OpenAI API Review 2026

What is OpenAI API?

GPT-6 Luna charges $0.10 for each million input tokens and $0.50 for each million output tokens, so a support bot answering 20,000 chats a month (3,000 tokens in, 500 out per chat) costs about $11 to run. Put the same traffic on GPT-6 Astra and the bill climbs to roughly $1,100. That hundredfold spread is the real story behind OpenAI API pricing per million tokens in late 2026, and it settles most buying decisions before features enter the discussion. This review draws on OpenAI's developer pricing page, its model cards, the rate-limit guide and the help centre to explain what you pay, what you get and where the catches sit.

Three GPT-6 models and a hundredfold price gap

The OpenAI API is the paid, programmatic route to the models behind ChatGPT. Your code sends text or images to an endpoint, the model sends text back, and you are billed by the token. OpenAI, the San Francisco company founded in 2015, currently lists three flagship models on its API page. GPT-6 Astra arrived on 3 September 2026 and is the flagship for hard reasoning and long coding jobs. GPT-6 Sol and GPT-6 Luna followed on 22 September, and a week later GPT-6.1 Sol took Sol's place on the price list. Luna is the cheap, fast option for high-volume work.

All three share a 1.05 million token context window and 128K tokens of maximum output. They read text and images and write text only. OpenAI's migration notes for Astra list what it gives up: custom temperature, top_p and the "none" reasoning effort.

OpenAI API pricing per million tokens, model by model

Below are standard rates from the developer pricing page, set beside Anthropic's nearest Claude model from its own price list. Prices are in US dollars.

Per single token the numbers look tiny: Sol input works out to $0.000002 a token. Several modifiers change the real bill:

• Batch and Flex processing cost half the standard rate.
• Fast mode, which OpenAI used to call Priority processing, doubles the price. Ultrafast, offered on 6.1 Sol and Astra, costs six times standard.
• Long prompts over 272K input tokens are charged 2x for input and cache and 1.5x for output, on the whole request.
• Cache writes are billed at 1.25x the input rate, while cache reads get 90 to 95% off.
• Regional processing endpoints add 10% on models released since 5 March 2026.

Extra meters for tools, audio, images and embeddings

Token prices are only part of a production bill. Built-in web search costs $10 per 1,000 calls plus the search content tokens. File search is $2.50 per 1,000 calls, with storage at $0.10 per GB per day after the first free GB. A hosted code container with 1 GB of memory costs $0.03 per 20-minute session.

For retrieval projects, embeddings from text-embedding-3-small cost $0.02 per million tokens, which is close to free at small scale. Speech work has moved on: the Whisper API model (whisper-1) shuts down on 26 February 2027, and its replacement gpt-transcribe runs at $0.0045 a minute. Voice agents built on the Realtime API with gpt-realtime-2.1 pay $32 per million audio input tokens and $64 for audio output, or $10 and $20 on the mini version. Image generation with gpt-image-2.5 is billed at $30 per million image output tokens.

OpenAI API free credits: what the free tier really runs

There is a Free usage tier. It needs no purchase, only an account in a supported country, and carries a $100 monthly usage cap. The catch shows up on the model cards: Astra, 6.1 Sol and Luna all list the Free tier as "not supported". The quickstart mentions a free test request and then points you to billing.

So treat $5 as the real entry fee. Prepaid credits start at $5 (the default top-up is $10), expire one year after purchase and are non-refundable, according to the help centre. Any promotional grant is spent before paid credit. Anthropic offers no ongoing free tier, just a one-time $5 credit that new Console users claim by verifying a phone number.

How to get an OpenAI API key and keep it out of GitHub

• Create a developer account on platform.openai.com and add at least $5 of credit.
• Generate a key on the API keys page of the dashboard and save it somewhere safe straight away.
• Store it as an environment variable called OPENAI_API_KEY rather than pasting it into code.
• Install an official SDK: pip install openai for Python or npm install openai for JavaScript. .NET, Java, Go and Ruby libraries also exist.
• Send a first request to the /v1/responses endpoint with GPT-6 Luna, which is the cheapest place to debug.

Use one key per project so a leak can be revoked without breaking everything. Never ship a key inside a mobile app or browser bundle; route calls through your own server. Leave auto-recharge off until you know your monthly burn.

Responses API first, Chat Completions kept, Assistants retired

New code should target the Responses API. Every example in the docs uses it, and tool calling on the GPT-6 models requires it. Chat Completions still works and still appears on the status page, which helps when porting older apps, though Luna supports function calling there only with reasoning effort set to none.

