Flight recorder
Research and engineering notes on model routing — including the Router Graveyard series: post-mortems of routing projects that didn't make it, and what that says about the ones that will.
The April 4, 2026 Anthropic ban kicked off a gold rush of 'route your existing consumer subscriptions' routers. At least seven shipped between mid-2025 and mid-2026. Most repeat the same structural mistakes as the 2023-2024 graveyard: ToS exposure, hosted concentration, learned routers that rot, and zero moat on policy. Here's the actual list, the failure modes, and the only pattern that survives.
For a local-first, one-time product, the payment processor is part of the product. Polar's license keys and no-pressure economics matched our soul pillars. LemonSqueezy's merchant-of-record model would have pushed us toward subscriptions and away from the thing we are building.
Bifrost published benchmarks claiming up to 54× faster routing than LiteLLM. The numbers are real for a narrow test case, but they measure the wrong thing for most production workloads. Raw proxy speed is rarely the constraint once you have subscriptions, policies, and real provider latency.
The May 29, 2026 acquisition of Portkey by Palo Alto Networks shows that at enterprise scale, 'routing' is being absorbed into governance, DLP, and compliance stacks — not sold as a standalone smart-picker. The durable surface is control, not intelligence.
The $1.3B post-money raise into OpenRouter proves that hosted routing has consolidated around billing aggregation and catalog breadth, not routing intelligence. The pure 'smart router' companies are dead or pivoted. Local policy is the only durable escape hatch.
The April 4, 2026 ban on using Claude Pro and Max subscriptions through third-party routers was not a surprise enforcement. It was the predictable end of an arbitrage that providers had been signaling against for months. This is the canary for the entire subscription-OAuth routing category.
Agent frameworks promised to orchestrate tools, memory, and multi-step reasoning. Most are gone, pivoted, or absorbed. What killed the category — and what actually endured.
Model aggregators promised one key, one invoice, every model. Most are gone, pivoted, or maintenance-mode. What killed the category — and what the survivors actually sell.
LLM routing looked like an obvious business in 2023. Three years later, most of the first wave is dormant, pivoted, or rebranded. What killed them — and what actually survives.
Get new posts + product updates
The Router Graveyard series and RoutePlane releases. No noise, unsubscribe anytime.