The cheapest token can produce the most expensive answer.
Compare the complete workflow bill: model usage, tools, storage, failed attempts, review and integration. Current catalogs are a starting point; account limits and actual task acceptance determine the buying decision. This report replaces uninspectable client-deployment and latency claims with a reproducible worksheet for your own workload.
What changed—and what the sources establish
OpenAI’s API price page separates usage categories by model and feature. Include applicable input, output and cached-input rates rather than treating one advertised token price as the entire bill. OpenAI API pricing.
Claude’s pricing documentation separately covers model usage and features such as caching and batch processing. Discounts have conditions; a workload must qualify and use the feature correctly to receive the assumed economics. Claude API pricing.
Kimi’s current model list includes K3, K2.7 Code and K2.6. A coding subscription or consumer plan is a different buying product from a direct API route, even when the underlying model family is familiar. Kimi API model list.
Choose the approach that fits the constraint
Choose the purchasing route from workload and operating requirements. A vendor shortlist should include acceptance conditions and exit costs, not just logos.
| Situation | Start here | Tradeoff / acceptance check |
|---|---|---|
| Individual knowledge work | Evaluate the relevant application subscription | Check sharing, export and workspace data controls |
| Application with variable demand | Direct API or an approved hosted route | Measure token mix, quotas and failure recovery |
| Large delay-tolerant batch | Supported batch processing | Confirm turnaround, limits and workload eligibility |
| Sensitive or specialized workload | Route satisfying actual control requirements | Review contractual terms before price ranking |
Turn a vendor quote into an auditable usage model
Gather request counts and input/output sizes from a representative period. Include the long tail: long documents, repeated tool results and customer sessions that require several attempts. Separate cached from uncached inputs only where the system actually achieves reuse. Model output costs can dominate an application that generates lengthy reports, while input dominates a document-processing pipeline. A single average token count hides both patterns.
List non-model charges and internal effort. Search calls, document ingestion, storage, orchestration, monitoring and human review can matter more than the first completion. Include minimum commitments, taxes or regional terms where they apply to the actual quote. Keep any promotional rate and its expiry separate from the normal rate. A business case that survives only an introductory discount needs an explicit renewal decision.
Measure latency in the experience customers receive
Record time to first useful output and time to completion separately. Streaming a greeting quickly does not establish that the answer or action is ready. For an agent, include retrieval, tool calls, queues, retries and approvals. Report tail latency and error frequency alongside median results. A fast model connected to a slow inventory service can still produce a slow support experience.
Run comparisons on the same task shapes and concurrency. Keep the region, access route, model settings and caching state visible. Do not mix a warm cached call from one vendor with a cold document upload from another. When a vendor reports speed, treat it as a vendor claim under its stated conditions; measure your own application before committing to a customer-facing service objective.
Write the purchase decision around accepted outcomes
A candidate must first pass data, feature and availability requirements. Then compare cost per accepted task, including escalation. Keep a separate severe-error gate for outputs or actions that the business cannot accept. Budget implementation and operational ownership, not only a monthly subscription. The smallest team may benefit more from a managed application with a reliable workflow than from a cheaper API requiring continuous integration work.
Add an exit test before signing a larger commitment. Can you export documents, evaluation data and useful configuration? Are model-specific parameters isolated? Can a second route satisfy the same output contract? Portability is not free, and maintaining several providers can add complexity. Decide which parts need an abstraction and which can remain vendor-specific because they provide a justified capability.
Worked example: illustrative assumptions, not a reported result
Illustrative monthly usage: 10 million uncached input tokens at an assumed $2 per million cost $20. Two million output tokens at an assumed $10 per million cost another $20. Add $100 for tools/storage and ten review hours at $40, producing a total planning cost of $540. These invented rates are not a current provider quote.
If the workflow completes 900 accepted tasks, total cost per accepted task is $0.60. Looking only at model usage would imply about $0.044. That gap explains why token-price comparisons can mislead. Recalculate with actual account rates, achieved caching, failed-task counts and review records before presenting the worksheet as a purchasing recommendation.
Failure modes to catch before scaling
- Promotional economics: a temporary price becomes a permanent forecast. Record expiry and normal-rate sensitivity.
- Latency theater: first-token speed replaces time to a useful completed outcome. Measure the entire tool chain.
- Missing review cost: cheap outputs require repeated correction. Track accepted-task cost and failure categories.
A purchasing report should be dated and reproducible. This guide contains no original latency benchmark, no verified survey of 37 deployments and no promise that a named provider will win your workload.
A concrete next step
Export a representative usage sample, complete a three-route cost sheet and test the same tasks against explicit acceptance checks. Include a renewal-price scenario and one exit exercise before committing to a broad rollout.
| Commercial | Current quote, usage categories and discount expiry |
|---|---|
| Technical | Exact model route, quota and output contract |
| Quality | Accepted outcomes, escalation and severe-error gate |
| Ownership | Integration, review and renewal decision owners |
Store the assumptions with the result: date, account plan, rates, task mix and review rate. Update the arithmetic when those inputs change. Make clear whether value means reduced cash spend, released capacity or additional revenue; those are different outcomes and should not be added without examining overlap.
Practical questions
Which token price should we compare?
Compare the actual mix of input, output, cached input and feature usage for the intended model and access route. Add tools, storage, retries and human review to calculate workflow cost.
Are consumer subscriptions equivalent to API access?
No. They are different products with different limits, billing and integration features. Confirm the exact route and plan the application or team will use.
What does a useful latency comparison measure?
Measure time to useful output and completion, including retrieval, tools, queues and retries. Use the same task shapes, concurrency and caching conditions and report tail latency and errors.
How often should the buying worksheet be updated?
Update it when prices, discounts, model mappings, workload mix or operating requirements change. A dated source check is not a permanent quote.
Primary sources checked 11 October 2026
- OpenAI API pricing — API pricing distinguishes model and usage categories, including input, output and cached input where supported.
- Claude API pricing — Claude API pricing distinguishes model, input/output tokens and caching or batch usage.
- Kimi API model list — Current official model documentation lists Kimi K3, K2.7 Code and K2.6.
The calculations in this guide are illustrative planning examples. Product documentation establishes supported features and conditions; it does not establish a measured return for your organization. Recheck account availability and release notes when implementing the recommendation.
Continue with Current model shortlist and Accepted-outcome ROI.
Research Path