How Providers Price Tokens: The Structure of an AI Bill
Major AI providers charge for tokens in dollars per million, abbreviated MTok, with separate rates for input and output. The model tier determines the per-token rate, and across every current-generation model from Anthropic and OpenAI, output tokens cost five times more than input tokens.
That ratio is not arbitrary. Output generation requires a full forward pass through the model for each token produced. Reading input can be parallelized. The pricing reflects what is actually happening computationally, and it has a direct implication for cost modeling: a workload that doubles its average output length does not double its bill, but it is still far more expensive than the input token count alone would suggest.
The table below shows current standard list rates for the models most commonly used in production, alongside the calculated cost for a workload of one million requests carrying 2,000 input tokens and 500 output tokens each :
| Provider | Model | Input ($/MTok) | Output ($/MTok) | Cost: 1M Requests (2K In / 500 Out) |
|---|---|---|---|---|
| Anthropic | Claude Haiku 4.5 | $1.00 | $5.00 | $4,500 |
| Anthropic | Claude Sonnet 4.6 | $3.00 | $15.00 | $13,500 |
| Anthropic | Claude Opus 4.8 | $5.00 | $25.00 | $22,500 |
| OpenAI | GPT-5.4 | $2.50 | $15.00 | $12,500 |
| OpenAI | GPT-5.5 | $5.00 | $30.00 | $25,000 |
Standard, non-cached API rates per million tokens (MTok) from Anthropic’s and OpenAI’s official pricing documentation, June 2026. Cost column formula: (1M x 2,000 / 1,000,000 x input rate) + (1M x 500 / 1,000,000 x output rate). Batch processing and prompt caching reduce these figures further.
Two things stand out in that table. First, the gap between model tiers is not a rounding error. Sending the same one million requests to Claude Opus 4.8 instead of Claude Haiku 4.5 costs five times more. If the work requires frontier-model capability, that premium has a justification. If it is classification or extraction work, it probably does not.
Second, output tokens concentrate the cost even when they are a small fraction of total tokens. In the workload above, output tokens are 500 of every 2,500 tokens per request, which is 20% of the token volume. On Haiku 4.5, they account for $2,500 of the $4,500 total, which is 56% of the cost. Any feature producing long responses without a specific reason to should show up on a cost review.
The rate card tells you what a token costs. The workload shape determines what you actually pay. A team managing AI spend without visibility into average output length, model tier mix, and caching utilization is working from an incomplete number.
Why AI Spend Does Not Behave Like Cloud Spend
Conventional cloud cost management works because the unit of consumption is stable and predictable. An EC2 instance has a known hourly rate. A gigabyte of storage has a known monthly rate. The cost of one unit does not change based on what is inside it. Token pricing does not work that way.
Each API call generates a different token count depending on the system prompt length, how much conversation history is carried, how much context was retrieved, and how long the model’s response turned out to be. There is no fixed cost per request. The cost is recalculated fresh for each call, based on exactly what that call contained. This makes AI spend behave more like metered electricity than like a software license, and it requires a different approach to monitoring and forecasting.
- Agentic workflows compound the variability further. When an AI agent handles a task through a sequence of model calls, the number of calls in that chain is a runtime decision, not a parameter set in advance. A single user action might trigger two model calls in a simple case and twelve in a complex one. The cost of an interaction is not deterministic before it happens.
- Model routing adds another layer. Applications that route requests across model tiers, sending simple requests to cheaper models and complex ones to frontier models, have a cost per feature that is a distribution, not a fixed number. That distribution shifts as user behavior and routing logic change.
- The aggregate shows up without attribution. All of this variable spend typically lands on an invoice as a single line item per model, per month. Before an organization builds attribution infrastructure, answering a question like “which product feature drove this month’s Bedrock costs” requires manual log correlation work that most teams have not had time to build.
The FinOps Foundation’s Token Economics working group identifies these structural issues as the defining challenges of managing AI spend at enterprise scale. The root causes are consistent: developer-led purchasing, billing data that aggregates rather than allocates, and pricing models that vary across model tiers and use cases in ways that do not map cleanly onto traditional cost management tools.
How Providers Price Tokens: The Structure of an AI Bill
Major AI providers charge for tokens in dollars per million, abbreviated MTok, with separate rates for input and output. The model tier determines the per-token rate, and across every current-generation model from Anthropic and OpenAI, output tokens cost five times more than input tokens.
That ratio is not arbitrary. Output generation requires a full forward pass through the model for each token produced. Reading input can be parallelized. The pricing reflects what is actually happening computationally, and it has a direct implication for cost modeling: a workload that doubles its average output length does not double its bill, but it is still far more expensive than the input token count alone would suggest.
