The ChatGPT (OpenAI) API bills per token, and Standard-tier rates run from $0.05 per 1M input tokens on GPT-5-nano to $600 per 1M output tokens on o1-pro. The flagship GPT-6 Astra costs $10 per 1M input tokens and $50 per 1M output tokens. Batch and Flex processing cut both rates in half.
Our team at Cabina.AI tracks pricing for OpenAI directly, since we give developers access to GPT-6 Astra alongside 55+ other models through one subscription – when a rate changes, we see it firsthand. We noticed Standard-tier pricing across the board is exactly double what it was on September 11, 2026, while Batch and Flex pricing held at last week’s numbers, so a workload already running on Batch saw no increase at all.
ChatGPT API Pricing Table – All Models
OpenAI is the company that sets and bills every rate on this page; ChatGPT is the product name most searches use, but API traffic runs through the same OpenAI account and billing as every other product the company sells. Every chatgpt api pricing figure below is Standard-tier, per 1M tokens, current as of September 18, 2026. OpenAI still runs older GPT-4 and GPT-3.5 completion models in production behind their own separate rate cards; the table below covers the lineup that answers most cost questions today.
| Model | Input | Cached input | Output |
|---|---|---|---|
| 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 | $1.25 | $0.125 | $10.00 |
| gpt-5-mini | $0.25 | $0.025 | $1.00 |
| gpt-5-nano | $0.05 | $0.005 | $0.40 |
| gpt-4o | $2.50 | $1.25 | $10.00 |
| gpt-4o-mini | $0.15 | $0.075 | $0.60 |
| o1 | $15.00 | $7.50 | $60.00 |
| o1-pro | $150.00 | – | $600.00 |
| o3 | $2.00 | $0.50 | $8.00 |
| o4-mini | $1.10 | $0.275 | $4.40 |
GPT-5.6 Sol current rate is promotional and holds at least through November 21, 2026, so budget for a possible increase after that date. Regional data residency endpoints add a flat 10% on top of every row above for models released on or after March 5, 2026.
A few specialty model IDs sit outside the general chat and reasoning line. gpt-5-search-api ($1.25 input / $0.125 cached / $10 output per 1M) powers search-grounded responses as its own billable model, distinct from the $10-per-1,000-calls web search tool add-on described below. chat-latest ($5 / $0.50 / $30 per 1M) mirrors whichever model chat.openai.com is currently serving, for callers who want ChatGPT exact production behavior instead of a version they pin themselves.
ChatGPT API Cost Calculator
Estimate only. Excludes web search tool calls, tool/function-call token overhead, data residency uplift, and Codex or embeddings usage.
Pick a model and a processing tier, enter monthly input and output tokens, and the widget below turns that into an estimated chatgpt api cost. Rates are pulled straight from the table above, so switching the tier just applies the documented multiplier rather than a separate hand-typed number.
ChatGPT API Request Example
OpenAI current quickstart runs requests through the Responses API rather than the older chat completions endpoint. Here is a minimal authenticated call to gpt-6-astra:
curl https://api.openai.com/v1/responses \
-H "Content-Type: application/json" \
-H "Authorization: Bearer your-api-key-here" \
-d '{
"model": "gpt-6-astra",
"input": "Summarize this in one sentence."
}'
The Responses API also accepts Model Context Protocol connections and standard tool/function definitions in the same request body. Neither carries its own line item on the pricing page above; a tool definition and any content a tool call returns simply count as ordinary input tokens on whichever model is handling the request, so a heavily tool-using agent call can cost noticeably more than the same prompt without tools attached.
OpenAI coding-agent product, Codex, calls its own model, gpt-5.3-codex, priced at $1.75 input / $0.175 cached / $14 output per 1M tokens – a separate row from every general-purpose model above. A “codex api pricing” search should not be answered by looking up gpt-6-astra; Codex usage bills against its own rate, not the chat-oriented flagship line.
ChatGPT API Price and Model History
Standard-tier pricing across this entire table doubled between September 11 and September 18, 2026; every model that carried a rate a week ago now costs exactly 2x what it did, while Batch and Flex pricing simply kept last week’s numbers as the new discounted floor. GPT-6 Astra itself is a recent arrival, succeeding the GPT-5.6 family – Sol and Terra as the mid-tier pair, plus the smaller Luna variant beneath them – as OpenAI top model, and the same generational pattern runs through the reasoning line – o1 gave way to o3, and o3 now sits alongside the cheaper o4-mini rather than replacing it outright. OpenAI does not retire old model IDs quickly: gpt-4-0613 and multiple gpt-3.5-turbo variants from 2023 and 2024 are still billable today, years after gpt-4o and the GPT-5 family made them obsolete for most new projects.
