OpenAI API Pricing Explained: Every Model, Every Hidden Cost
Last updated: October 2026
Short answer: OpenAI API pricing is pay-as-you-go, billed per 1 million tokens, with three prices on every bill: input, cached input, and output. The cheapest current model is GPT-6 Luna at $0.10 per 1M input tokens and $0.50 per 1M output tokens. The workhorse is GPT-6.1 Sol at $2.00 input and $10.00 output. The flagship GPT-6 Astra costs $10.00 input and $50.00 output. The number that matters most is output: on Sol and Astra, output tokens cost five times input tokens, so the words the model writes — not the prompt you send — usually decide your bill.
There is no subscription fee for API access and no per-request fee. You pay for tokens, and each model has its own rate card. This post covers the full current lineup from OpenAI's official pricing page, the extras that surprise first-time builders, and which model fits which workload. Run your own token mix through our AI API Cost Calculator before committing to any tier.
The Price Table: Every Current OpenAI Model
All prices below are USD per 1M tokens, from OpenAI's official pricing page. "Cached input" is the price for input tokens that match a previously seen prefix — OpenAI's prompt cache. The two "write" columns are what you pay to place tokens into the cache for 30 minutes or 12 hours.
| Model | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|
| GPT-6 Astra (flagship) | $10.00 | $1.00 | $50.00 |
| GPT-6.1 Sol (workhorse) | $2.00 | $0.10 | $10.00 |
| GPT-6 Luna (fast, cheapest) | $0.10 | $0.01 | $0.50 |
| GPT-5.6 Sol | $4.00 | $0.40 | $20.00 |
| GPT-5.6 Cyber | $12.50 | $1.25 | $75.00 |
| GPT-Rosalind Research | $5.00 | $0.50 | $25.00 |
| GPT-Rosalind Discovery | $5.00 | $0.50 | $25.00 |
Two things stand out. First, the gap between tiers is enormous: Luna's output costs one-hundredth of Astra's. Second, cached input is 10x to 20x cheaper than fresh input on every model. On Luna, a cached input token costs $0.01 per million — effectively free. If your workload repeats context, caching is the single biggest lever on your bill, which is why our prompt caching guide comes before any model comparison.
One deadline to note: GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026, per OpenAI's own page. If you are budgeting on the $4.00/$20.00 Sol rate, read our GPT-5.6 pricing deadline post before that date passes. The Rosalind research models are a special case: billing began October 5, 2026, and access is limited to approved research through OpenAI's trusted-access program — most developers will never see them on an invoice.
The Three Prices on Every Bill
1. Input tokens: what you send
Every token in your prompt, system instructions, tool definitions, and conversation history bills at the input rate. Long system prompts and big retrieved contexts are the usual reason a bill is higher than expected. On Sol, 1M fresh input tokens cost $2.00 — so a 100K-token prompt costs $0.20 before the model writes a word.
2. Cached input: what you repeat
When a request reuses a prefix OpenAI has already seen, those tokens bill at the cached rate instead. On Sol that is $0.10 per million versus $2.00 fresh — a 95% cut. Agent loops, which re-send repository context and tool definitions on every step, live or die by this line. The cache-write columns (for example, $2.50 for a 30-minute write on Sol) are the small fee for placing tokens into the cache in the first place.
3. Output tokens: what the model writes
Output is where bills explode, because output costs five times input on Astra and Sol, and five times input on Luna too ($0.50 versus $0.10). A model that explains itself at length, or an agent that reasons out loud for dozens of steps, can easily spend more on output than on everything else combined. Verbosity settings and step counts are budget settings, not just quality settings.
What a Real Month Costs: Two Worked Examples
Example 1 — a simple support chatbot on Sol. Suppose it handles 500,000 requests in a month, each with 2,000 fresh input tokens and 400 output tokens. That is 1,000M input tokens and 200M output tokens. Input: 1,000 × $2.00 = $2,000. Output: 200 × $10.00 = $2,000. Total: $4,000 — split evenly, even though output is only one-sixth of the token volume. Move the same workload to Luna and it costs 1,000 × $0.10 + 200 × $0.50 = $100 + $100 = $200, which is why simple, high-volume work should never sit on a flagship tier.
Example 2 — a coding agent on Sol with caching. Suppose it processes 10M input tokens per month, 85% of them cached, and produces 1M output tokens. Fresh input: 1.5M × $2.00 = $3.00. Cached input: 8.5M × $0.10 = $0.85. Output: 1M × $10.00 = $10.00. Total: $13.85. Run the identical workload on Astra and it costs 1.5M × $10.00 + 8.5M × $1.00 + 1M × $50.00 = $15.00 + $8.50 + $50.00 = $73.50 — more than five times as much. The full comparison of coding workloads across providers is in our Claude vs GPT vs Gemini coding cost post, and the per-model detail for Sol is in our GPT-6.1 Sol pricing post.
