2026 AI API Cost Benchmark

Why Listed API Price Is Not the Final Product Cost

A provider's per-token, per-image or per-second price is the base unit, but real product costs diverge from it through several compounding factors:

• Input-to-output ratio — Prompt-heavy applications with short responses cost more per useful output than balanced request-response flows.
• Repeated conversation context — Multi-turn conversations re-send prior turns as context. The same 10-turn conversation can cost 5–10× a single-turn exchange.
• Coding-agent tool results — Each tool call returns data that is injected into the next model call as context. A single coding task can accumulate thousands of tool-result tokens.
• Cache write and read — Some providers charge for tokens stored as cache and credit cache reads at a significant discount. Using cache effectively can reduce input costs by 50–90%.
• Retries — Each retry doubles the cost of the original operation. Unhandled errors, timeouts and rate-limit backoff can multiply spend unexpectedly.
• Duplicate or overlapped jobs — A client timeout followed by a new submission creates two billed jobs instead of one.
• Failed or moderated media tasks — Some providers bill job creation regardless of outcome. Moderation blocks, invalid inputs and queue timeouts all create billable events.
• User rejection and regeneration — If users reject outputs and regenerate, each attempt is a new API call. Image and video workflows are particularly sensitive to this.
• Gateway and marketplace routing — OpenRouter and similar marketplaces add routing overhead and may charge differently from direct provider APIs.
• Billing reconciliation lag — Dashboard usage may not reflect billed amounts until hours after the request, making real-time cost monitoring difficult.

This calculator separates the official unit rate from a planning budget that multiplies the base cost by retry rate and a buffer. Always cross-reference with your actual organization usage data before committing a budget.

LLM API Cost

LLM API cost is driven by tokens. Every request sends input tokens (prompt + conversation context) and receives output tokens (completion). Prices are quoted per million tokens (M tokens).

monthly_requests =
  monthly_active_users
  × active_days_per_month
  × requests_per_user_per_day

input_cost =
  monthly_requests
  × average_input_tokens
  ÷ 1,000,000
  × input_price_per_million

output_cost =
  monthly_requests
  × average_output_tokens
  ÷ 1,000,000
  × output_price_per_million

base_cost = input_cost + output_cost

retry_adjusted_cost =
  base_cost × (1 + retry_rate_as_decimal)

planned_budget =
  retry_adjusted_cost × (1 + buffer_as_decimal)

Coding-Agent Cost

Coding agents cost more than simple chat requests because each completed task involves multiple model calls, large context windows and tool-result token accumulation.

Tool calls themselves are not free operations — they generate output tokens and inject result data into subsequent calls, increasing the input token count of every following request. Specific costs depend on the model, tool definitions, context window size, cache usage and provider billing rules. Use your actual usage records to calibrate these estimates.

monthly_model_calls =
  tasks_per_month
  × calls_per_task

total_input_tokens_per_call =
  prompt_tokens + tool_result_tokens

base_call_cost =
  (prompt_tokens + tool_result_tokens) ÷ 1,000,000 × input_price
  + output_tokens ÷ 1,000,000 × output_price

monthly_base_cost =
  monthly_model_calls × base_call_cost

effective_input_cost =
  (1 - cache_read_pct) × full_price
  + cache_read_pct × cache_discount_price

planned_budget =
  monthly_base_cost × (1 + retry_rate) × (1 + buffer)

Image Generation Cost

Image generation is usually billed per accepted output or per submitted job. The effective cost per accepted image is higher than the listed per-generation price because rejected or re-generated attempts also consume budget.

submitted_jobs =
  accepted_images × average_attempts_per_accepted

effective_cost_per_accepted =
  listed_price × average_attempts_per_accepted

monthly_generation_cost =
  submitted_jobs × listed_price

effective_output_cost =
  monthly_generation_cost ÷ accepted_images

planned_budget =
  submitted_jobs × listed_price × (1 + buffer)

Do not assume all failed or retried image generation jobs are billed or free. Inspect the provider's failure policy and check the billing dashboard before assuming a uniform treatment.

Video Generation Cost

Video generation is typically billed by generated duration (seconds) or per job. Async job lifecycle — creation, polling, completion, failure and webhook delivery — means that timeouts, retries and moderation blocks can all create billable events.

submitted_jobs =
  accepted_videos × attempts_per_accepted

total_generated_seconds =
  submitted_jobs × seconds_per_job

base_spend =
  (billing_mode == per-second)
    ? total_generated_seconds × price_per_second
    : submitted_jobs × price_per_job

planned_budget =
  base_spend × (1 + buffer)

Benchmark Methodology

Each data point on this page is classified as one of:

• Official fact — Directly quoted from the provider's current official documentation.
• Pricing snapshot — Verified from the official pricing page on the review date. Prices may have changed since then.
• Editable assumption — Default values in the calculator are illustrative planning estimates, not provider guarantees.
• Derived calculation — Results computed from the formulas above. Not audited by any provider.
• Provider-specific policy — Billing behavior that varies by provider, model and account tier.

Benchmark Data

All prices are pricing snapshots reviewed on June 17, 2026 from official provider pages. Prices change — always verify against the live provider pricing page before production use. Video generation models are deprecated or scheduled for shutdown; check current availability.

Download CSV

Export the benchmark data as a CSV file for use in your own cost model or spreadsheet. The file includes category, provider, model, billing unit, prices and source URLs.

Official Sources

The benchmark data on this page was verified against official provider documentation as of June 17, 2026. Pricing, model availability and API terms may have changed since then. Verify against the live documentation before making integration decisions.

• Anthropic API Docs: Pricing — Claude model rates and prompt caching
• OpenAI API Docs: Pricing — GPT-4o, o3, o4-mini, gpt-image-2, Sora 2
• Google Gemini API: Pricing — Gemini model rates
• OpenRouter Docs: Models — Marketplace routing and per-model rates
• Runway Developer API: Pricing — Video and image generation credits

Related Guides

Frequently Asked Questions