Other OpenAI API Costs Beyond Chat
Chat and reasoning models are not the only thing OpenAI meters. Embeddings bill separately and far cheaper than any chat model: text-embedding-3-small costs $0.02 per 1M tokens and text-embedding-3-large costs $0.13. The older text-embedding-ada-002, still billable but superseded by the 3-series, costs $0.10. Whisper audio transcription bills by the minute instead of by the token, at $0.006 per minute, and the omni-moderation-latest content-moderation model is free to call.
OpenAI is also winding down its fine-tuning platform. It no longer accepts new fine-tuning jobs. Existing customers keep access through a limited transition window, and models already fine-tuned stay available for inference only until their underlying base model is retired.
Pricing Tiers for OpenAI, Explained
Batch and Flex Processing
Batch API requests, submitted asynchronously for jobs that are not time-sensitive, bill at exactly half of Standard-tier pricing across chatgpt models api pricing as a whole – gpt-6-astra drops from $10/$50 to $5/$25 per 1M tokens. Flex processing matches the Batch rate for most models and trades a lower price for higher latency rather than an async queue, so the choice between the two comes down to whether the workload can tolerate a delayed batch job or needs a live (if slower) response. The same 50% cut applies down the lineup – gpt-4o-mini goes from $0.15/$0.60 Standard to $0.075/$0.30 on either tier, which matters more in absolute terms on a high-volume nano-class workload than it does on a handful of GPT-6 Astra calls.
Fast Mode (Priority Processing)
OpenAI renamed Priority processing to Fast mode on July 30, 2026; either service_tier: "priority" or service_tier: "fast" still works in a request. Fast mode roughly doubles the Standard rate – gpt-6-astra runs $20/$100 per 1M tokens under Fast mode instead of $10/$50 – in exchange for lower latency across every request, short-context calls included. Fast mode is not available for GPT-6 Astra requests routed through EU data residency; those fall back to Standard processing instead. OpenAI models running on Amazon Bedrock bill through AWS rather than through OpenAI directly, at the same rate as the first-party API for equivalent services – useful for a team that already centralizes cloud spend on AWS and would rather not open a second billing relationship. Large accounts running enough volume to outgrow every tier above can also negotiate a custom enterprise agreement directly with OpenAI; the resulting rate is not published, so treat any specific enterprise discount figure quoted online as unconfirmed until OpenAI own sales team states it in writing.
The built-in web search tool adds its own charge on top of whatever model calls it: $10.00 per 1,000 searches, plus the tokens of whatever search content gets pulled back into the prompt, billed at that model’s normal input rate. It is easy to model a chat workload’s cost from the pricing table alone and miss this line entirely, since it only shows up once a request actually invokes the tool.
Prompt Caching on the ChatGPT API
Cached input is the one discount that does not follow a single fixed multiplier across the lineup. GPT-6 Astra cached input runs at $1 against a $10 standard rate, a 90% discount, while gpt-4o cached input is $1.25 against a $2.50 standard rate, only a 50% discount. From tracking this table update to update, that gap is easy to miss if a cost estimate assumes every OpenAI model caches at the same rate GPT-6 Astra does. A related, separate question is how much is a token in ChatGPT to begin with: OpenAI counts tokens with its own tiktoken encoding, which does not produce the same token count as Anthropic or DeepSeek tokenizer for identical text, so a per-token price is not directly comparable across providers without converting through actual token counts first.
GPT-6 Astra vs GPT-5-nano: Which OpenAI Model Should You Use?
GPT-5-nano is the cheapest way into the lineup at $0.05 input and $0.40 output, built for high-volume classification and extraction work, the kind where a smaller model already clears the bar. GPT-6 Astra costs 200x as much per input token and earns that premium on harder reasoning and long agentic runs, where a wrong answer downstream is expensive enough that token cost stops being the limiting factor. Between the two sits gpt-5, at $1.25 input and $10 output, a reasonable default for general-purpose chat work that does not need Astra ceiling or nano floor. In our view, the chatgpt api costs per model line up cleanly with that split, so the practical move is to prototype on gpt-5 or nano and upgrade a specific call path to Astra only once it demonstrably needs the extra capability.
| Model | Input / output per 1M | Reasonable fit |
|---|---|---|
| gpt-5-nano | $0.05 / $0.40 | High-volume classification and extraction |
| gpt-5 | $1.25 / $10.00 | General-purpose chat and coding assistance |
| gpt-4o | $2.50 / $10.00 | Multimodal input where vision matters more than latest reasoning |
| gpt-6-astra | $10.00 / $50.00 | Long agentic runs, hard reasoning, high-stakes output |
Is the ChatGPT API Free?
