Sources

Every price in the dataset links to the provider's official page and carries a verification date. Every formula is documented on the methodology page. Prices are publicly listed list prices and are convenience defaults — overridable in every tool.

API pricing

• Claude Sonnet-class (Anthropic) — $3/$15 per 1M (in/out), verified Jun 25, 2026. Official pricing →
• GPT-4o-class (OpenAI) — $2.5/$10 per 1M (in/out), verified Jun 25, 2026. Official pricing →
• Gemini Flash-class (Google) — $0.3/$2.5 per 1M (in/out), verified Jun 25, 2026. Official pricing →
• Llama-70B via managed API (e.g. Groq/Together) — $0.6/$0.8 per 1M (in/out), verified Jun 25, 2026. Official pricing →

GPU pricing

• NVIDIA A100 80GB — 80 GB, from $1.59/hr (cloud), verified Jun 25, 2026. Source →
• NVIDIA H100 80GB — 80 GB, from $2.99/hr (cloud), verified Jun 25, 2026. Source →
• NVIDIA L40S 48GB — 48 GB, from $0.99/hr (cloud), verified Jun 25, 2026. Source →
• NVIDIA RTX 4090 24GB — 24 GB, from $0.69/hr (cloud), verified Jun 25, 2026. Source →

Open-weight models

• Llama-3 70B — 70B parameters, Llama Community. Model card →
• Mixtral 8x7B — 47B parameters, Apache-2.0. Model card →
• Llama-3 8B — 8B parameters, Llama Community. Model card →
• Qwen2 7B — 7B parameters, Apache-2.0. Model card →

Formulas & references

• Token pricing and $/1M-token math — see methodology §1.
• GPU amortization and cost-per-token — methodology §2–§6.
• VRAM estimation (bytes-per-parameter by quantization, KV-cache headroom) — methodology §7.
• FX rates (EUR/GBP): Frankfurter, cached daily with a bundled fallback.

Spotted a stale price or an error? Let me know — corrections are credited.