| Model | Provider | In / Out |
|---|---|---|
| Claude Fable 5.1 | Anthropic, US | $10 / $50 |
| Claude Opus 5 | Anthropic, US | $5 / $25 |
| GPT-5.6 Sol | OpenAI, US | $4 / $20 |
| Kimi K3 | Moonshot, China | $3 / $15 |
| Gemini 3.1 Pro | Google, US | $2 / $12 |
| Claude Sonnet 5 | Anthropic, US | $2 / $10 |
| Mistral Medium 3.5 | Mistral, France | $1.50 / $7.50 |
| Grok 4.5 | xAI, US | $2 / $6 |
| Claude Haiku 4.5 | Anthropic, US | $1 / $5 |
| GLM 5.2 | Zhipu, China | $1.40 / $4.40 |
| Gemini 3.7 Flash | Google, US | $0.75 / $3.75 |
| Mistral Large 3 | Mistral, France | $0.50 / $1.50 |
| GPT-5.6 Luna | OpenAI, US | $0.20 / $1.20 |
| DeepSeek V4-Pro | DeepSeek, China | $0.44 / $0.87 |
| Mistral Small 4 | Mistral, France | $0.15 / $0.60 |
| Grok 4 Fast | xAI, US | $0.20 / $0.50 |
| Gemini Flash-Lite | Google, US | $0.10 / $0.40 |
| DeepSeek V4-Flash | DeepSeek, China | $0.14 / $0.28 |
| Amazon Nova Micro | Amazon, US | $0.035 / $0.14 |
| Qwen3.7 Flash | Alibaba, China | $0.03 / $0.13 |
| Ministral 3, 3B | Mistral, France | $0.10 / $0.10 |
Gemini 3.1 Pro and both Grok models charge double past 200K tokens of context.
| Option | Jurisdiction | Price |
|---|---|---|
| US frontier via EU region | Still US. The CLOUD Act reaches data held anywhere by a US company | Same as direct |
| AWS European Sovereign Cloud | German operating entity, stronger controls, some residual exposure | Premium |
Data residency and legal jurisdiction are not the same thing. Where the servers sit does not change which government can compel the company holding the keys.
| Provider | What they run | Note |
|---|---|---|
| Mistral, serving GLM 5.2 | Zhipu's model on Mistral's infrastructure, France | $1.40 / $4.40 |
| Verda, Helsinki | LLM Gateway, sells to Finnish finance and healthcare | On request |
| Scaleway | France, runs Mistral workloads | On request |
| OVHcloud | France, largest European-owned cloud | On request |
| Nebius | Amsterdam, data centres in Finland | On request |
| IONOS | Germany | On request |
| Hetzner | Germany | On request |
| EuroHPC AI Factories | EU public compute, 19 sites selected | Free |
EuroHPC is worth knowing about and almost nobody uses it. AI startups and SMEs can get GPU hours on European AI Factories at no charge through the access calls. Playground access is always open with a small allocation and approval inside two working days. Fast Lane goes up to 50,000 GPU hours with approval inside four working days. One caveat your accountant will raise: free compute has measurable value and its state aid treatment is unsettled.
This is the row most cost comparisons leave out. A Chinese-built model running on European machines puts your data under European law at close to Chinese prices.
| Option | What decides it | Reality check |
|---|---|---|
| Rented GPU, open weights | Utilisation. A card costs the same idle as busy | 50-60% load needed |
| Own hardware, on premise | Jurisdiction. Nothing leaves the building | 5-12M tokens/day to break even vs frontier |
| Workstation or Mac | Zero marginal cost after purchase | Pays back in 3-8 months at €200/mo spend |
| On device, small models | Good enough for sorting, tagging, extraction | gpt-oss-20b fits in 16GB |
One worked example: 50 million tokens a day costs about $2,250 a month on a cheap service and about $5,175 on four A10G cards. Against budget services, self-hosting almost never wins on cost. Against the most expensive models it can, at high and steady volume. Open models run roughly a generation behind the best closed ones, which for most ordinary work does not matter.
| Model | Terms |
|---|---|
| gpt-oss 20b and 120b | Apache 2.0, properly permissive in both sizes |
| Mistral open models | Apache 2.0 for research and individual use. Production needs a commercial licence |
| GLM | Moved off clean MIT terms, conditions now apply to large hosts |