How the cost — and energy — of a fixed AI & data stack has changed since 2023, even as the model behind each licence was swapped underneath you.
Enterprise AI is bought two ways, and the two prices behave differently. Per-user licences (Copilot, Gemini, Claude) have barely moved since 2023: you pay per head, whatever model sits underneath. API token prices have been far more volatile. They fell through 2023–25, then OpenAI's flagship rate doubled in April 2026 (GPT-5.5 at $5/$30 per million tokens, against GPT-5.4's $2.50/$15). Since August they have come back down. OpenAI cut GPT-5.6 Sol to $4/$20 from 21 August, Google put its Flash models on a half-price introductory rate, Opus 5.5 launched 20% below Opus 5, and GPT-6 Sol arrived on 22 September at $2/$10. Two of those cuts are temporary: GPT-5.6 Sol's rate runs to at least 21 November 2026, and Gemini Flash returns to $1.50/$7.50 on 1 January 2027. Energy per prompt has broadly fallen throughout. The charts below show where a licence is still the better deal and where paying per token now wins. Sources are listed at the bottom of the page.
A licence price barely moves while the model underneath it is swapped — the markers carry that story. Enterprise types are regional list per licence; your licence count drives the expected enterprise-agreement discount on the E5 add-on (≈20% at 5,000) — set a real rate in Licence price override to pin it. Personal subscriptions (Claude Pro / Max) are flat USD fees, FX-converted. Early access is now part of the licence type: the E7 Frontier Suite receives new models on release day, standard enterprise licences ≈6 weeks later, and personal Claude plans on release day. (The free Frontier program — early preview features — is not the paid E7 Frontier Suite; same name, different thing.) The Main model list greys out models the selected licence can't access — M365 Copilot has offered Anthropic models alongside OpenAI's since Sep 2025.
Hover the chart to read the values at any month.
Seat pricing is insensitive to your input/output ratio; API pricing is not — the ratio control above moves the API line and the break-even, never the licence line. That asymmetry is part of what the seat premium buys.
Price and energy are set independently: a model's $/token reflects the vendor's pricing decisions and its Wh/prompt reflects its engineering, so a new model can be pricier and more efficient at the same time. GPT-5.6 launched at GPT-5.5's price on similar energy; Grok 4.3 cost about 79% less than Grok 4 on similar energy. Price can also change with no new model at all, as with GPT-5.6 Sol's August promotion, which is a cost step with no energy step. CO₂ and water are fixed multiples of energy, so those three lines always move together. Only cost moves on its own.
Each service right now — its monthly cost set against the environmental footprint of the same workload, for every model in your stack at the current settings.