The Assistants API was switched off on 26 August 2026, with Responses plus the Conversations API as the migration path. Function calling and structured outputs (JSON that matches a schema you supply) work on all three GPT-6 models. The Batch API suits overnight jobs such as tagging a product catalogue at half price. Fine-tuning is being wound down: new users can no longer reach it, and active customers lose the ability to start new jobs on 6 January 2027. A newer Decisions API, in beta with Luna since 6 October, returns typed answers for classification-style tasks.

OpenAI API rate limits after the move to three paid tiers

On 6 October 2026 OpenAI cut its paid usage tiers from five to three. Upgrades happen automatically once your total credit purchases cross each threshold:

• Build ($5 purchased): 5,000 requests per minute and 1 million tokens per minute on Astra and Sol, 2 million TPM on Luna. Monthly spend cap $500.
• Launch ($100): 10,000 RPM; 4 million TPM on Astra and Sol, 10 million on Luna. Cap $5,000.
• Grow ($500): 15,000 RPM and 40 million TPM on Astra and Sol; Luna gets 30,000 RPM and 180 million TPM. Cap $200,000.

Limits also apply per day and, for image models, per minute of images. Whichever ceiling you touch first triggers the block.

Fixing a 429 error without making it worse

A 429 has two very different causes. If the message is about quota or billing, retrying is pointless; top up credit or raise your spend cap. If it says slow_down, your request rate rose too fast. The rate-limit guide recommends honouring the Retry-After header plus a small random delay, then falling back to exponential backoff with jitter. Cap retries by count and total time. If you write your own retry loop, switch off the SDK's built-in retries so the two do not multiply. Failed calls still count against your per-minute limit. Past roughly 1 million input tokens a minute, grow traffic by no more than 50% every 15 minutes, and push anything that can wait into the Batch API.

OpenAI API vs Anthropic API on the same coding workload

Which one is cheaper? On list price the two firms now mirror each other: Luna matches Haiku 5.5, 6.1 Sol matches Sonnet 5.5 and Astra matches Fable 5.1. Anthropic adds Claude Opus 5.5 at $4 and $20, a middle step OpenAI does not offer. Both cut batch jobs by 50%.

The gap comes from caching and prompt size. Take a coding agent that sends 50 million input tokens a month, 80% of them cache hits, and gets 5 million output tokens back. On 6.1 Sol that is $20 for fresh input, $4 for cached input and $50 for output, about $74 before cache-write charges. Long prompts matter too: Haiku 5.5 jumps to $0.50 and $2.50 above 100K tokens, while Luna's surcharge starts at 272K. OpenAI also sells embeddings, transcription, realtime voice and image models in one account, which Anthropic's price list does not.

Paying from India in dollars, plus the in.api.openai.com option

India is on OpenAI's list of supported API countries. The INR and UPI billing you see in ChatGPT does not carry over, though; the help centre says API credit purchases are limited to USD. Expect to need a card that allows international payments, and remember the one-year credit expiry when buying in bulk. For companies with data rules, an India endpoint at in.api.openai.com stores data in the country but does not process it there, and it needs Modified Abuse Monitoring or zero data retention approval.

What happens to the prompts you send

OpenAI's data guide says API traffic is not used for model training unless you opt in. Abuse-monitoring logs are kept for up to 30 days by default. Eligible customers can apply for zero data retention, though files, conversations and fine-tuning endpoints still keep state. Under the Services Agreement you own your output and may use it commercially, but you may not use it to build a competing model (small internal classifiers and embeddings are excepted). Since 5 October, eligible organisations can accept a HIPAA business associate agreement inside the dashboard, and live service health sits on status.openai.com.

Where building on OpenAI gets frustrating

The platform changes fast, and that churn costs engineering time. In just over a year the Assistants API closed, fine-tuning started winding down, whisper-1 got a shutdown date and the tier system was rebuilt. Billing is USD-only for Indian teams. The free tier cannot call a single flagship model. Small accounts mostly rely on help articles rather than a named contact.

Who should build on GPT-6, and which alternatives deserve a test

The API makes most sense for SaaS teams adding chat or agents to a product, agencies running support bots for clients, and anyone who wants voice, transcription, images and embeddings from one bill. Start on Luna, log token counts for a week, and move only the hard prompts up to Sol or Astra. For comparison, test Anthropic's Claude API (Opus 5.5 fills the price gap), the Google Gemini API, and DeepSeek, whose deepseek-flash costs $0.15 and $0.60 per million tokens off-peak but stores data in China.