OpenAI Doesn’t Have a Price. It Has Dozens.
Ask ten people what OpenAI costs and you’ll get ten answers, most of them correct. A freelance designer paying $20 a month for ChatGPT and a startup spending $6,000 a month on the API are both OpenAI customers, and the two bills have almost nothing in common beyond the logo on the invoice.
That split is where the confusion begins. Subscriptions are flat fees you can predict. API usage is metered by the token, priced differently for every model, and shaped by decisions your developers make at 11pm. Understanding both is the difference between a budget you can defend and a monthly surprise.
The Subscription Side: Flat Fees, Clear Tiers
ChatGPT plans are the simple half. You pick a tier, you pay monthly, and the number doesn’t move. As of writing, the lineup looks like this:
- Free costs nothing, with limited access to the newest models, capped usage, and slower responses at peak times.
- Plus at $20/month is the default for individuals: higher limits, access to advanced models, image generation, voice, and file analysis.
- Pro at $200/month targets heavy users who hit Plus limits regularly, with substantially higher caps and priority access to the most demanding features.
- Team runs roughly $25–30 per seat per month, depending on whether you pay annually or monthly, and adds a shared workspace, admin controls, and a guarantee that your data isn’t used for training.
- Enterprise is quoted per organisation and adds compliance, SSO, and dedicated support.
The practical question for most teams isn’t “which tier” but “how many seats.” Five Plus accounts at $20 each is $100 a month, and it’s often the cheapest way to test whether AI fits your workflow at all. There’s a fuller look at what each plan actually unlocks in this guide to ChatGPT and OpenAI in 2025, including which features are genuinely worth paying for.
The API Side: You’re Buying Tokens, Not Access
The API bills by token, which is roughly three-quarters of a word in English. Every request has an input (your prompt, system instructions, conversation history, attached documents) and an output (the model’s reply). Both are counted, and output is consistently the more expensive half.
Approximate rates per million tokens
Numbers shift every few months, so treat these as order-of-magnitude guides rather than gospel. At the time of writing, a broadly capable flagship model like GPT-4o sits around $2.50 per million input tokens and $10 per million output tokens. The mid-tier models land near $0.40/$1.60 to $2/$8 depending on generation. The small, fast models drop to roughly $0.10–$0.15 input and $0.40–$0.60 output. Reasoning models that “think” before answering cost several times more per token, because they generate hidden reasoning tokens you pay for but never see.
Embeddings are almost free by comparison, around $0.02 per million tokens for the small model. Image generation and realtime audio are priced on entirely different scales again, often per image or per audio minute.
Why output costs more than input
Input tokens are processed in parallel, which is efficient. Output tokens are generated one at a time, sequentially, which is expensive. A model that writes 2,000 words costs far more compute than one that reads 2,000 words. If you ever wonder why a summarisation feature feels cheap and a long-form writing feature feels expensive, that asymmetry is the answer.
What a Real Monthly Bill Looks Like
Concrete math beats abstract rates. Imagine a customer support bot handling 500 conversations a day. Each conversation sends about 1,200 input tokens (system prompt, help-desk context, conversation history) and gets back roughly 300 output tokens.
That’s 18 million input tokens and 4.5 million output tokens per month. On a flagship model at $2.50/$10, you’d pay about $45 in input and $45 in output: roughly $90 a month. Run the identical workload on a small model at $0.15/$0.60 and the bill drops to about $5.40. Same product, same users, a 17x difference in cost.
That gap is the single most important fact about OpenAI pricing. Model choice dominates everything else, which is why it’s worth mapping which OpenAI model actually fits your job before you start optimising anything else.
Where Bills Quietly Balloon
Most overruns don’t come from a single dramatic mistake. They come from a handful of small patterns compounding.
- Resent context. In a multi-turn chat, the entire history is re-sent with every message. A 20-message conversation might cost twenty times what a single message does, even though the user typed far less.
- Bloated system prompts. A 2,000-token instruction block attached to every single call across a million calls is 2 billion tokens of pure overhead.
- Retries and timeouts. Every transient failure and automatic retry is a paid request. Poorly handled rate limits can double your token count without producing a single extra answer.
- Re-embedding everything. Rebuilding a vector index from scratch each time a document changes gets expensive fast at scale, even at $0.02 per million tokens.
- Fine-tuning overhead. You pay for training runs, then an hourly storage fee for the tuned model, then a premium on every inference call. It’s rarely worth it unless prompt engineering has genuinely hit a wall.
Legitimate Ways to Cut the Bill in Half
OpenAI offers real discounts, and most teams leave them on the table.
The Batch API processes requests within a 24-hour window at a flat 50% discount. Anything that isn’t user-facing, like nightly classification, backfills, report generation, or bulk content tagging, belongs there. Prompt caching is the other big lever: if your requests share a long common prefix, such as a detailed system prompt or a reference document, the cached portion bills at a fraction of the normal input rate. For a support bot with a fixed instruction set, that alone can cut input costs by half or more.
Then there’s the unglamorous stuff. Cap max_tokens so a runaway generation can’t produce a 4,000-token answer to a yes/no question. Summarise old conversation turns instead of carrying the full history forever. Route the easy 80% of traffic to a cheap model and escalate only the hard cases. These are engineering decisions, not pricing ones, and they matter more than any rate card. There’s a solid walkthrough of structuring this properly in building with the OpenAI API from first call to production.
When Azure OpenAI Changes the Arithmetic
If you’re inside a Microsoft shop, Azure OpenAI deserves a look. The per-token rates are broadly comparable to OpenAI’s direct pricing, but the billing runs through your existing Azure agreement, which means committed-spend discounts, enterprise invoicing, and data residency options can shift the effective cost. For some organisations the procurement simplification is worth more than the token savings. Others find the regional availability and deployment model a constraint. The trade-offs are covered in detail in what Azure OpenAI Service is and when to use it.
Forecast Your Own Bill in an Afternoon
Generic calculators are close to useless because your prompts aren’t generic. A better approach costs almost nothing.
Take thirty to fifty real requests from your actual workload. Run them through the model you intend to use. Log the input and output token counts returned in each API response, which are included by default. Average them out, multiply by your expected monthly volume, and apply the current rates. You’ll land within 20% of reality, which is far better than a guess and enough to sanity-check a budget conversation.
Then keep watching. The usage dashboard breaks spend down by model and by day, and a weekly glance catches the pattern where one endpoint quietly grows from 2% to 40% of your bill. Set a hard monthly spend limit in the platform settings while you’re there. Not because you expect to hit it, but because the alternative is finding out from an invoice.
One last habit worth building: log token counts per user or per feature from day one. When someone asks why costs tripled last month, “the summariser feature” is a useful answer and “we’re not sure” is not. Attribution turns pricing from a monthly shock into a number you can actually manage.