The table below shows current standard list rates for the models most commonly used in production, alongside the calculated cost for a workload of one million requests carrying 2,000 input tokens and 500 output tokens each :
| Provider | Model | Input ($/MTok) | Output ($/MTok) | Cost: 1M Requests (2K In / 500 Out) |
|---|---|---|---|---|
| Anthropic | Claude Haiku 4.5 | $1.00 | $5.00 | $4,500 |
| Anthropic | Claude Sonnet 4.6 | $3.00 | $15.00 | $13,500 |
| Anthropic | Claude Opus 4.8 | $5.00 | $25.00 | $22,500 |
| OpenAI | GPT-5.4 | $2.50 | $15.00 | $12,500 |
| OpenAI | GPT-5.5 | $5.00 | $30.00 | $25,000 |
Standard, non-cached API rates per million tokens (MTok) from Anthropic’s and OpenAI’s official pricing documentation, June 2026. Cost column formula: (1M x 2,000 / 1,000,000 x input rate) + (1M x 500 / 1,000,000 x output rate). Batch processing and prompt caching reduce these figures further.
Two things stand out in that table. First, the gap between model tiers is not a rounding error. Sending the same one million requests to Claude Opus 4.8 instead of Claude Haiku 4.5 costs five times more. If the work requires frontier-model capability, that premium has a justification. If it is classification or extraction work, it probably does not.
Second, output tokens concentrate the cost even when they are a small fraction of total tokens. In the workload above, output tokens are 500 of every 2,500 tokens per request, which is 20% of the token volume. On Haiku 4.5, they account for $2,500 of the $4,500 total, which is 56% of the cost. Any feature producing long responses without a specific reason to should show up on a cost review.
The rate card tells you what a token costs. The workload shape determines what you actually pay. A team managing AI spend without visibility into average output length, model tier mix, and caching utilization is working from an incomplete number.
Why AI Spend Does Not Behave Like Cloud Spend
Conventional cloud cost management works because the unit of consumption is stable and predictable. An EC2 instance has a known hourly rate. A gigabyte of storage has a known monthly rate. The cost of one unit does not change based on what is inside it. Token pricing does not work that way.
Each API call generates a different token count depending on the system prompt length, how much conversation history is carried, how much context was retrieved, and how long the model’s response turned out to be. There is no fixed cost per request. The cost is recalculated fresh for each call, based on exactly what that call contained. This makes AI spend behave more like metered electricity than like a software license, and it requires a different approach to monitoring and forecasting.
- Agentic workflows compound the variability further. When an AI agent handles a task through a sequence of model calls, the number of calls in that chain is a runtime decision, not a parameter set in advance. A single user action might trigger two model calls in a simple case and twelve in a complex one. The cost of an interaction is not deterministic before it happens.
- Model routing adds another layer. Applications that route requests across model tiers, sending simple requests to cheaper models and complex ones to frontier models, have a cost per feature that is a distribution, not a fixed number. That distribution shifts as user behavior and routing logic change.
- The aggregate shows up without attribution. All of this variable spend typically lands on an invoice as a single line item per model, per month. Before an organization builds attribution infrastructure, answering a question like “which product feature drove this month’s Bedrock costs” requires manual log correlation work that most teams have not had time to build.
The FinOps Foundation’s Token Economics working group identifies these structural issues as the defining challenges of managing AI spend at enterprise scale. The root causes are consistent: developer-led purchasing, billing data that aggregates rather than allocates, and pricing models that vary across model tiers and use cases in ways that do not map cleanly onto traditional cost management tools.
What Attribution Looks Like Today
Attributing AI spend to the teams, features, and workloads that generated it is the next problem after establishing what the spend is. AWS addressed a significant piece of this in April 2026 with the introduction of granular cost attribution for Amazon Bedrock. The feature automatically captures which IAM principal made each Bedrock API call and surfaces that identity, along with tags applied to it, in Cost and Usage Report data. Before this launch, AWS itself described the process of connecting Bedrock costs to specific teams as requiring manual reconciliation of CloudTrail logs against billing data: time-consuming, error-prone, and difficult to maintain at scale.
The table below outlines the three attribution mechanisms now available for Bedrock workloads and what each one is best suited for :
| Mechanism | What It Attributes By | Best Suited For |
|---|---|---|
| IAM principal-based allocation |
Caller identity: IAM user, role, or federated identity. New line_item_iam_principal column in CUR 2.0.
|
Teams with direct API callers, dedicated service accounts per application, or role-per-service architectures |
| Cost allocation tags on IAM principals | Tags attached to IAM users or roles: team, project, cost center, or any custom dimension | Aggregating costs across multiple principals under a shared business dimension (e.g., department-level chargeback) |
| Model invocation logs | Per-request metadata including token counts, model ID, and timestamp | Per-prompt, per-feature, or per-experiment attribution requiring individual request-level detail |
Source: AWS Bedrock cost management documentation and IAM principal-based cost allocation launch, April 2026. IAM principal attribution requires CUR 2.0 (AWS Data Exports). Per AWS guidance, cost allocation tags take up to 24 hours to appear in Cost Explorer after activation and are not applied retroactively.
