★ EUROPE'S SOVEREIGNTY
TokenFloor · Insights
Findings from the floor — hidden taxes, price disruptions, vendor markups.
OpenAI's own open model is cheapest on EU-sovereign infrastructure
gpt-oss-120b runs at $0.09 in / $0.47 out per million tokens on OVHcloud's EU-sovereign French infrastructure — 22% below the same model on US Together, and outside the reach of the US CLOUD Act. The token floor is now open-weight and European.
The frontier benchmark tax: GPT-5.5 Pro costs up to 191× the open-weight floor
GPT-5.5 Pro lists at $15/$90 per million tokens. The same class of work on gpt-oss-120b — OpenAI's own open model — costs $0.09/$0.47 on EU-sovereign OVHcloud: 167× cheaper input, 191× cheaper output. Closed frontier models are a benchmark, not the floor.
Anthropic's +35% tokenizer tax on code and JSON
Claude Opus 4.7 counts roughly 35% more tokens on JSON and code than Opus 4.6 — silently inflating the bill on agentic and structured workloads with no announcement, no price-list footnote.
Azure's 6.2× H100 markup vs RunPod
Microsoft Azure lists the same NVIDIA H100 SXM5 chip at $12.29/h on-demand. RunPod sells identical hardware at $1.99/h. Same chip, same power envelope — 6.2× the bill.
DeepSeek V4 Pro undercuts the Western frontier by an order of magnitude
At $2.10 in / $4.40 out per million tokens, DeepSeek V4 Pro prices the same 1.6T-MoE class roughly 10× below Claude Opus 4.7 and 14× below GPT-5.5 Pro — without sacrificing benchmark parity on coding and reasoning.