Commercial AI Token Costs: Budgeting Beyond List Price

Your spreadsheet says one model is cheaper. Your invoice says otherwise. The gap appears because the spreadsheet priced the prompt and final answer, while production also paid for reasoning, repeated instructions, failed tool calls, retries, discarded drafts, and cache behavior.

If you are choosing a commercial AI model or defending an AI budget, compare cost per accepted outcome, not cost per million tokens. That change turns a rate card into a forecast you can actually use.

A token price is only one layer of your production cost

Published input and output prices tell you the rate applied to certain tokens. They do not tell you how many tokens the model will consume before your application gets an acceptable result. A useful cost model therefore has three layers:

• Unit rates: the applicable prices for input, output, reasoning, cache reads, cache writes, and any long-context tier.
• Consumption: the number of tokens used by the prompt, retrieved context, system instructions, tool definitions, intermediate reasoning, and response.
• Completion efficiency: how many attempts, revisions, and tool calls you pay for before the result passes your acceptance checks.

The third layer causes many budget misses. A cheap attempt is not a cheap task if the attempt is rejected and repeated. Nor is a successful API response necessarily a completed business task. A coding agent that returns malformed code, a content model that produces an unusable draft, or a schema generator that fails validation has consumed tokens without delivering the outcome you intended to buy.

In measured 2026 production usage, the categories commonly omitted from simple estimates represented 52.5% of billed tokens and added 70.4% above a list-price-only estimate. These percentages are not universal overhead rates. They are a practical checklist of what your own logging needs to capture.

Do not solve this by applying one generic markup to every vendor quote. Instrument each category instead. A reasoning-heavy model, a tool-using agent, and a short classification call can have radically different overhead even when their visible prompts look similar.

Falling rate-card prices do not remove this problem. Within a constant-capability mid-tier series from Q1 2023 through Q3 2026, the list-price index fell 91.4%, but real cost per completed task fell only 62.9%. Token consumption per completed task rose 4.3 times. The completed-task cost reached its low point in Q4 2024 and then increased 80% by Q3 2026 even as published rates generally continued downward. More capable reasoning behavior can consume part of the saving advertised on the price sheet.

Compare models by accepted task, not by token rate

A model comparison becomes useful only after the denominator represents something your business accepts. From May 4 through August 21, 2026, a standardized set of 14 production tasks was run across 11 commercial models. The resulting cost included billed reasoning, prompt repetition, cache activity, retries, and discarded output. The September 2026 prices and measured completed-task costs show why rate-card ranking and production ranking can diverge.

Several reversals matter when you shortlist a model. GPT-5.4 mini had lower published input and output prices than Claude Haiku 4.5, yet its measured task cost was $0.0627 versus $0.0474. Claude Sonnet 5 had higher published rates than Gemini 3.6 Flash but completed the task set for $0.0848 instead of $0.1040. At the frontier end, GPT-5.6 Sol and Claude Opus 5 shared the same $5.00 input price, but Sol cost 62% more per completed task, with the difference driven almost entirely by output volume.

These results do not make one model universally cheaper. Your prompts, tools, input-to-output ratio, quality threshold, and retry policy may reverse the ranking again. Use published comparisons to choose candidates, then reproduce the comparison on your own workflow.

1. Define completion before testing. For JSON-LD, completion might require parsable output that passes your validation checks. For a content brief, it might require every mandatory field and entity. An HTTP success code is not an acceptance criterion.
2. Freeze a representative task set. Give every candidate the same source material, system instructions, tools, output requirements, and acceptance tests.
3. Record every billable attempt. Keep rejected generations, malformed output, repair prompts, tool-call failures, and fallback calls in the numerator.
4. Separate visible output from total usage. Store every usage field the provider exposes, including reasoning and cache categories where available.
5. Compare only models that meet the quality gate. A low-cost result that cannot be used is a failed attempt, not a bargain.
6. Divide total model spend by accepted completions. That figure is your effective task cost and the basis for a credible monthly forecast.

Content costs multiply after the first draft

Content teams often estimate AI spend from the tokens in one draft. That calculation stops before the expensive part: revisions, replacement drafts, citation repair, structural fixes, and output that never reaches publication.

For 1,000 words of finished, publishable copy, the measured token cost included revision rounds and discarded generations. The difference between first-draft and finished cost was substantial across every tested model.

The cheapest and most expensive first drafts were separated by roughly 25 to 1. After revisions and discards, finished costs were separated by about 10 to 1. Draft rejection narrowed the apparent advantage of the cheapest models.

Discard rate was also more useful than list price for anticipating finished cost. Claude Sonnet 5 started at $0.032 per 1,000 words, above Gemini 3.6 Flash at $0.024. Sonnet finished lower, at $0.074 versus $0.084, because its discarded-draft rate was 13% rather than 26%.

