OpenAI API pricing in 2026, checked 13 Oct: price per million tokens and what a chatbot, RAG, summary or agent costs per user per month.
Short answer. OpenAI's API is priced per million tokens. On the standard tier, the three current flagship models cost $0.10 input / $0.50 output (gpt-6-luna), $2 / $10 (gpt-6.1-sol) and $10 / $50 (gpt-6-astra). In practice a typical chatbot user costs about $0.02, $0.36 or $1.80 a month on those three models, before any savings. A heavy AI agent user can cost up to $14 a month on the top model, which is why model choice matters more than anything else on this page.
Per-token prices mean little until you convert them into something a founder budgets for: the cost of one user, per month, for one feature. This article takes OpenAI's published price table, states the token assumptions for four common AI features, and shows the arithmetic so you can change the numbers for your own product. Prices are in US dollars, as OpenAI lists them, and exclude VAT. Last checked: 13 October 2026.
What does the OpenAI API cost per million tokens in 2026?
The table shows OpenAI's current flagship models on the standard tier with short context, per 1 million tokens. The same page lists Batch, Flex, Fast and Ultrafast tiers.
| Model | Input | Cached input | Cache writes | Output |
|---|---|---|---|---|
| gpt-6-luna | $0.10 | $0.01 | $0.125 | $0.50 |
| gpt-6.1-sol | $2.00 | $0.10 | $2.50 | $10.00 |
| gpt-6-astra | $10.00 | $1.00 | $12.50 | $50.00 |
Source: OpenAI API pricing, standard tier, short context, USD per 1M tokens, checked 13 October 2026.
Four things on OpenAI's page change the bill and are easy to miss:
- Batch is half price. The Batch tier lists gpt-6.1-sol at $1.00 input and $5.00 output, and gpt-6-luna at $0.05 and $0.25.
- Long context costs more. For prompts above the short-context limit, gpt-6.1-sol is listed at $4.00 input and $15.00 output, which is double the input price and 1.5 times the output price.
- Speed costs more. Fast mode (previously called Priority) lists gpt-6.1-sol at $4.00 input and $20.00 output, double the standard rate.
- Data residency costs 10% more. OpenAI charges a 10% uplift on regional processing endpoints for eligible models released on or after 5 March 2026, which matters if you need processing kept in a particular region.
OpenAI also notes that promotional pricing on one of its models runs at least until 21 November 2026, so confirm the live page before you fix a budget.
How do you turn a price per token into a cost per user?
Take the tokens one user sends and receives in a month, multiply by the price per million, and add the two. We use four features and write the assumptions out, because every AI cost estimate hides its assumptions somewhere. Treat them as starting points to be replaced by your own logs. Token counts are Magora planning estimates, not OpenAI figures.
| Feature | Usage per active user per month | Input tokens | Output tokens |
|---|---|---|---|
| Support chatbot | 10 conversations of about 6 turns. Each turn resends a 1,500-token instruction and knowledge prompt plus the history so far; about 1,200 tokens of replies per conversation. | 120,000 | 12,000 |
| Document Q&A (RAG) | 30 questions. Each sends about 4,000 tokens of retrieved passages, 500 of instructions and the question, and returns about 400 tokens. | 138,000 | 12,000 |
| Summarisation | 20 documents of roughly 4,500 words (6,000 tokens), each summarised in about 500 tokens. | 120,000 | 10,000 |
| AI agent with tool calls | 15 tasks. Each task runs about 8 steps; every step resends the growing context (about 10,000 tokens on average) and writes about 375 tokens. | 1,200,000 | 45,000 |
The agent row is the one to notice. An agent calls the model repeatedly and resends its working context each time, so input tokens grow far faster than the answer does.
What does each feature cost per user per month?
