Cost Per 1K Users: AI Unit-Economics Benchmarks

Average AI spending reached $2,068 per employee in 2026, up 50% from $1,358 in 2025. But the distribution is extreme: more than half of companies expect to spend no more than $200 per employee, while the top 10% plan to invest at least $2,800 per employee (Federal Reserve Bank of Atlanta via Rize, 2026). That 14x gap shows how differently companies are adopting AI.

For a SaaS company building an AI feature, "cost per user" means total AI infrastructure spend divided by active users who actually consume the feature. It is not the same as the per-seat price of a Copilot or ChatGPT Enterprise license. It includes:

• Model API spend (input and output tokens).
• Embedding models and vector database costs.
• Orchestration, caching, and retrieval infrastructure.
• Observability, guardrails, and fallback systems.

The denominator matters. Use monthly active users of the AI feature, not registered accounts or total employees. A user who logs in but never clicks the AI assistant does not consume tokens. A power user who runs 50 AI queries per day does. If your product has 10,000 MAU but only 2,000 use the AI feature, your per-user cost is five times higher when calculated against total MAU.

Model choice is the single biggest lever on your AI bill. OpenAI's GPT-5.4 Nano costs $0.20 per million input tokens and $1.25 per million output tokens. GPT-5.5 costs $5.00 and $30.00 respectively. That is a 25x spread on input and a 24x spread on output for the same token volume ( OpenAI API pricing, 2026).

Anthropic follows a similar tier structure. Claude Opus 5 inputs at $5.00 per million and outputs at $25.00 per million. Sonnet 5 inputs at $2.00 and outputs at $10.00. Haiku 4.5 inputs at $1.00 and outputs at $5.00 ( Anthropic pricing, 2026). The pattern is consistent across providers: output tokens cost 4–6x more than input tokens because generation requires more compute.

That output premium is why long responses are expensive. A customer-support answer that returns 800 output tokens costs roughly 4–6x more than the 800 input tokens that described the ticket. For a deeper dive on per-token pricing, see our LLM inference cost per million tokens guide.

A simple Q&A bot averages 500–2,000 tokens per session. A multi-turn chatbot runs 2,000–10,000 tokens. Document analysis can hit 5,000–50,000 tokens. Agentic workflows can burn 10,000–200,000+ tokens in a single run (FinOps Foundation, Tokenomics working group, 2026). The same user base can produce wildly different costs depending on what the AI feature actually does.

Here are three worked examples using real 2026 OpenAI pricing.

Light AI assistant: ~$0.30 per user per month

A user makes 100 queries per month, averaging 1,500 tokens total per query (input plus output), routed to GPT-5.4 Mini at $0.75 input and $4.50 output per million tokens. Assuming a 2:1 input-to-output ratio, that is 1,000 input tokens and 500 output tokens per query.

Support chatbot: ~$0.05 per active user at scale

A SaaS company routes 10,000 support queries per day through GPT-5.4 Mini, averaging 500 input tokens and 300 output tokens per query. Based on CloudZero's worked example using 2026 pricing, that workload costs about $518 per month. At 10,000 active users, that is roughly $0.05 per active user per month ("active" meaning users who actually submit tickets).

Document analysis: ~$1.50–$3.00 per user per month

A legal-tech or research feature generates 50 AI reports per user per month, each averaging 10,000 tokens (mostly input context). On GPT-5.4 Standard at $2.50 input and $15.00 output per million tokens, the cost lands around $1.50–$3.00 per user per month depending on output length. Move the same workload to GPT-5.4 Mini and the cost drops to roughly $0.45–$0.90 per user.

The curve looks similar to cloud cost per user, but the driver is different. Cloud costs fall because fixed infrastructure spreads across more users. AI costs fall because the same model API spend is divided by more users, while retrieval and orchestration costs stay relatively flat. For more on why agents blow up the token math, see AI agent token costs.

Tokens are a cost unit. Outcomes are a value unit. The most useful AI metric is cost per outcome: total AI feature cost divided by the number of completed, valuable actions. A support resolution, a generated report, a code review, or a sales email are all outcomes. Vendors are already pricing this way: Salesforce Agentforce charges $2 per conversation, and Intercom FinAI Agent charges $0.99 per successful resolution (Zylo 2026 SaaS Management Index, citing High Alpha).