That last point on retroactivity is worth underlining. Tags applied today do not reach back and label last month’s spend. Every month an organization runs Bedrock workloads without an active tagging strategy in place is a month of spend that cannot be allocated after the fact. The AWS guidance on Bedrock cost allocation is explicit on this. The cost of delay compounds.
How CloudInvent Identifies This
CloudInvent’s 200+ AI-driven algorithms are built to surface the structural gaps in AI spend that standard billing dashboards do not make visible. Our analysis works entirely from billing and usage metadata, with no access required to prompts, completions, or model outputs, and we evaluate AI spend across the same three dimensions we apply to cloud infrastructure: utilization patterns, pricing structure, and workload architecture.
For organizations with meaningful AI spend already in production, the first pass of analysis typically identifies model routing decisions that default to a more expensive tier than the work actually requires, output verbosity inflating the output-side bill without clear justification, and attribution gaps that make it impossible to connect spend to the teams or features driving it. These findings sit alongside instance-level and rate-based optimization work, because the underlying principle is the same regardless of whether the resource is a compute instance or a model endpoint: attribute the cost, match consumption to the cheapest option that clears the quality bar, and measure the result.
Our work with customers across AWS, Azure, and GCP consistently produces 25% to 40% reductions in cloud IT spend, achievable in the first 30 days. As AI spend becomes a larger share of the technology footprint, the same metadata-driven approach that has always driven those results extends directly into token-level cost management.
Practical Considerations Before You Build a Token Budget
- Establish attribution infrastructure before usage scales. Tags are not retroactive. Spend that occurs before IAM principal tags are activated on Bedrock cannot be allocated after the fact. Set up the tagging strategy, activate the tags, and enable CUR 2.0 data exports before usage grows, not after it already has.
- Separate input and output in every cost model. Because output tokens cost five times more than input, a cost model that treats them as equivalent will consistently underestimate bills for any workload that generates substantive responses. Track the two separately from day one.
- Treat model selection as a routing decision, not an architecture choice. The cheapest model that clears the quality bar for a given task is rarely the same model across every task the system performs. Static model assignment is one of the most consistent sources of unnecessary AI spend, and it is one of the most straightforward things to change.
- Log token counts on every request in production. API response metadata includes exact input, output, and cached token counts. Logging this data per request, per model, and per feature gives you the raw material for cost attribution, anomaly detection, and forecast modeling. Without it, the only number available is the aggregate monthly invoice.
- Watch output length as a first-order cost signal. Given the 5:1 output-to-input price ratio, average output token length is one of the strongest predictors of total spend for generation-heavy workloads. Cap output where responses do not need to be long, and audit any feature where average output has grown without a corresponding change in user behavior or requirements.
The Takeaway
AI spend is not a smaller version of cloud spend with different numbers. It runs on a pricing model that charges per unit of text processed, bills input and output at different rates, and generates cost that varies with every single request. The organizations managing it well are not necessarily the ones with the most sophisticated tooling. They are the ones that built attribution first, understand the input-output cost split, and treat model tier selection as an ongoing routing decision rather than a one-time setup.
If your AI spend is currently showing up as a single line item on a monthly invoice, that is the starting condition, not the normal one. The tools to break it down by team, feature, model, and workload already exist. The question is whether anyone has put them to work.
Managing AI spend like cloud spend means treating tokens the way FinOps treats compute hours: meter them, attribute them, and optimize against the number you actually have, not the one on the rate card.
References
- FinOps Foundation. State of FinOps 2026 Report.
- Gartner, Inc. Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026 (May 19, 2026).
- FinOps Foundation. Token Economics: Managing AI Value in SaaS Model Token Costs.
- FinOps Foundation. FinOps X 2026 Day 1 Keynote: The Wild West of AI, Token Economics and the Evolving Role of FinOps.
- Anthropic. Pricing Documentation, Claude Platform Docs.
- OpenAI. Pricing, OpenAI API Documentation.
- Amazon Web Services. Introducing Granular Cost Attribution for Amazon Bedrock (April 17, 2026).
- Amazon Web Services. Track Amazon Bedrock Costs by Caller Identity with IAM Principal-Based Cost Allocation (April 17, 2026).
- Amazon Web Services. Understanding Your Amazon Bedrock Cost and Usage Report Data.
- Amazon Web Services. Amazon Bedrock Now Supports Cost Allocation by IAM User and Role.
- Stanford Institute for Human-Centered Artificial Intelligence. The 2025 AI Index Report.
- FinOps Foundation. State of FinOps 2025 Report.