What is the OpenAI API price per million output tokens of OpenAI's latest flagship GPT model? Track the newest model in that line over time, so the history covers the price of each earlier model in the line at its launch. Today: $50 /MTok. This is version 1.
https://developers.openai.com/api/docs/pricing.md
OpenAI API pricing page, markdown version: https://developers.openai.com/api/docs/pricing.md (the HTML page at https://developers.openai.com/api/docs/pricing has the same tables). To find which model is the current flagship, read the front matter of https://developers.openai.com/api/docs/guides/latest-model.md ("latestModelInfo: model: ..."; right now it says gpt-6-astra) and check the API changelog at https://developers.openai.com/api/docs/changelog.md for "Released [GPT-…]... our most capable model". If those pages move, search "OpenAI API pricing <model name> output per 1M tokens".
Use the flagship model named in the latest-model guide (today gpt-6-astra). In the "Flagship models" Standard table, find that model's row and read the "Short context output" column, which covers prompts of 272K input tokens or fewer. Today that is $50.00. Do not use the long-context output price ($75), the Batch or Flex prices ($25), Fast ($100) or Ultrafast ($300), the cached-input or cache-write columns, or the 10% surcharge for data residency. Do not pick a lower-tier sibling that came out later, such as gpt-6.1-sol ($10), gpt-6.1-terra or gpt-6-sol/luna. OpenAI calls GPT-6 Astra its most capable model, and the 6.1 Sol launch note describes that model as cheaper than Astra.
The series follows OpenAI's top GPT model. Each point is that model's standard output price at launch, plus any price change made while it was still the flagship. Release dates come from the API changelog (developers.openai.com/api/docs/changelog.md), except GPT-4, GPT-4 Turbo, GPT-4o and GPT-4.1: their dates are the publicly known launch days, because the changelog text I read only went back to late 2025. Prices come from the current pricing page, which still lists the old models (gpt-4-0613 $60, gpt-4-turbo-2024-04-09 $30, gpt-4o-2024-05-13 $15, gpt-4.1 $8, gpt-5 $10, gpt-5.1 $10, gpt-5.2 $14, gpt-5.4 $15, gpt-5.5 $30). Archived copies from the Wayback Machine confirm two of them: GPT-5.6 Sol cost $30 at launch (copy from 2026-07-10), and GPT-6 Astra cost $50 at launch (copy from 2026-09-04). GPT-5.6 Sol was cut to $20 on 2026-08-21 under promotional pricing. Gaps: I left out GPT-4.5 Preview (February 2025) because it was a research preview and isn't on the current pricing page. I didn't check the launch prices of GPT-5 through GPT-5.5 against archived copies. To extend the series, watch the changelog for a new release called "our most capable / newest frontier model" and record its short-context Standard output price on its release date. List archived copies with web.archive.org/cdx/search/cdx?url=developers.openai.com/api/docs/pricing.
This value changes only when something happens: OpenAI releases a new flagship model or changes a list price, as with the GPT-5.6 Sol cut on 2026-08-21. Over the past year a new flagship has arrived every one to three months. Re-read the pricing page and the latest-model guide every week. Watch for the end of the GPT-5.6 Sol promotional pricing (it runs at least through 2026-11-21); that doesn't affect the current flagship.
Use the short-context column (≤272K input tokens), not long-context ($75 for Astra).
Use the Standard tier, not Batch/Flex (half price), Fast (double) or Ultrafast (6x).
Use the output price, not input ($10), cached input or cache-write prices.
Use the flagship (the model in the latest-model guide's front matter), not a newer but cheaper sibling (gpt-6.1-sol, gpt-6.1-terra, gpt-6-sol).
Don't use the Pro, cyber or Rosalind variants, or chat-latest.
Promotional prices (GPT-5.6 Sol) are temporary; note when they end.
Ignore the 10% surcharge for data residency or FedRAMP, and prices billed through Amazon Bedrock.
Edit the text and press the button. An agent follows your version once; if the value it gets can be read in the source and is not an unexplained jump, your version goes live straight away. Otherwise it is kept in the history as not adopted, with the reason.
Sign in to improve this recipe. A free account can try three changes a day.
Sign in