← All writing
Strategy

The 18-month window before agents commoditize

Model providers can't credibly route to cheaper competitors — it cannibalizes their margin. The structural moat is real, but it's a window, not a permanent condition.

April 30, 2026 · 6 min read

OpenAI and Anthropic can’t route your work to cheaper competitor models. It’s not that they wouldn’t if they could — it’s that doing so cannibalizes their own pricing tiers and undermines the whole business. Their incentive is to keep you on their newest, most expensive thing.

That structural mismatch is sitting underneath every multi-model platform you can think of right now, including Sulla. It’s the moat. And it’s a window, not a permanent condition.

The mechanics

Look at what’s happening:

  • OpenAI is reportedly testing internal model routing — the GPT-5 “router” that picks between fast and reasoning modes
  • Anthropic ships Sonnet vs. Haiku selection
  • Microsoft Copilot routes between models based on task
  • Grok already does multi-model routing across providers

In ~18 months, every major provider will offer some form of routing. The “we route differently” pitch becomes false.

But here’s the thing nobody’s saying: even when they all route, they’ll never route to competitors. OpenAI will never recommend GPT-4o-mini for the task and then quietly hand it to a Haiku call. The routing stays within their own model line. That’s still a real differentiator for genuinely model-neutral platforms — but it’s narrower than the moat we have today.

What I’m betting on instead

The durable position isn’t “we route differently” — it’s “you control the routing AND own the data infrastructure.” Two things that stay true even after everyone routes:

  1. You decide the policy. Cost-optimize? Speed-optimize? Quality-optimize? Your platform, your call.
  2. Your data lives where you say. Not in OpenAI’s cloud. Not in Anthropic’s cloud. On your infra.

That’s not a 12-month window. That’s a structural answer to a question the giants can’t credibly give a different answer to without cannibalizing themselves.

Why this matters for what you ship now

If you’re building anything that depends on multi-model neutrality as a feature, ship it now. The window is real and it’s open. By Q2 2027, the conversation will be different. By 2028, “we route between models” will be table stakes.

The race is to build the next differentiator while the current one is still earning. That’s what every platform is doing right now whether they’re saying it out loud or not.

— J