OpenAI API Pricing: Which Model Should You Pay For?
OpenAI API costs vary sharply by model. The expensive mistake is using the frontier tier for every request. This breakdown separates premium reasoning, balanced production and lower-cost general workloads so buyers can route tasks instead of paying one rate for everything.
By Aniruddh PanvalkarPricing verified Oct 18, 2026
Plans and pricing
GPT-5.6 Sol
US$2 / 1M input, roughly US$0.20 cached input, US$10 / 1M outputOpenAI's frontier model for complex professional work, coding and high-value reasoning.
Best for
Hard coding, complex analysis and tasks where failure costs more than inference.
Not for
Routine extraction, classification, summaries or high-volume low-risk traffic.
GPT-5.6 Terra
US$2 / 1M input, roughly US$0.20 cached input, US$12 / 1M outputSibling GPT-5.6 tier at the same input rate as Sol with a slightly higher output rate; choose between them on your own evals.
Best for
General production, coding assistance and mixed-complexity business workflows.
Not for
Ultra-low-cost bulk jobs or the hardest reasoning tasks.
GPT-5.4 mini
US$0.75 / 1M input, cached input roughly one tenth of list, US$4.50 / 1M outputLower-cost GPT-5.4 tier with a 400K context window for high-volume, lower-risk workloads.
Best for
Summaries, extraction, classification, drafting and bulk agent steps that do not need frontier reasoning.
Not for
Hard coding, multi-step reasoning or anything where a wrong answer is expensive to catch.
GPT-5.6 and GPT-5.4 mini rates come from the OverpayingForAI catalogue sync on 2026-09-11. Tool calls and long-context surcharges may add cost; verify the official rate card before deployment.
How to avoid overpaying for OpenAI API usage
Do not make the frontier model your universal default. Start with the lowest-cost model that passes a written acceptance test, then escalate only failed or high-value tasks. Prompt caching and batch processing should be evaluated before negotiating volume discounts.
Calculate your exact cost →Who is overpaying for OpenAI API?
- Teams routing summaries, extraction and classification to a frontier model
- Products that ignore cached-input discounts on repeated system prompts
- Buyers comparing only input price while producing long outputs
Sources
Prices on this page are checked against these vendor pricing pages and catalogue entries.
- developers.openai.com/api/docs/models
- developers.openai.com/api/docs/models/compare
- openai.com/business/chatgpt-pricing
- openai.com/chatgpt/pricing
- openrouter.ai/openai/gpt-4-turbo
- openrouter.ai/openai/gpt-4o
- openrouter.ai/openai/gpt-4o-mini
- openrouter.ai/openai/gpt-5.4
- openrouter.ai/openai/gpt-5.4-mini
- openrouter.ai/openai/gpt-5.5
- openrouter.ai/openai/gpt-5.5-pro
- openrouter.ai/openai/o3
Frequently asked questions
Which OpenAI API model is the best default?
Use the least expensive model that consistently passes your real workload tests. At catalogue rates GPT-5.6 Sol ($2 / $10) and Terra ($2 / $12) are priced within a few percent of each other, so pick on your own eval results rather than the rate card, and route routine calls to GPT-5.4 mini ($0.75 / $4.50).
Does cached input reduce OpenAI API cost?
Yes. Current GPT-5.6 model pages list cached-input rates at one tenth of standard input pricing, making repeated prefixes and stable system prompts important cost levers.