The Replit–Anthropic Split Has Four Data Points. I Read All of Them.

Four. That is the total number of discrete facts contained in the report that triggered this week's cycle of commentary about Replit ending its use of Anthropic's models. No publication timestamp. No named author. No named replacement vendor. No statement from Anthropic. The piece ran on a crypto outlet covering a subject with no on-chain component whatsoever, and it leans on a single source: Replit's chief executive, speaking on a platform where he has an audience and a commercial interest. Trace the story back and the provenance is worse than the content. A crypto publication, covering an AI vendor dispute, citing a social post, with no date stamp and no second source. In my line of work we call that a dangling reference. You cannot audit a claim with no block height, no timestamp, and no counterparty confirmation, and you certainly cannot build a position on it. Tracing the ghost funds from the genesis block is only useful when there is a genesis block. Here there is a screenshot. I have spent most of my career building dashboards that people use to check whether I am lying to them. The rule is simple. If I cannot reproduce it, I do not publish it. So treat what follows as an audit rather than a column. Start with structure, because structure is the entire story. Replit sells an agentic coding product: a browser environment where an autonomous loop writes, executes, and deploys code. That loop consumes tokens at a rate that would have looked absurd in 2023. A multi-step session with long context and repeated tool calls burns tens to hundreds of thousands of tokens per task. At flagship pricing in the three-dollar input, fifteen-dollar output range per million tokens, one hard task costs somewhere between ten cents and more than a dollar in raw inference. Replit's entry subscription sits near twenty dollars a month, agent usage included. Do the arithmetic. A subscriber who leans on the agent for an hour a day does not need to be malicious to become unprofitable. They need to be enthusiastic. That is the frame. Not capability. Not alignment. Cost of goods sold. The second structural fact is stated only implicitly in the source: Anthropic sells the model and also ships Claude Code, a competing agentic coding product. Replit was therefore paying a supplier that was simultaneously bidding for Replit's customers. In supply-chain terms this is not a partnership. It is a hostage arrangement with an invoice attached. I have watched this pattern in crypto for years. The RPC provider that also runs a competing indexer. The oracle network that also operates a lending market. The moment your vendor's roadmap includes your product, the vendor relationship has become a governance problem, and no service-level agreement survives that. Here is where the source material stops and the engineering starts. Multi-model routing is not a procurement decision. It is a re-platforming. Every model has a different tokenizer, a different tolerance for long context, a different failure mode when a tool call returns malformed JSON. Moving an agent from one frontier model to another means rebuilding prompt scaffolds, re-running eval baselines, re-tuning retry logic, and re-measuring latency at every hop. The switching cost is measured in engineer-months, not purchase orders. A company does not do this because a competitor shipped a product. It does this because the unit economics stopped closing. I have audited this exact failure mode before. In 2017, working as a junior engineer in Tokyo, I reviewed fifteen pre-sale contracts for a boutique security firm and found a reentrancy path in one of them before the public launch. The team's response was instructive. They did not argue with the finding. They argued with the timing. Nobody disputes a balance sheet when the numbers are already on the screen. That is why I insist on reading the deployment, not the announcement. An announcement is a statement of intent. A migration is a commit history. Now the routing problem itself, and I want to be precise, because the analogy is usually drawn badly. Anyone proposing that multi-model routing solves vendor dependency should study the Lightning Network first. Lightning has been routing payments for seven years, and its routing layer still fails often enough that most users never notice, because they never route. They custody with a large hub and treat the network as a settlement rail with extra steps. The lesson is not that routing is impossible. The lesson is that routing works well at low volume and low value, and degrades exactly where you need it most: at the tail, under load, with time pressure. An agentic coding session is a tail event by construction. It is long, expensive, stateful, and unforgiving of a mid-loop failure. Routing between three providers does not eliminate that risk. It distributes the risk, and distributes the debugging burden along with it. Failover that works in a status-page incident review rarely works in a forty-step agent loop that has already written eleven files. The oracle parallel is the one I keep returning to. DeFi spent five years learning that a price feed is only as trustworthy as the incentive structure of the parties operating it. Chainlink answered the decentralization question by adding node operators, then concentrated the meaningful work in a small set of them and called the result robust. Latency was never the real problem. The real problem was that the entity supplying the number held a position in the outcome. When the oracle bleeds, the chain holds the knife, and the chain is holding it because the oracle handed it over. A model provider that competes with its own customers is an oracle with a position. Which brings me to the part of the discourse I find least defensible: the rush to declare multi-model routing a new infrastructure category. I have watched this movie in crypto. Data availability layers were pitched as the indispensable substrate for a rollup future. The demand curve never arrived at the projected scale, because most rollups do not produce enough data to justify a dedicated layer, and the ones that do can afford to build their own. The middleware got built before the demand was measured. That is not infrastructure. That is a bet dressed as infrastructure. Model routing will follow the same curve: genuinely load-bearing for a handful of high-volume application companies, marginally useful to everyone else, and priced by the market as though the second group were the first. The financial logic still points one way, and it is worth spelling out, because it explains why this decision was probably inevitable regardless of who shipped what. Tiered routing, cheap open-weight models for scaffolding and boilerplate and file edits, frontier models for the reasoning-heavy steps, is the only real lever an application company has over gross margin. It is not a technology strategy. It is a survival strategy. Liquidity flows are just money with a pulse, and inference spend is liquidity leaving the building one token at a time. The underreported cost is observability. Once you route across three providers, you own an evaluation harness, a regression suite, and a per-route latency budget. You also own the silent failures, where a model produces syntactically valid code that passes tests and violates an assumption nobody wrote down. Those do not surface in a changelog. They surface in churn, six weeks later, and by then the narrative has moved on. The consensus read is that Replit is punishing a competitor and hedging inference costs. Both are plausible. Neither is verified. Consider a competing explanation the coverage has not entertained: this may be a negotiation conducted in public. A founder with a large audience saying we are diversifying our model suppliers is not a disclosure. It is leverage. It costs nothing to say and it moves the counterparty's willingness to discount. Announcements are cheap. Deployments are expensive. Until a replacement vendor is named, with a contract and an eval regression report attached, the correct position is that no migration has been confirmed. There is also a selection problem. We are hearing from the company that chose to talk. The ones quietly renewing their contracts are invisible, and they are almost certainly the majority. Correlation is not causation, and a press cycle is not a data set. The ledger does not lie, only the auditors do. What to watch is not whether Replit switches. It is whether the switch survives contact with an evaluation suite, and whether anyone else follows. If Cursor and Windsurf stay put, this is one company's margin problem. If they move within a quarter, it is a repricing of the model layer's entire distribution channel. In crypto I can fact-check the hype with cold, hard chain data. Here I have four sentences and a founder's timeline. Watch the commit history, not the quote.