The Extras That Surprise People
Text tokens are only part of an OpenAI invoice. Voice and audio are priced separately: GPT-Live 1 voice sessions bill at $0.05 per minute, charged per second. The Realtime models are far pricier per token — GPT-Realtime 2.1 charges $32.00 per 1M audio input tokens and $64.00 per 1M audio output tokens, though cached audio input drops to $0.40. If you are building a voice agent, the audio rows, not the text rows, are your bill.
Built-in tools bill on top of tokens. Web search costs $10.00 per 1,000 calls, plus the search content tokens at model rates. File search storage is $0.10 per GB per day (1 GB free), with tool calls at $2.50 per 1,000 calls. Transcription is per minute: GPT-Transcribe at $0.0045 per minute, GPT-4o Transcribe at $0.006, and GPT-4o Mini Transcribe at $0.003. None of these appear in a text-only estimate, which is why a prototype's bill and a production bill can look like different products.
Finally, two surcharges to check before you budget. Regional processing (data residency) endpoints and FedRAMP endpoints each carry a 10% uplift on models released on or after March 5, 2026. And OpenAI renamed Priority processing to Fast mode on July 30, 2026 — if an old dashboard or invoice still says "Priority," it is the same Fast mode, not a separate product.
Which OpenAI Model Should You Pick?
For most production work, Sol is the default: $2.00/$10.00 is cheap enough for real traffic and strong enough for complex tasks. Luna is the pick for high-volume, simple work — classification, extraction, autocomplete-style steps, and triage — where $0.10/$0.50 makes costs disappear. Astra is for the hardest jobs only, where a failed answer costs more than a $50-per-million output rate. The cheapest model overall is not always the cheapest bill: a weak model that needs three retries costs more than a strong model that answers once. Our cheapest LLM API comparison ranks the same question across all providers, not just OpenAI.
Three Ways to Pay Less (Without Switching Providers)
First, turn on prompt caching and design for it: stable system prompts and repeated context are what make the $0.10 cached rate real. Second, route simple steps down to Luna and keep Sol or Astra for the steps that think — most agents waste flagship tokens on work a small model does identically. Third, measure before you scale: our 9 moves to reduce your OpenAI bill covers the full checklist, and our guide to calculating your AI API bill shows how to turn token counts into a monthly forecast.
FAQ
What is the cheapest OpenAI API model?
GPT-6 Luna, at $0.10 per 1M input tokens and $0.50 per 1M output tokens, with cached input at $0.01 per 1M. It is built for fast, high-volume work rather than the hardest reasoning tasks.
Is the OpenAI API a subscription?
No. API usage is pay-as-you-go per token and billed separately from ChatGPT subscriptions. There is no monthly API fee on standard pricing.
Why is my OpenAI bill higher than the input price suggests?
Almost always output tokens, which cost five times input on Sol and Astra, or tool charges (web search, file search, audio) that sit outside the text rates. Check your output volume and tool usage before blaming the input rate.
Does GPT-5.6 Sol's price change after November 21, 2026?
OpenAI lists its promotional pricing as available at least through November 21, 2026, with no follow-on pricing announced on the pricing page. Budget on the assumption that the $4.00/$20.00 rate may change after that date.
What is cached input, and how much does it save?
Cached input is text OpenAI has already seen in an earlier request — repeated system prompts, conversation history, or retrieved context. On GPT-6.1 Sol it bills at $0.10 per 1M tokens instead of $2.00 fresh — a 95% cut. Agent loops that re-send the same context on every step save the most.
Which OpenAI model is best for coding?
For most coding work, GPT-6.1 Sol at $2.00/$10.00 is the default — strong enough for complex tasks and cheap enough for real traffic. Use Luna at $0.10/$0.50 for simple steps like autocomplete and triage, and Astra at $10.00/$50.00 only for the hardest jobs where a failed answer costs more than the output rate.
How much does a small chatbot cost per month?
It depends on volume. Our worked example: 500,000 requests with 2,000 input and 400 output tokens each costs $4,000 a month on Sol — but only $200 on Luna, because Luna's rates are 20x lower. Match the model to the job before you scale.
What charges appear besides text tokens?
Voice and audio (GPT-Live 1 at $0.05 per minute, Realtime audio from $32.00 per 1M tokens), built-in tools (web search $10.00 per 1,000 calls, file search storage $0.10 per GB per day), and a 10% uplift on regional and FedRAMP endpoints for newer models. Check these before budgeting a voice agent or search-heavy app.
Sources: OpenAI API pricing page (developers.openai.com/api/docs/pricing), read 26 Oct 2026. All prices USD per 1M tokens unless stated per minute or per call. Prices change — re-check the official page before budgeting.