Using the standard prices above and no savings at all, the monthly cost of one active user is:
| Feature | gpt-6-luna | gpt-6.1-sol | gpt-6-astra |
|---|---|---|---|
| Support chatbot | $0.02 | $0.36 | $1.80 |
| Document Q&A (RAG) | $0.02 | $0.40 | $1.98 |
| Summarisation | $0.02 | $0.34 | $1.70 |
| AI agent with tool calls | $0.14 | $2.85 | $14.25 |
Our calculation from OpenAI's standard prices and the Magora planning estimates above, rounded; no caching or batch discount applied.
Worked example for the support chatbot on gpt-6.1-sol: 120,000 input tokens at $2 per million is $0.24, and 12,000 output tokens at $10 per million is $0.12, giving $0.36. At 1,000 active users that is about $360 a month, which is small next to the cost of building the feature. Do the same sum for the agent at $2.85 and 1,000 users and you reach about $2,850 a month, enough to affect your pricing.
A word on interpretation. These are costs per active user, not per registered user, and they cover model usage only. Hosting, vector database storage, monitoring and, for agents, any tool charges sit on top. OpenAI lists web search at $10 per 1,000 calls plus the tokens retrieved, and file search tool calls at $2.50 per 1,000 calls, for example.
How do you cut the cost of an AI feature?
Four levers do most of the work. Here they are in the order we usually pull them.
1. Choose the cheapest model that passes your tests
On input price, gpt-6.1-sol is 20 times gpt-6-luna, and gpt-6-astra is a further 5 times gpt-6.1-sol. A common pattern is to route simple requests (classification, short replies, extraction) to the cheap model and send only hard cases to a larger one. The test is accuracy on a set of your own real cases, not a benchmark table. For the support chatbot above, moving from gpt-6.1-sol to gpt-6-luna takes $0.36 to $0.02 per user.
2. Use prompt caching for repeated prompts
Cached input on gpt-6.1-sol is $0.10 per million against $2.00. If 70% of the support chatbot's input tokens are repeated instructions and earlier turns that hit the cache, the monthly cost falls from $0.36 to about $0.20. For the agent, with the same 70%, it falls from $2.85 to about $1.25. Two cautions: caching rewards a stable prompt prefix, so put fixed instructions first and variable content last; and writing to the cache costs 25% more than normal input on this model ($2.50 against $2.00), so it pays back only when the prefix is reused. The 70% hit rate is a Magora planning estimate; measure yours.
3. Move non-urgent work to the Batch API
Batch halves the price. Summarising 20 documents per user on gpt-6.1-sol falls from $0.34 to $0.17 if the summaries can wait. It is a poor fit for live chat and a good fit for overnight reports, backlog classification and evaluation runs.
4. Keep prompts and retrieved context short
You pay for every token you resend. In document Q&A, retrieving 2,000 tokens of well-chosen passages instead of 4,000 cuts the monthly cost per user on gpt-6.1-sol from $0.40 to about $0.28. For an agent, trimming resent context by 40% (1.2 million to 720,000 input tokens) takes it from $2.85 to about $0.93 when combined with the 70% cache assumption. Better retrieval is cheaper and usually more accurate, which is why it is a design task rather than a tuning afterthought.
We build AI assistants, automation and machine-learning features into real products, from proof of concept to production. First release in 6–8 weeks; developers from £49/hour.
How do OpenAI prices compare with Anthropic?
Anthropic publishes its own table on its pricing page. On the standard rate, Claude Haiku 5.5 is listed at $0.10 input and $0.50 output per million tokens (for prompts up to 100,000 tokens), Claude Sonnet 5.5 at $2 and $10, and Claude Fable 5.1 at $10 and $50. Those match gpt-6-luna, gpt-6.1-sol and gpt-6-astra tier for tier, so at this level the choice between providers turns on quality for your task, latency and data terms, not list price. Anthropic lists its Batch API at a 50% discount, as OpenAI does. We have not compared Google's prices here because we did not verify them for this article.
