ONLINEAGENT_OPS 2026.Q3 HOME ARTICLES CRAFT RECORD BLOG MAP HUBS FAQ SEARCH
HOMEARTICLESWHAT THE OPENAI API ACTUALLY COSTS
ARTICLES · COMPARISON

What the OpenAI API Actually Costs

Four things decide your OpenAI bill and only one is the headline price: the model ladder, the long-context cliff, caching, and the spread between service tiers.

READ4 min
WORDS762
SECTIONS5
TYPECOMPARISON
CHECKED16 SEP 26

Four numbers decide your OpenAI bill, and only one of them is the headline price. The other three are context length, whether the input was cached, and which service tier you sent it on. Each can double or halve the cost of the same request.OpenAI, Pricing, read at source 16 Sep 2026: “Prices per 1M tokens.” Every figure on this page uses that unit.

TL;DR — THE SHORT VERSION

The list price is the worst case. Three levers move it.

  • GPT-6 Astra is the flagship: $10 in, $50 out per million tokens.OpenAI, Pricing, read at source 16 Sep 2026: gpt-6-astra short context input “$10.00”, output “$50.00”.
  • Long context doubles the input and adds half to the output — $20 and $75 on the same model.OpenAI, Pricing, read at source 16 Sep 2026: gpt-6-astra long context input “$20.00” and output “$75.00”.
  • Cached input is a tenth of the price, which is the biggest single saving on the page.
  • Batch halves everything. Fast doubles it.OpenAI, GPT-6 Astra model page, read at source 22 Sep 2026: “Batch and Flex are priced at 50% of Standard rates. Fast mode is priced at 2x the applicable rates.”
  • The cheap models are very cheap: Luna is $0.20 in, and nano is five cents.OpenAI, Pricing, read at source 16 Sep 2026: gpt-6-astra long context input “$20.00”, output “$75.00”; gpt-5.6-luna “$0.20”; gpt-5-nano “$0.05”.
◈ PRICES ON THIS PAGE

Every price here was correct when this site last checked it, between 16 Sep and 22 Sep 2026. Prices change often and this site no longer updates them, so check the vendor’s own page before you rely on one: OpenAI pricing · Google pricing · xAI pricing.

The model ladder

Standard tier, short context, US dollars per 1M tokens. Read at source 16 Sep 2026
ModelInputCached inputOutput
gpt-6-astra$10.00$1.00$50.00
gpt-5.6-sol$4.00$0.40$20.00
gpt-5.6-terra$2.00$0.20$12.00
gpt-5.6-luna$0.20$0.02$1.20
gpt-5-nano$0.05$0.005$0.40

OpenAI, Pricing, read at source 16 Sep 2026, from the table headed “Standard pricing data”. The page lists more than thirty models; these are the current family plus the cheapest option still sold.

INPUT PRICE · PER MILLION TOKENS
The ladder above drawn to scale, standard tier, short context, read at source 16 Sep 2026. The two cheapest bars barely register next to Astra.
gpt-6-astra$10.00
gpt-5.6-sol$4.00
gpt-5.6-terra$2.00
gpt-5.6-luna$0.20
gpt-5-nano$0.05
OpenAI, Pricing, read at source 16 Sep 2026: gpt-6-astra short context input “$10.00”; gpt-5.6-luna “$0.20”; gpt-5-nano “$0.05”.

A caveat on the nano row. The snapshot behind gpt-5-nano is scheduled to shut down on 11 December 2026, and OpenAI names gpt-5.6-luna as its replacement. Budget on Luna’s price, not nano’s, for anything meant to outlast the year.OpenAI, GPT-5 nano model page, read at source 22 Sep 2026: “Default snapshot: gpt-5-nano-2025-08-07” · OpenAI, Deprecations, read at source 22 Sep 2026: shutdown date “Dec 11, 2026” for gpt-5-nano-2025-08-07, recommended replacement gpt-5.6-luna.

