How to Compete With Anthropic

You don’t.

Most startups building on Claude think they’re in a race. They’re in a parade. Float belongs to someone else.

April 4, Anthropic cuts off Claude subscription access for third-party agent frameworks. April 8, they ship Managed Agents. Eight cents per session-hour. Notion, Rakuten, Asana onboard. April 13, someone notices you can build full-stack apps inside Claude. Lovable-like. No code.

Three moves in nine days.

Eric Simons built Bolt.new to $40M ARR in five months, 45 employees, zero marketing spend. Was at an Anthropic dinner last fall, seated next to Lovable’s CTO. Joked to the Anthropic people: “When are you gonna put me and this guy out of business?” Their faces said it all.

“The easy wins are officially over.”

He’s being gracious.

Platform starts generous. APIs open. Support responsive. Dev rel people remember your name. “We love builders!”

Platform hits critical mass. Pricing tightens. Rate limits appear.

Platform ships the thing you built.

Startups that built on “Anthropic hasn’t done this yet” confused absence with intention. If your moat is the platform’s absent feature, you have a countdown.

Customers don’t experience this as Anthropic eating their vendor. One fewer integration to manage. Customer chose Anthropic’s API first. Every decision after that flows downhill. Adding Managed Agents inside an existing Claude deployment costs zero integration effort, zero procurement, zero security review. Buying that from a startup costs all three plus a second vendor relationship someone has to manage quarterly.

Startup isn’t competing on product. It’s competing against the sequence of commitments the customer already made.

Claude Code half works. You load credits, pull the lever. Sometimes magic, sometimes garbage. Pull again. BF Skinner studied this in the 1950s with pigeons and rats. Variable reward schedules create the strongest behavioral persistence. You don’t get addicted to things that always work. You get addicted to things that sometimes work brilliantly.

Your codebase ends up on Claude. Your muscle memory. Your team’s daily habits. The question shifts from “should I build on this platform?” to “how do I build more on this?”

Founders know about platform risk. They’ve read the threads. AWS ate hosting startups. Stripe ate payment wrappers. Salesforce ate every CRM-adjacent tool that forgot who owned distribution. Still build deeper. Variable rewards recruit attention before the analytical brain catches up.

If this were about the model, Google would have won already. Best researchers. Infinite compute. Three billion daily users on Search alone. They couldn’t pull it together. Model was never the thing.

Anthropic got the product experience right. The habit loop. Claude’s system card reads like a company trying to align something that’s already too sticky to stop using. They keep publishing alignment research and safety reports because the model plus the harness plus the experience creates a concoction that’s genuinely addictive. Same model inside Cursor feels completely different from the same model inside Claude Code. Same engine, completely different car. Google has the engine. Anthropic built the car people can’t stop driving.

That’s what makes the venus flytrap work. Variable rewards delivered through a harness that remembers your context, learns your patterns, fits your workflow like a glove that tightens one finger at a time. The model is the nectar. The harness is the jaw.

Thirty days. That’s roughly the window. After that you’re rationalizing.

Models went from novel to commodity in about three years. Orchestration went from custom-built to commodity in nine months. Nine months ago orchestration was a venture-backable business. Now it’s a line item on an API bill.

Eight cents per session-hour. Anthropic commoditizing the harness layer on purpose. At eight cents, the math of building your own orchestration stack stops working. Every startup that could have competed on price just lost. Price will not stay at eight cents. The dependency will.

What becomes possible now that orchestration is cheap? What was uneconomical when it cost real engineering time? Those things nobody’s built yet.

Two kinds of knowledge.

Parallel. You acquire it by hiring smart people working simultaneously. Building an orchestration layer is parallel knowledge. Anthropic hired engineers. Built it. Sprint complete.

Sequential. Can only be acquired through experience accumulated in order. Failure after failure in a specific domain, each one teaching you something the next one builds on. No amount of money compresses this. Knowing which compliance edge cases trigger false positives in which regulatory jurisdictions. How a specific customer’s CFO wants data formatted down to the column headers. Which legacy COBOL patterns break which modernization tools and why the fix is never what the documentation says.

Is your core asset parallel or sequential. If parallel, you’re a feature on the platform’s roadmap.

Bun passed the accumulation test. Anthropic acquired them for $900M because replicating what Bun had would take years. Not the code. The accumulated operational knowledge embedded in the code.

Where in your value chain would removing you break the customer’s operation. Not inconvenience. Break. If the answer is nowhere, you’re a feature.

Simons said it at the dinner table before Anthropic said it with code. Go deep on specific workflows with high retention. Raise enough to be dangerous.

Anthropic will build the generic version of everything. They will never build the version that knows your customer’s regulatory filing deadlines, their specific data schemas, their CFO’s reporting format. Market’s too small. I’m still working out how far that holds when the platform starts acquiring vertical players too.

Every prompt you send to Claude teaches Claude about your domain. Every context window packed with proprietary data becomes signal for their system.

Goldman embedded Anthropic engineers inside the bank for six months. Six months of learning how accounting reconciliation works at Goldman. How compliance flags get triaged. Which edge cases the automated systems miss. That knowledge won’t stay at Goldman. Can’t. Goldman paid for the education. Anthropic keeps the degree.

Every API call is a lesson. Every context window is a briefing.

Open-weight models caught up. Qwen3-Coder-Next. 80B total parameters, 3B active per inference step. Runs quantized on a Mac with 64GB unified memory. 70.6% on SWE-bench Verified. Claude Sonnet 4.6 scores 79.6%. Gap is closing fast and for most production coding tasks the difference is negligible. GLM-5 for reasoning. DeepSeek V3.2 for heavy-context work. Apache 2.0. Own the weights. Nobody turns them off. Nobody changes the terms of service on a Saturday.

Mem0 self-hosted. Sits between your app and the model. Memory stays on your infrastructure.

OpenClaw for coding agents. Or build your own pointed at your local model through the same OpenAI-compatible API that vLLM exposes. Harness is where sequential knowledge lives. Part you never outsource.

MCP is open standard. Same protocol whether the model is Claude or Qwen or DeepSeek. Swappable.

Contabo VPS. Hetzner bare metal. Mac Studios in a closet. Cost of one month of Claude API at enterprise scale buys an H100 that runs for three years. Air-gapped if you want. No data leaving your network.

Anthropic’s moat is the accumulation of context. Every conversation, every workflow pattern flowing through their infrastructure teaches their system. Self-host and you cut off the intelligence supply.

Open strategy. Publish the whole architecture, it gets stronger.

Can’t fight it by making Claude better. Open community replicates within months. Qwen3 is already there on coding benchmarks. Can’t fight it by making Claude cheaper. Self-hosted is 10x cheaper at moderate scale. vLLM is a few commands

Only weapon left is the venus flytrap. Variable rewards. Magic moments.

Startups spent their own money, their own time, took all the risk to figure out which AI product categories actually work. Bolt proved vibe coding is a real market at $40M ARR. Lovable proved it scales to $300M. Hundreds of agent harness startups proved orchestration is valuable.

Anthropic watched all of this happen on their API. Saw the usage data. Saw which categories grew. Saw what customers actually paid for.

Then built it themselves.

Startups ran the experiments. Anthropic collected the results. Reconnaissance was free. Startups paid for the mission. Anthropic got the intelligence report.

Commoditize the model with open weights. Keep context and memory on your own infra. Build the harness yourself.

You’re somewhere around day 900. vLLM is a few commands. Mem0 is an afternoon. Architecture decision takes less time than the meeting where you’d discuss it.

Platform tend to gobble up innovation done on top of them.

Open stack is one way to not lose it.