Anthropic's platform squeeze: when your model supplier ships your product, trust becomes a pricing input

The Wall Street Journal reports that resentment toward Anthropic is building in Silicon Valley, and it is not primarily about model quality. Startup founders, software company executives and AI researchers told the paper they started moving to cheaper models after watching Anthropic ship products that compete with tools sold by other companies. The dek names three grievances: competitive tactics, guardrails, and the company's lack of support for open-weight models — that is, models whose parameters are published so anyone can run, inspect and fine-tune them on their own hardware.

The case cited is Claude Design, released in April, which some read as a direct competitor to design company Figma — one of Anthropic's own partners. At an event shortly afterward, Figma CEO Dylan Field said Anthropic had not been "consistently candid in their communications," according to attendees quoted by the Journal. Sarah Sachs, head of AI at Notion, who was present, framed the lesson bluntly: "The whole industry learned a lot of lessons from how that dynamic played out. Some companies might have been overly trusting." These are characterizations from partners and attendees, not established findings; Anthropic's side is not represented in the excerpt.

Strip away the AI vocabulary and this is the oldest fight in platform economics. Every layer that hosts other people's businesses eventually discovers that the most attractive adjacent market is one of its tenants — the same dynamic that soured developers on the Twitter API in the 2010s and that has followed AWS for a decade. What is new is the speed. A foundation-model vendor can ship a competing product in weeks because the product *is* the model plus an interface, so the distance between "our partner's feature" and "our feature" has collapsed to a prompt and a UI. Partners who built on Claude did not sign up for a supplier with a two-week path into their category.

The reason the complaint has teeth in mid-2026, though, is substitution. By our own tracking, Anthropic still sits at the top of the Zendoric Quality index — Fable 5 at 90, ahead of GPT-5.6 Sol at 79 and Zhipu's open-weight GLM-5.2 at 77. A year ago that spread was a moat. Today it is a judgment call: for most production workloads, a model 13 points back that costs a fraction and carries no strategic conflict is a defensible choice. That is what "turning to cheaper AI models" actually means. Trust has stopped being a soft value and become a line item in the build-versus-buy spreadsheet, and open weights are the hedge — not because they are better, but because they cannot be repriced or turned into a competitor.

The guardrails charge deserves more care than the others. Refusal behavior and safety friction are genuine costs for developers, but they are also precisely why a slice of the enterprise market picked Anthropic in the first place. "Too restrictive" and "the responsible lab" are the same design decision seen from two P&Ls. Bundling it into the same indictment as competitive tactics is convenient for rivals, and some of this backlash is straightforward competitive positioning by companies that would like customers to switch.

Our reading: this is a healthy, clarifying conflict, and the important consequence is architectural. Enterprises are learning to treat models as swappable components rather than platform commitments — multi-model routing, portable abstractions, an open-weight fallback in reserve. That discipline is exactly what stops any single lab from owning the on-ramp to the technology, and it is the mechanism by which frontier capability becomes cheap infrastructure instead of rented privilege. The near term will be messy: opportunistic churn, partnerships renegotiated in public, some genuinely damaged relationships. The longer arc is better — power dispersed across layers, not concentrated in whoever trains the best model. What to watch next is whether Anthropic changes how it pre-announces launches to partners, and whether the defections show up in retention data or stay at the level of conference-hallway grievance.