The frontier benchmark tax: GPT-5.5 Pro costs up to 191× the open-weight floor

The US frontier sets the ceiling, not the floor. GPT-5.5 Pro lists at $15/Mtok input and $90/Mtok output; GPT-5.5 at $12.5/$75; Claude Opus 4.7 at $5/$25 — and Opus carries a +35% tokenizer tax on JSON and code, pushing its effective rate to roughly $6.75/$33.75.

Against the EU open-weight floor — gpt-oss-120b on EU-sovereign OVHcloud at $0.09/$0.47 — GPT-5.5 Pro is 167× more expensive on input and 191× on output. For the large majority of production workloads (RAG, extraction, classification, routing, agents), the open-weight floor on sovereign EU infrastructure wins on cost and on data sovereignty.

That is why the matrix now reads as two markets: a featured EU Open-Weight Market that finds the floor, and a dimmed Showcase · frontier benchmark where the closed US models belong — useful for measuring quality, not for paying the bill.