TAKEAWAY

Astra costs 200 times what nano costs on input. Almost no workload needs the top of the ladder for every call, and the gap is wide enough that routing cheap work downward is the whole optimisation.

The long-context cliff

OpenAI publishes two prices for every current model: one for short context, one for long. On the long one, input and cached input double and output rises by half.

GPT-6 Astra, standard tier, per 1M tokens
Short contextLong context
Input$10.00$20.00
Output$50.00$75.00
Cached input$1.00$2.00

OpenAI, Pricing, read at source 16 Sep 2026.

Long starts above 272,000 input tokens, and then the whole request is billed at the higher rate. The pricing page only labels the columns; the threshold is on the model pages. So one long document does not just cost more for its own tokens — it moves every token in that request, output included, into the right-hand column.OpenAI, GPT-6 Astra model page, read at source 22 Sep 2026: “Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request.” The GPT-5.6 Luna model page, same day: “Prompts with >272K input tokens are priced at 2x input and 1.5x output for the full request.”First-hand: until 22 Sep 2026 this paragraph said the pricing page does not say where short becomes long. That was true of the pricing page; the model pages state it.

Google and xAI do the same thing at 200,000 tokens, which is covered on the Gemini and Grok comparison.Google, Gemini Developer API pricing, read in a browser 16 Sep 2026 (“$4.00, prompts > 200k tokens”) · xAI, Pricing, read at source the same day.

The lever most people ignore

Cached input costs a tenth of fresh input on every current model: $1.00 against $10.00 on Astra, two cents against twenty on Luna. If your prompts share a long stable prefix — a system prompt, a document, a tool schema — that is the difference between a viable product and an expensive one.

There is a catch the table makes plain. Writing to the cache costs more than the input itself: $12.50 against $10.00 on Astra. So caching pays when a prefix is reused, and costs you when it is not.OpenAI, Pricing, read at source 16 Sep 2026: gpt-6-astra cached input “$1.00” against input “$10.00”, cache writes “$12.50”; gpt-5.6-luna cached input “$0.02” against “$0.20”.

TAKEAWAY

Cache a prefix you will send again. Caching something you send once is strictly worse than not caching it — the write costs more than the read would have.

Batch, Flex and Fast

The same model has four prices depending on how you send the request.

GPT-6 Astra input price by service tier, per 1M tokens
TierInput
Batch$5.00
Flex$5.00
Standard$10.00
Fast$20.00

OpenAI, Pricing, read at source 16 Sep 2026, from the tables headed “Batch pricing data”, “Flex pricing data” and “Fast pricing data”.OpenAI, GPT-6 Astra model page, read at source 22 Sep 2026: “Batch and Flex are priced at 50% of Standard rates. Fast mode is priced at 2x the applicable rates.”First-hand: until 22 Sep 2026 this note called the half-and-double ratio an observed pattern, because the pricing page does not state it as a rule. The Astra model page does.

That is a fourfold spread on identical output. If the work is not interactive — overnight summarisation, backfills, evaluation runs — the cheapest tier is the same model.

Free or paid? These are API prices; for ChatGPT’s own plans see free vs paid AI chat.

What this page could not verify

  • Whether every model uses the same 272,000 threshold. Read on the Astra and Luna model pages only.OpenAI, GPT-6 Astra model page, read at source 22 Sep 2026: “more than 272K input tokens”.
  • What Batch and Flex cost you in latency. Published elsewhere, not read here.
  • Real spend for any workload. Price per token says nothing about how many tokens a model uses.
  • Enterprise terms. Not published.
SOURCES

OpenAI, Pricing, read at source on 16 September 2026. The markdown version of that page was used, which OpenAI offers by appending .md to any docs URL.

API prices are the most volatile figures on this site. Re-read before you budget.

ABOUTMETHODVERIFYPRIVACYCONTACTINDEXAI PROMPT GENEER · EVERY ARTICLE CARRIES ITS OWN CHECKED DATE