No. The API bills per token from the first request, and OpenAI pricing page does not currently list a signup credit for new developer accounts. That is a different question from how much do ChatGPT tokens cost inside the consumer app: ChatGPT.com own free tier gives casual users a limited amount of chat access at no charge, but that quota does not extend to API calls, and API usage on any of the models above is billed separately from the start.
ChatGPT Plus/Pro vs ChatGPT API: Which Is Cheaper?
ChatGPT Plus costs $20 a month for expanded access to the consumer app; ChatGPT Pro runs $100 to $200 a month depending on tier, though OpenAI temporarily paused new sign-ups for the $200 tier in September 2026 (confirm current availability before publishing). Neither plan translates cleanly into a per-token API rate, since the subscriptions bundle usage rather than metering it – a subscription tends to be cheaper for a single person’s day-to-day chat habit, while the API tends to win once an application needs to serve many users or automate a workload the flat-fee plans were never built to price.
If you would rather not commit to a single provider while testing, GPT-6 Astra is also available on Cabina.AI alongside Claude and 50+ other models under one subscription, useful for comparing outputs before committing to an API integration. See how it stacks up against Claude at Chatgpt vs Claude compare.
ChatGPT API vs Claude API Pricing
Reading chatgpt api pricing per million tokens next to Anthropic own table puts the two flagships close together: GPT-6 Astra costs $10 input / $50 output, and Claude Opus 5 costs $5 input / $25 output, so Astra runs roughly 2x Opus 5 at both ends. The gap narrows further down each lineup – GPT-5-nano at $0.05/$0.40 undercuts Claude Haiku 4.5 $1/$5 by a wide margin, while Claude most aggressive prompt-caching discount, 97.5% off on Fable 5.1, still beats Astra 90% cache discount on paper. Running both APIs side by side, the cheaper flagship on the input column is not always the cheaper one once real caching rates get applied, so a fair comparison has to price a specific workload rather than read the two headline rows next to each other.
| Model | Input per 1M | Output per 1M |
|---|---|---|
| GPT-5-nano | $0.05 | $0.40 |
| Claude Haiku 4.5 | $1.00 | $5.00 |
| GPT-6 Astra | $10.00 | $50.00 |
| Claude Opus 5 | $5.00 | $25.00 |
ChatGPT API Pricing FAQ
Is the ChatGPT API free?
No. Every request is billed per token from the first call. This is separate from ChatGPT.com own free chat tier, which does not cover API usage, and OpenAI does not currently list a signup credit for new API accounts.
How much does the ChatGPT API cost per 1M tokens?
- Standard-tier rates run from $0.05 input / $0.40 output on GPT-5-nano to $150 input / $600 output on o1-pro.
- The flagship GPT-6 Astra costs $10 input and $50 output per 1M tokens.
- Batch and Flex processing cut those rates in half;
- Fast mode roughly doubles them.
What’s the cheapest and most expensive OpenAI model?
GPT-5-nano is the cheapest at $0.05 per 1M input tokens. o1-pro is the most expensive, at $600 per 1M output tokens on the Standard tier.
Does OpenAI offer a caching or batch discount?
Yes to both. Cached input ranges from a 50% to a 90% discount depending on the model. Batch and Flex processing both cut the Standard rate in half; Fast mode charges roughly double instead.
What is the newest OpenAI model and how much does it cost?
GPT-6 Astra is the current flagship, at $10 per 1M input tokens and $50 per 1M output tokens on the Standard tier. GPT-5.6 Sol and Terra form the tier below it, with the smaller Luna variant beneath them; Sol current rate is promotional pricing that runs through at least November 21, 2026.
Is the ChatGPT API cheaper than ChatGPT Plus?
It depends on usage. ChatGPT Plus is a flat $20 a month for personal chat access. The API bills per token with no monthly cap, so it is usually cheaper for light, occasional use and more expensive than a subscription once a single heavy user’s token volume climbs high enough – the crossover point depends entirely on how many tokens that person actually sends.
