Enterprise AI usage pricing: token costs & cost modelling.
Most enterprise AI cost projections are wrong by a factor of two to five — almost never because the wrong vendor was chosen, but because token economics were never modelled properly. This note explains how token pricing actually works, where models break, and the framework that forecasts spend before you commit to enterprise volumes.
There is no single AI cost number — there is a model, and most are built wrong. Output tokens cost 3–5× input, system prompts and RAG context inflate every call, and multi-turn history compounds silently. Model selection moves cost more than negotiation: GPT-4o mini is ~94% cheaper than GPT-4o. Model the drivers first; negotiate the rate second.
01 Key findings
Token pricing is structurally unlike software licensing. Spend is driven per million tokens, split into input (text you send) and output (text generated) — and the gap between projection and actual routinely surprises even sophisticated procurement teams.
Output tokens cost 3 to 5× input tokens across every frontier provider. GPT-4o is $2.50/M in, $10.00/M out — a 4:1 ratio; Claude 3.5 Sonnet runs 5:1. Naive models anchor on input volume and miss where the money is.
The costly tokens are the invisible ones. System prompts (500–2,000 tokens per call), retrieved RAG chunks, and accumulated conversation history never appear in naive models yet dominate production bills.
Model selection beats negotiation. Choosing GPT-4o mini over GPT-4o for suitable workloads cuts cost ~94% at list price — a larger lever than any discount a procurement team can win.
Accurate models come from pilot data, not spreadsheets. System-prompt length, real input/output ratios, turn distribution and RAG volume can only be measured from a representative pilot at meaningful scale.
02 Token pricing mechanics
A token is roughly 3 to 4 characters of text, or about 0.75 words; a 1,000-word document is approximately 1,300 to 1,400 tokens. Models are priced per million tokens processed, with separate rates for input tokens and output tokens.
The critical asymmetry: output tokens cost 3 to 5× more than input tokens across all frontier providers. For GPT-4o the input rate is $2.50/M and the output rate $10.00/M — a 4:1 ratio; Claude 3.5 Sonnet is closer to 5:1. This reflects genuine computational cost in inference, and it changes modelling because most estimates focus on text sent to the model, not text generated in response.
Get the input/output ratio wrong by 2× and you double your cost-model error at a 4:1 pricing ratio. Output length varies wildly by use case — a 200-token summary of a 2,000-token document, versus a 2,000-token code generation from a 100-token spec.
03 Provider rate card
List rates for the models that carry most enterprise workloads. Input and output are quoted per one million tokens; the ratio column is the multiplier that quietly governs your bill.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Output / input ratio |
|---|---|---|---|
| GPT-4o | $2.50 | $10.00 | 4.0× |
| GPT-4o mini | $0.15 | $0.60 | 4.0× |
| o1 | $15.00 | $60.00 | 4.0× |
| Claude 3.5 Sonnet | $3.00 | $15.00 | 5.0× |
| Gemini 1.5 Pro | $1.25 | $5.00 | 4.0× |
| AWS Bedrock (Claude) | $3.00 | $15.00 | 5.0× |
04 Cost at scale
Because output volume drives spend, model choice has more impact than any negotiation. Output-token list rates span roughly 100× from the cheapest small model to the most capable reasoning model — the single largest cost lever available before a contract is signed.
The task is identifying which use cases genuinely need frontier capability and which are served adequately by smaller models. High reasoning complexity, high quality sensitivity and low volume favour frontier models; low complexity, moderate sensitivity and high volume strongly favour small ones. See our GPT vs Claude vs Gemini comparison.
05 The output & context trap
The costs that break budgets are the tokens nobody counts. Five patterns recur across the enterprise models we review, each causing systematic underestimation.
| Hidden driver | Mechanism | Impact if unmodelled |
|---|---|---|
| System prompts | 500–2,000 input tokens on every call, unchanging | 100k calls/day × 1,500 tokens = 150M tokens/day; ~$137,000/yr at $2.50/M |
| Output volume | Generated length modelled from input, not output | 2× ratio error doubles cost-model error at 4:1 pricing |
| Model sprawl | Teams drift to more capable, costlier models | Uniform-tier assumptions over- or under-state spend |
| Retries & refusals | Timeouts, rate limits and safety refusals still bill input | 3% retry rate adds 3%; refusals 0.5–2% on top |
| Context growth | Multi-turn history re-billed as input each turn | Underestimates conversational cost by 3 to 5× |
The biggest surprise is retrieval-augmented generation (RAG). A system retrieving 5 chunks of 500 tokens adds 2,500 input tokens to every query — on top of the system prompt and the user's question. Enterprise RAG at 500,000 queries per month can consume 2 to 3 billion tokens monthly in retrieved context alone. This is almost never in the initial cost model.
06 Cost-modelling framework
A defensible model needs four inputs most organisations lack at the start: real system-prompt length, realistic input/output ratios by use case, turn distribution for multi-turn systems, and RAG context volume. All four come from a pilot at meaningful scale, not a spreadsheet estimate.
Measure a 30-day pilot
Run representative workloads on the intended model. Record input, output, system-prompt, context and retry tokens separately — not as a single blended figure.
Apply a scale factor
Extrapolate pilot volume to production, accounting for seasonal peaks and adoption growth rather than assuming a flat linear ramp.
Model the tier split
Set the percentage of queries on lower-cost versus frontier models, and guard the split against silent model sprawl over time.
Add overhead, then commit
Layer in retries (2–5%), safety refusals (0.5–2%) and eval calls; apply your negotiated rate and commit volume at the 80th percentile of the modelled range.
07 Cost-reduction levers
Three levers move effective cost without touching a negotiated rate — and one hidden expense must be modelled explicitly.
All major providers process asynchronous requests at 50% of real-time pricing. Any workload that tolerates delay — document processing, overnight analysis, bulk classification — qualifies. Organisations that identify 40 to 60% of workload as batch-eligible cut effective AI cost by 20 to 30% with no negotiation.
Choosing GPT-4o mini over GPT-4o for appropriate use cases reduces cost ~94% at list price. Route by reasoning complexity, quality sensitivity and volume scale rather than defaulting every workload to a frontier model.
RAG and semantic search need embeddings billed separately from inference. A 10M-page base at $0.13/M tokens (~300 tokens/page) costs ~$390 to embed once — but re-embedding monthly makes it $4,680/year. For frequently updated bases, model it explicitly.
For the strategy context see the AI Procurement Guide and our AI procurement advisory service; for benchmarks, our OpenAI Enterprise pricing analysis.
08 Our recommendation
Build the cost model from pilot data, capturing system prompts, RAG context and multi-turn history. The drivers — not the headline rate — determine whether a budget holds.
Match each workload to the smallest model that meets its quality bar, and govern against model sprawl. This lever outperforms any rate you can negotiate.
Add batch eligibility and infrastructure overhead, then commit volume at the 80th percentile of the modelled range so peaks are covered without stranding spend.
Do not commit to volumes you have not modelled
Our AI practice builds accurate cost models from your pilot data, benchmarks vendor pricing, and structures commitments that protect you from overspend.
The Licensing Edge
Weekly AI pricing intelligence and cost-management insight for enterprise procurement teams. 3,000+ subscribers.