Skip to content

Router Lab

High impactLive published free item

Frontier pricing converged, so the real choice moved down the stack

GPT-6 Astra and Claude Fable 5.1 now cost the same per token. With the top tier price-matched, the decision that moves your burn is which work never reaches the top tier at all.

Published 11 Sept 2026, 07:30Verified 11 Sept 2026, 07:30Sources 0

What changed

Within a week, OpenAI priced its new frontier model identically to Anthropic's at $10 and $50 per million tokens, while DeepSeek released a one-million-context open-weights model at $0.15 per million input off-peak and $0.003 per million cached. The spread between top and bottom tier is now roughly 65x on input.

Why it matters

When the two frontier options cost the same, choosing between them stops being an efficiency decision. The efficiency decision is the routing boundary: which tasks justify the top tier at all. A 65x spread means that moving even a modest share of routine turns off the frontier tier does more for your burn than any choice made inside it. This is the same argument the routing vendors are making, and for once their interest and yours point the same way.

Take

GPT-6 Astra and Claude Fable 5.1 now cost the same per token. With the top tier price-matched, the decision that moves your burn is which work never reaches the top tier at all.

What to do

Sample one week of your own work and sort it into two piles: tasks where a wrong answer costs you real time, and tasks where you would accept a good-enough draft. Price the second pile at cheap-tier rates. That number is your routing opportunity, and it is almost always larger than expected.

Evidence

No public sources are linked yet.