How to Compare Claude, OpenAI and Gemini API Costs

Updated September 21, 2026

This guide removes stale model prices, fixed exchange rates, undocumented client outcomes and universal savings claims. Its purpose is not to declare a permanent winner. It provides a decision system that can be rerun against current official prices and a fixed workload.

Quick answer

Sticker price is not a purchasing decision. Uncached input, cached input, output, tool calls, retries, latency and human review must be measured together. The useful comparison is total cost per accepted task after the quality gate.

Create a dated pricing record

Open the primary source on the day you run the calculation:

• OpenAI API pricing
• Anthropic Claude pricing
• Google Gemini API pricing

Do not treat a copied price as a future guarantee. Version the source timestamp with the calculation workbook.

Use a full-cost formula

API cost =
  uncached_input_tokens × uncached_input_rate
  + cache_write_tokens × cache_write_rate
  + cache_read_tokens × cache_read_rate
  + output_tokens × output_rate
  + tool_call_cost

Total task cost =
  API cost
  + review_time × loaded_labor_rate
  + rework_cost
  + infrastructure_and_observability_share

Cost per accepted task =
  total_cost_for_all_attempts / accepted_task_count

Compare the same task set

Build a workload that represents production traffic without personal or confidential data. Give each candidate the same input, tool permissions, output schema and timeout. If one model receives an extra prompt revision, repeat an equivalent optimization round for every candidate.

Blind reviewers to the model name where practical. Preserve the raw output, reviewer decision and reason for rejection. A low-cost failed output is not a saving.

Evaluate caching, batch and routing honestly

Caching is useful when a large block of stable context repeats. Batch processing belongs to latency-tolerant work. Multi-model routing can help when task classes are distinguishable, but its classifier, fallback call and operational complexity are part of the cost.

Streaming can improve perceived latency, but it does not by itself prove a lower bill. Stopping generation early saves cost only when the API actually records fewer billable output tokens and the result still passes the acceptance gate.

Publish a bounded result

Before publishing a savings claim, disclose the task set, dates, model IDs, pricing-source snapshot, token counters, acceptance rubric, human labor and excluded charges. Report uncertainty and failed attempts. Do not generalize one pilot to every company or workload.

The goal is not to crown the cheapest label. It is to choose a workflow whose sources are traceable, quality is accepted and total cost can be recalculated.