Token-based pricing: how AI usage billing works (2026)

In August 2026, OpenAI rewired billing for new ChatGPT Enterprise agreements. Per-message rates are gone: Chat now bills by tokens and model, Codex and ChatGPT Work meter the same way (Codex made the move back in April), and ChatGPT for PowerPoint started charging by token once its free period ended on August 6, 2026. GPT-6 Astra, launched September 3, runs $10 per million input tokens and $50 per million output tokens in Codex.

OpenAI’s Enterprise rate card says the change was made “so that usage pricing is standardized across all product surfaces.” Seats are a number you set once a year. Tokens are a meter that runs every second your products do, which means the budget owner who approved the contract in January has no lever on what it costs in September.

Token-based pricing: key facts

What is token-based pricing?

Token-based pricing is a usage-based billing model where AI providers charge for the text their models read and write, measured in tokens of roughly 4 characters each. About 1,000 tokens covers 750 words of English. As of 2026, it is the default billing model across OpenAI, Anthropic, and Google AI services.

Every request consumes tokens in two directions. Your prompt, plus any documents, conversation history, or system instructions riding along with it, bills as input. The model’s reply bills as output, at a higher rate. Our explainer on what AI tokens are breaks down the unit itself; the short version is you pay for every word in and every word out.

Providers landed on tokens because their costs scale with compute, and compute scales with text volume. The pass-through is honest. It’s also alien to how software budgets were built: nobody signs a contract for tokens. You sign for a model, and the tokens show up later, on the invoice.

The scale of adoption is why the term stopped being niche. When the three largest AI vendors all meter the same unit, token literacy becomes a finance skill, the way reading compute instances became one during the cloud decade.

How do tokens turn into dollars?

Every token-based bill reduces to one formula: total cost equals input tokens times the input rate, plus output tokens times the output rate, with rates quoted per million tokens (MTok). Providers that discount repeated content add a third meter, cached input, at roughly a tenth of the standard input price.

Run the math at Claude Sonnet 5’s September 2026 rate of $2 per million input tokens and $10 per million output. A 2,000-token prompt with a 500-token reply costs $0.004 in and $0.005 out: $0.009 total, less than a cent.

Individual requests round to nothing, so nobody watches them. At 100,000 requests a day, modest volume for a production chatbot, that same request costs $900 daily, roughly $27,000 a month, from a feature that looked free in testing. Cost per token matters at the sixth decimal place; cost per month is what reaches the board deck.

Estimating your own workloads takes minutes with an LLM cost calculator, and our LLM API pricing comparison shows the formula across every major provider. If the bill has already outrun every estimate, schedule a demo and see which features and customers are consuming the tokens.

How do you estimate token usage before the bill arrives?

You can estimate token usage three ways: word counts for planning, vendor tokenizer tools for precision, and the usage data returned with every API response for ground truth. English text averages about 1.33 tokens per word, so expected words times 1.33 gives a defensible planning number before a single request ships.

Tokenizers differ by vendor, and the difference has teeth. The same paragraph produces different token counts on OpenAI, Anthropic, and Google models, so a cross-vendor comparison built on one vendor’s counts will quietly mislead you. Every major provider publishes a tokenizer or token-counting endpoint; use each vendor’s own.

In production, the truth is in the response. Every API reply includes a usage breakdown listing input, cached, and output tokens for that request. Logging that field per feature and per customer is the cheapest instrumentation you’ll ever add, and it replaces month-end reconciliation with same-day visibility.

The forecast then follows from volume: tokens per request, times requests per day, times the per-million rate, split by input and output. Build it per workload. A support bot, a coding agent, and a document summarizer have wildly different input-to-output shapes on the same model.

For finance, the estimate doubles as a negotiation baseline. Enterprise agreements and committed-use discounts price against projected token volume, so a team that can forecast tokens per workload walks into a vendor conversation with real numbers instead of last quarter’s invoice and a guess.

Why do output tokens cost more than input tokens?

Output tokens cost 5x to 6x more than input tokens on the flagship and mid-tier models at every major provider as of September 2026, because generating text takes more compute than reading it. A model processes your input in parallel but writes its reply one token at a time, running a full forward pass for each one.

The ratio holds across flagship lineups. Anthropic prices output at 5x input from Haiku 4.5 at $1/$5 to Fable 5.1 at $10/$50. OpenAI’s GPT-5.6 Sol carries the same 5x at $4/$20. Google’s Gemini 3.1 Pro runs 6x at $2/$12. Budget tiers drift higher: Gemini 3.5 Flash-Lite’s output runs 8.3x its input at $0.30/$2.50.

Reasoning makes it worse. Thinking tokens bill as output on every major provider, and default reasoning budgets can burn tens of thousands of tokens per request before a single visible word appears. Most teams never change that default, which means the verbosity setting nobody owns is setting the output bill.

What do tokens cost across major providers in 2026?

Current AI token list rates span 50x on input alone, from $0.20 per million tokens on OpenAI’s GPT-5.6 Luna to $10 on GPT-6 Astra (launched September 3, 2026) and Claude Fable 5.1. All rates are standard API list prices in USD per million tokens.

Promotional rates hide in that table. GPT-5.6 Sol’s $4/$20 reflects a cut OpenAI shipped in August 2026, 20% off input and 33% off output, guaranteed at least through November 21, 2026; the standard rate is $5/$30. Gemini 3.8 Flash’s $0.75/$3.75 is introductory pricing that doubles to $1.50/$7.50 on January 1, 2027, per Google’s published schedule.

