Anthropic hit $30B — your stack is on the wrong model

The headline number masks the real shift. Anthropic went from $1B ARR in January 2025 to $30B in April 2026 — 30x in fifteen months, and it did it spending roughly 4x less on training than OpenAI. More importantly, 80% of revenue is enterprise versus OpenAI's consumer-heavy mix. Enterprise revenue has higher retention, better expansion economics, and it compounds. The $1M+ contract count doubling from 500 to 1,000 in under two months is the tell: this isn't a spike, it's a land-and-expand engine hitting its second derivative.

The part most operators are missing: the market has chosen a winner for enterprise workloads, and most founder stacks were architected between Q2 2023 and Q4 2024, when betting on OpenAI was the obvious call. That choice has quietly become a tax. Claude Opus 4.7 and Sonnet 4.6 have opened measurable instruction-following and long-context deltas on the exact workloads enterprises buy — code, legal review, customer ops, multi-step agents. If your production traffic is still 90%+ on GPT-class models, you're paying a premium for worse output on the workflows that matter most.

The pricing side is where this gets interesting. Salesforce just shipped AELA — Agentic Enterprise License Agreement — flat fee, unlimited agent usage, multi-year lock. Adecco signed in March for unlimited Agentforce across 60+ countries. OpenAI is quietly cutting 25-40% off list for volume commits with data isolation and SLA upgrades. Anthropic is doing custom $1M+ deals with named Slack support. The per-token moat is collapsing. Vendors are taking losses to secure multi-year footprint because they know switching costs compound inside an enterprise once agents are wired into workflows.

Here is the contrarian read: the winning move is not 'switch to Claude.' It's building switchable. The operators we see cutting 30-40% off AI spend in Q1 didn't pick a model — they picked an abstraction. A thin router layer, a proper eval harness, and quarterly renegotiation cycles where they walk into both account teams with benchmarks and invoice data. That posture flips the leverage. When two vendors are both burning cash to lock you in, the founder who can credibly walk gets priced like a $1M customer at $300k of actual spend.

The ones who lose this cycle are the founders who bet right on one model in 2024 and are now emotionally anchored to the choice. The ones who win are boring: model-agnostic architecture, monthly eval runs, and a standing 'Q2 model review' meeting with both account teams on the calendar. The technology is not the moat. The procurement discipline is.