Here is a concrete comparison. A human support agent costs roughly $8 per ticket when you include salary, benefits, and overhead. An AI agent that resolves a ticket on GPT-5.4 Mini might consume 500 input tokens and 300 output tokens, costing about $0.18 per resolution. Even after adding retrieval, orchestration, and quality-assurance overhead, the AI outcome is typically 10–40x cheaper than the human equivalent.

The catch is that not every AI interaction produces an outcome. A user who asks three follow-up questions before getting a useful answer consumes three queries for one outcome. That is why cost per outcome is a stricter test than cost per user. It forces you to measure whether the AI feature actually finishes the job.

For products using retrieval-augmented generation, remember that the harness around the model call — vector database, embeddings, orchestration, observability — can represent 40–60% of total feature spend (FinOps Foundation, 2026). A cost-per-outcome calculation that only counts tokens will understate real cost by half. See our production RAG system cost breakdown for a full component map.

Seventy-eight percent of IT leaders report unexpected charges on SaaS due to consumption-based or AI pricing models (Zylo, 2026). The surprise comes from four sources that do not exist in traditional software.

1. Output length is unpredictable

You control the input prompt, but the model controls the output length. A user asking for a one-sentence summary might get a three-paragraph essay. Output tokens cost 4–6x more than input tokens, so a single verbose response can dominate the day's spend.

2. Context windows accumulate

Conversational AI resends the full conversation history on every turn. A session that starts at 1,000 tokens can grow to 10,000 tokens as the user asks follow-up questions. Each subsequent turn costs 10x the first. Without context pruning, costs grow linearly with session length.

3. Shadow AI duplicates spend

Shadow AI — employees using unapproved AI tools — costs companies an average of $412,000 per year. Seventy-eight percent of workers use unapproved tools, and 34% of that shadow spending duplicates tools the company already pays for (HelpNetSecurity via Rize, 2026). The real AI cost per user is the vendor invoice plus the shadow spend.

4. Vendors apply an AI tax at renewal

Tropic's 2026 data shows AI-driven price increases of 20–37% at renewal, far above the typical 3–9% uplift. Negotiation reduces those asks by roughly 55%, with final uplifts averaging ~12%. Credit-based pricing and forced SKU migrations make the true cost harder to benchmark.

The result is that AI budgets based on per-seat assumptions are usually wrong. The only reliable approach is to instrument usage, establish a baseline, and set budgets at 110–120% of baseline with alerts at 80% and 100% (FinOps Foundation, 2026).

Model right-sizing is the highest-impact optimization available. Routing simple queries to a mini or nano model and reserving frontier models for complex work can cut average cost per query by 60–90% (FinOps Foundation, 2026). The FinOps Foundation also reports typical savings of 50% from batch APIs, 50–90% on cached tokens from prompt caching, and 20–60% from context window management.

1. Route by task complexity

Not every query needs a frontier model. Use GPT-5.4 Nano or Haiku 4.5 for classification, extraction, and short answers. Use GPT-5.4 Mini or Sonnet 5 for multi-turn chat and summarization. Reserve GPT-5.5 or Opus 5 for tasks where evaluations prove a measurable quality lift. A routing layer that escalates only ambiguous queries can reduce average cost per call by 60–80%.

2. Cache what repeats

Prompt caching from OpenAI and Anthropic stores stable prefixes like system prompts and reference documents at a 50–90% discount. Semantic caching returns answers for similar queries without calling the model at all. For customer-facing features with repetitive questions, cache hit rates of 20–50% are achievable.

3. Cap output length

Set max_tokens and instruct the model to be concise. Unconstrained outputs can cost 4–6x more than necessary. A simple instruction like "answer in under 150 words" can cut output token spend by 10–40% without hurting usefulness.

4. Use batch for offline work

OpenAI's Batch API processes requests asynchronously within 24 hours at 50% off standard pricing. Any non-real-time workload — nightly report generation, document classification, content moderation — should use batch or flex processing.

5. Instrument before you optimize

You cannot optimize what you cannot attribute. Deploy API key governance, a proxy or observability layer, and unit-cost dashboards showing cost per query, per user, and per outcome. Without this visibility, pricing decisions are guesswork.

For a broader cost-optimization playbook, see our cloud cost optimization checklist. Many of the same disciplines — tagging, right-sizing, commitment pricing — apply to AI spend once you have visibility.