loud when Anthropic moves in the right direction. This matters more than the criticism. If they ship a unified surface, amplify it. If they relax a restriction, thank them. If they publish concrete allotments, celebrate that. Negative feedback dominates because it’s easier to write; positive feedback compounds because it tells the people inside the company that the changes they fought for were worth fighting for. The path to course-correction runs through the Anthropic employees who already want these things to happen. Make their internal case for them.
Don’t migrate silently. If you’re shifting work to Codex because the friction with Claude has gotten too high, say why. Specifically, publicly, in a way someone at Anthropic could read. A silent migration is a data point Anthropic can’t act on. A public migration with specific reasons is feedback they can use.
When I left Google, the soul wasn’t gone — Google is full of brilliant, caring people doing meaningful work, and there’s still real soul there. But it was being oppressed by a system that had stopped serving it. The artifacts that used to channel that soul — TGIF, 20% time, “Don’t be evil” — were still nominally around, but each had been hollowed out one decision at a time. By the time anyone noticed the dinosaur outside the Googleplex was a warning, the patterns that erode the soul had been entrenched for years. The fragmentation in the product was downstream of fragmentation in the company, and the fragmentation in the company was downstream of decisions the leadership had stopped questioning.
Anthropic is not at that point. They’re at the much earlier stage where the symptoms are subtle and easy to dismiss. The product fragmentation is starting. The billing fragmentation is starting. The trust between the company and its developer community is fraying at specific, identifiable places. The mission is still present in the rhetoric. The question is whether it stays present in the actual decisions.
The window for course-correction is widest at the beginning, when the patterns are just starting to show. That window is open right now. It won’t stay open forever. And the people who can keep it open longest are the builders who care enough to be specific, loud, and present about what kind of company they want Anthropic to be — and who refuse to let the convenience of silent migration substitute for the harder work of telling Anthropic, directly, what’s wrong and what we want instead.
I want them to choose well. I want us, as a community, to help them choose well. Not by demanding more or accepting less, but by being honest about what we’re seeing, specific about what we want, and present when it matters.
Let’s make sure they don’t lose their soul along the way.