Build that distinction into your content operations. Give every generated asset a final status such as accepted, revised, or discarded, and associate all attempts with the same job identifier. Then calculate finished token cost from all spend attached to accepted copy, divided by accepted word count and multiplied by 1,000. Counting only the last successful generation erases the waste you are trying to manage.

Keep the quality gate explicit. For an SEO or GEO workflow, your requirements may cover factual accuracy, source support, search intent, entity coverage, structure, brand constraints, and valid structured output. The exact rubric is yours, but it must be stable across models. Otherwise, a permissive review process can make a weak model look artificially inexpensive.

The figures above cover model-token spend. They do not represent a fully loaded content cost. Your internal budget should add editorial review, fact-checking, workflow infrastructure, monitoring, and any human repair work rather than treating a low token figure as the total cost of publication.

Budget by workload, then route each job to the right tier

A single company-wide average hides the workflows most likely to break your budget. Agentic coding, customer support, retrieval-based research, document processing, sales personalization, and content production have different volumes, context sizes, output patterns, and failure modes.

For a modeled 50-person company, the same mix of 157,400 monthly tasks cost $6,610 at the economy tier, $19,150 at the mid tier, and $48,670 at the frontier tier. That is a 7.4-times spread before changing the workload itself.

Volume alone does not reveal the expensive workflow. The coding agent ranked fourth by task count but was the largest monthly cost. At the frontier tier, it cost $1.24 per completed task, compared with $0.16 for customer support. Agentic workflows repeatedly call models, tools, and validation steps, so a task can contain much more billable activity than one support interaction.

Build your forecast from accepted workload volume

Your budget sheet should have one row per distinct workflow, not one row per provider. Separate content briefs from finished drafts, retrieval answers from document ingestion, and schema generation from schema repair. They may use the same API while having different cost behavior.

• Workload identity: team, application, task type, model, and model version.
• Demand: expected completed tasks, not merely API requests.
• Usage: input, output, reasoning, cache-read, and cache-write tokens where exposed.
• Workflow overhead: attempts, tool calls, validation failures, fallback calls, and discarded results.
• Outcome: accepted, repaired, rejected, or abandoned.
• Unit economics: total billed spend divided by accepted completions.

Forecast monthly model spend by multiplying expected accepted-task volume by your measured cost per accepted task. Keep the rate-card calculation beside it as a reconciliation check, not as the primary forecast. A widening gap between the two tells you to investigate prompt growth, longer retrieved context, increased reasoning, lower cache reuse, tool failures, or a rising retry rate.

Recalculate after changes to the model version, system prompt, tool schema, context strategy, output format, or acceptance threshold. Each can alter consumption or completion efficiency even when the published token rate stays fixed.

Use routing instead of choosing one model for everything

Model tier should be a workload decision. Economy models are strongest candidates when the task is constrained, output can be checked automatically, and failure is cheap to retry. Mid-tier models suit broader production work where reliability and cost both matter. Frontier models deserve the jobs whose ambiguity or quality requirement produces a measurable improvement worth their higher completed-task cost.

That does not require moving every workflow downmarket. In the modeled company, moving only the two highest-volume workloads – customer support and sales personalization – to economy models while leaving the other four at the frontier tier reduced total monthly spend by 26%. Selective routing captured savings without imposing one capability tier on every task.

Put a quality gate after the lower-cost route and send only failed or uncertain cases to a stronger model. Count both calls when escalation occurs. Otherwise, the first model appears cheaper in your dashboard while the fallback cost disappears into another service or team.

Key takeaways

• Published cost per million tokens is a unit rate. Your actionable metric is total billed spend per accepted task.
• Log reasoning, repeated system context, cache activity, retries, discarded output, tool failures, and long-context pricing instead of hiding them in a generic contingency.
• For content, calculate cost per 1,000 accepted words from every draft and revision associated with the finished asset.
• Benchmark candidates on the same tasks and acceptance criteria. Compare costs only among models that clear the required quality threshold.
• Route by workload. High-volume, tightly validated tasks may justify an economy model, while ambiguous or high-impact work may justify a more capable tier.
• Refresh the forecast whenever the model, prompt, tools, context, output contract, or quality gate changes.

Start with one workflow that already generates meaningful volume. Attach every billable attempt to an accepted or rejected outcome, calculate its effective cost, and use that result to challenge the rate-card estimate. Once the accounting works for one workflow, extend the same measurement to the rest of your AI stack and route each task on evidence rather than model reputation.

References

• First Page Sage Blog – AI Token Usage Cost by Model: OpenAI vs. Anthropic vs. Google