Two footnotes cut the other way. Anthropic scheduled Sonnet 5’s introductory $2/$10 to rise to $3/$15 on September 1, 2026, then cancelled the increase and made the low rate permanent, a change now reflected on Anthropic’s pricing page. And Gemini 3.1 Pro doubles input to $4 and lifts output to $18 the moment a single prompt exceeds 200K tokens of context.

Three of those repricings occured within ten days: Sonnet 5 made permanent on August 11, Gemini 3.8 Flash introductory pricing effective August 13, and the Sol cut on August 21. The Gemini 3.1 Pro context tier is not a repricing at all, it is a standing surcharge.

Our guides to OpenAI API pricing, Claude pricing, Google Vertex AI pricing, and Azure OpenAI pricing track each vendor’s rate card.

Which discounts and multipliers change the effective rate?

List price is where a token bill starts, and almost never where it ends. As of 2026, every major provider layers discounts and surcharges onto the base rate, and for many workloads these modifiers move the bill more than model choice does.

Stacking is the real lever. Caching pays for itself fast: on Anthropic’s terms, a 5-minute cache write costs 1.25x input and every read costs 0.1x, so the cache breaks even after a single hit, while the 1-hour tier at 2x breaks even on the second read. Batch halves whatever remains, and the two discounts combine by design.

Batched traffic with heavily repeated context can run at a small fraction of list price, which is why two teams on the same model can carry wildly different effective rates. The gap between them isn’t negotiating skill. It’s workload design.

And tokens aren’t the only meter. OpenAI bills web search at $10 per thousand runs and voice at $0.05 per connected minute on top of model tokens. Google meters generated video two ways: Omni Flash bills 5,792 tokens per second of 720p output, while Veo 3.1 charges $0.40 per second outright.

Why is token-based AI spend hard to forecast?

Token-based AI spend resists forecasting because the workload sets the bill, not the contract. A seat license costs the same whether an employee works 4 hours or 40. Token pricing has no floor and no ceiling. Usage patterns, prompt design, and model behavior all move the number, and most of them move it silently.

Context drift is the one nobody watches

One user request can fan out into dozens of model calls. Retrieval pipelines attach larger documents. Conversation history accumulates. Each addition rides along as input on every single request, so a chatbot that cost $0.009 per exchange in January can cost double that by June with identical traffic.

Agentic workloads multiply the problem

One user request can fan out into dozens of model calls as an agent plans, uses tools, checks its work, and retries. Multi-agent systems run several context windows in parallel, and reasoning tokens bill at output rates the whole way. Our guide to AI agent pricing models covers why per-task costs vary so widely.

When Anthropic shipped Claude Opus 4.7 in April 2026, it introduced a tokenizer now used across Claude 4.7 and later models. Anthropic says it produces approximately 30% more tokens for the same text. The rate card never moved, but identical workloads cost more.

A price increase with no price change is something traditional procurement has no playbook for.

Model routing adds one more moving part. Well-run teams send routine traffic to cheap models and escalate only hard cases to flagships, which is smart engineering and hard forecasting: the blended rate per request now depends on the routing mix, and the mix shifts as prompts, models, and traffic evolve.

The result is a line item where variance is structural. Getting ahead of it starts with AI usage tracking granular enough to explain why the number moved.

How does token pricing compare to seats, credits, and outcome models?

Token pricing is one of four models competing to bill AI in 2026, and most companies now pay through several at once. Seat pricing charges per user per month, predictable but disconnected from usage; heavy and idle users cost the same. Credit systems sell prepaid units that abstract over tokens, smoothing procurement but hiding the underlying meter.

Outcome and per-action pricing charges for completed work, a resolved ticket or an executed task, and it’s gaining ground for agents where buyers want to pay for results rather than compute. The trade-off inverts: your costs become predictable per unit of value, while the vendor absorbs the token variance.

In practice, most enterprises now run a hybrid stack without ever deciding to. Employees hold seat subscriptions to chat assistants, product teams pay token meters on APIs, and procurement holds prepaid credit commitments with one or more providers. Each contract looks manageable alone. Together they form an AI spend portfolio that no single invoice describes.

The direction of travel is toward the meter. When OpenAI moved ChatGPT Enterprise off per-message rates in August, it conceded that even seat-adjacent products settle up in tokens.

Our overview of AI pricing models maps the full landscape, but the practical reality for finance is simpler: whatever the packaging says, tokens are underneath it.

How can finance teams govern token-based AI spend?

Governing token-based spend means treating it like the cloud meter it is: measure it, allocate it, and tie it to unit economics before it grows. Waiting for the invoice means managing with a 30-day lag on a cost that compounds daily.

Start with allocation. AI token cost only becomes actionable when it’s attributed to a product, feature, team, or customer. Aggregate spend tells you the number went up; allocated spend tells you the new agent feature costs $0.40 per user session while the pricing page charges $0.30. That’s a margin conversation, not an infrastructure one.

Set token budgets before you need them. Per-feature and per-team token budgets with alerts catch looping agents and context drift in hours instead of at month-end. One CloudZero customer running more than 50 LLMs in production found over $1 million in immediate savings by optimizing inference workloads and leaning on token caching, alongside a 50%+ reduction in compute spend.

Then connect spend to revenue. Cost per customer, cost per request, and cost per feature turn an unpredictable meter into unit economics a CFO can plan around, the same discipline covered in our guide to LLM cost management.

Companies like Skyscanner, Upstart, and Drift use CloudZero to see exactly which products and customers drive their cloud and AI spend. CloudZero is the AI ROI company. If token-based bills are already surprising you, request a demo to see your token spend allocated down to the feature and customer level.

Frequently asked questions about token-based pricing

Author Spotlight

Lyne Carolyne

Lyne Carolyne has several years of experience in AI and cloud economics and brings that understanding into the content she creates. Outside work, she's an avid explorer.