Source: Anthropic pricing documentation, checked 13 October 2026. Both vendors change prices often; recheck before budgeting.
What does it cost to build the AI feature itself?
Running costs are only part of the bill. Magora's 2026 starting prices in GBP are: a free working prototype in 5-10 days, Discovery from £3,000 (two weeks), AI agents from £20,000 for one internal workflow, and a RAG pipeline with an API from £30,000. Our AI and ML development page has the full list, and support after launch starts from £1,000 a month. Model usage is billed separately by the provider. For wider cost-control habits, see how AI startups keep model costs under control.
For UK businesses, hosting is the other question. Application data, vector stores and logs can sit in UK or EU cloud regions, and where model processing must stay in-region we choose a provider and model that support it, or deploy through Azure OpenAI, AWS Bedrock or a self-hosted open-source model. In Discovery we measure real cost per request on your data, so the number you take to a board comes from a test, not from a table like this one.
Frequently asked questions
How much does the OpenAI API cost in 2026?
OpenAI charges per million tokens, and the price depends on the model. On the standard tier, as listed on 13 October 2026, gpt-6-luna costs $0.10 input and $0.50 output, gpt-6.1-sol costs $2 input and $10 output, and gpt-6-astra costs $10 input and $50 output. For a typical user of a support chatbot, that is roughly $0.02, $0.36 and $1.80 a month respectively, using the token assumptions in this article.
What is a token, and how many words is that?
A token is a chunk of text, often a short word or part of a longer one. In English, 1,000 tokens is roughly 750 words, which is a rule of thumb rather than an OpenAI-published figure. Input tokens (what you send) and output tokens (what the model writes) are priced separately, and output is five times dearer than input on all three flagship models.
Does a ChatGPT subscription include API access?
No. A ChatGPT subscription and the API are billed separately. The API is pay-as-you-go, billed in tokens, and is what you use to build AI features into your own product.
How much does cached input save?
OpenAI lists a separate, lower price for cached input: $0.10 per million tokens on gpt-6.1-sol against $2.00 uncached, $1.00 against $10.00 on gpt-6-astra and $0.01 against $0.10 on gpt-6-luna. It applies to the repeated start of a prompt, such as a long system prompt or a fixed set of instructions. Writing to the cache is priced slightly above normal input ($2.50 per million on gpt-6.1-sol).
What does the Batch API save, and when can I use it?
The Batch API prices the same models at half the standard rate. It suits work nobody is waiting for, such as overnight summaries, classification of a backlog or evaluation runs. It does not suit live chat.
Is there a UK or EU data residency option, and does it cost more?
OpenAI lists regional processing (data residency) endpoints with a 10% uplift for eligible models released on or after 5 March 2026. Which regions and models are supported is set out in OpenAI's own data guide, so check it against your requirements. Azure OpenAI and Amazon Bedrock are alternative routes, billed through those services.
How much does it cost to build an AI feature, as opposed to running one?
Running cost is usually the smaller number. At Magora, an AI agent for one internal workflow starts from £20,000 and a RAG pipeline with an API from £30,000, after a free prototype and a Discovery sprint from £3,000. Model usage is then billed on top, which is why this article exists: so you can estimate it before you commit.
Want a running-cost estimate for your own AI feature? Send us your brief for a free prototype in 5-10 days, try the app cost calculator, or start with a Discovery sprint from £3,000.
Sources: OpenAI API pricing (openai.com/api/pricing, served at developers.openai.com/api/docs/pricing), Anthropic pricing documentation (platform.claude.com/docs/en/about-claude/pricing), Magora pricing pages. Last checked 13 October 2026. Token assumptions and per-user costs are Magora planning estimates.
Alex Galaktionov
A seasoned technology expert and agile advocate, Alexey brings over a decade of transformative expertise in the IT sector guiding startups through effective fundraising strategies, tech innovation and development methodologies for groundbreaking products.
LinkedIn