Blog · 2026-10-16 · By Dhrumil Barot, Founder

OpenAI API Pricing: Every GPT Model, Cache and Batch Rate

OpenAI API pricing verified 2026-10-16: gpt-6-luna $0.10/$0.50 to gpt-6-astra $10/$50 per 1M tokens. Batch, Flex, Fast, caching, long context and cost per task with every assumption shown.

TL;DR: OpenAI API pricing on 2026-10-16 runs from $0.10 input / $0.50 output per million tokens (gpt-6-luna) to $10 / $50 (gpt-6-astra). The workhorse gpt-6.1-sol costs $2 / $10. Batch and Flex halve every rate, Fast doubles it, and a cache hit on gpt-6.1-sol costs $0.10, 5% of the input price.

OpenAI API pricing is a per-token rate card: each model has a price per 1 million input tokens, cached input tokens, cache writes and output tokens, in USD. Prices verified 2026-10-16 against the OpenAI API pricing docs. On a representative task of 10,000 input tokens and 1,000 output tokens, one call costs $0.0015 on gpt-6-luna, $0.03 on gpt-6.1-sol, $0.06 on gpt-5.6-sol and $0.15 on gpt-6-astra. Below: the full rate card, the modifiers that move a real bill (caching, Batch, Flex, Fast, long context, data residency, reasoning tokens, tools) and cost per task with every assumption shown.

How much does the OpenAI API cost per million tokens?

The OpenAI API price is set per model, per 1 million tokens, at standard (synchronous) processing. These are the current flagship models, short context (272K input tokens or fewer):

Output costs 5x input on every GPT-6 model and 5x to 6x on GPT-5.6. For chat and agent workloads, output is usually the line that dominates the bill. gpt-6.1-sol and gpt-6-sol have identical input and output prices; the only difference is the cache read ($0.10 vs $0.20), so anyone on gpt-6-sol with a cached prefix should move. GPT-5.6 Sol is on promotional pricing that OpenAI says runs at least through 2026-11-21, so recheck it after that date.

What do older GPT models cost?

Older OpenAI models stay on the API at their own rates, and several now cost more than newer, stronger models:

gpt-5.5 at $5 / $30 is 2.5x the input and 3x the output of gpt-6.1-sol. gpt-4o at $2.50 / $10 now costs more on input than gpt-6.1-sol. gpt-5-nano at $0.05 / $0.40 is still the lowest text rate on the card, but gpt-6-luna at $0.10 / $0.50 is a newer generation for a small premium. Check model IDs against the /models catalog before switching.

How does OpenAI prompt caching pricing work?

Prompt caching is automatic discounting of repeated prompt prefixes. For GPT-5.6 and later, per the OpenAI prompt caching guide:

• A cache read costs 0.1x the uncached input rate (0.05x on gpt-6.1-sol).
• A cache write costs 1.25x the uncached input rate. Models before GPT-5.6 have no write charge.
• The minimum cacheable prefix is 1,024 visible input tokens.
• prompt_cache_options.ttl supports one value, 30m: the prefix stays reusable for at least 30 minutes after its last write or read.
• Older models use prompt_cache_retention with in_memory (typically 5 to 10 minutes) or 24h.

Break-even: a write costs 0.25x extra once, and each read saves 0.9x. One cache hit pays back the write. A prefix shorter than 1,024 tokens never caches, which is why OpenAI's own guide suggests padding a stable prefix past the minimum when reuse is high. Our prompt caching break-even analysis works through more shapes.

How much do Batch and Flex save?

The Batch API is asynchronous processing at 50% of standard prices with a 24-hour completion window, per the Batch guide. Flex processing charges the same Batch rates on normal Responses or Chat Completions calls, in exchange for slower responses and occasional resource-unavailable errors, per the Flex guide. Batch and Flex rates for the flagships:

Use Flex when you want half price without rewriting to a JSONL batch file: set service_tier: "flex" and raise your SDK timeout to 15 minutes, as the guide shows. Use Batch when you have thousands of rows and can wait up to 24 hours. Our Batch APIs comparison covers the Anthropic side.

What do Fast and Ultrafast mode cost?

Fast mode is OpenAI's paid low-latency tier, renamed from Priority processing on 2026-07-30. It costs 2x standard: gpt-6.1-sol is $4 / $20 and gpt-6-astra is $20 / $100. Ultrafast exists for two models at 6x: gpt-6-astra at $60 / $300 and gpt-6.1-sol at $12 / $60. A request that would cost $0.03 on standard gpt-6.1-sol costs $0.06 on Fast and $0.18 on Ultrafast. Pay for it only on user-facing paths where latency loses revenue.

What does long context and data residency add?

Long context pricing applies when a request has more than 272K input tokens. On GPT-6 and GPT-5.6 models, input doubles and output rises 1.5x: gpt-6.1-sol goes from $2 / $10 to $4 / $15, and gpt-6-astra from $10 / $50 to $20 / $75. The whole request is billed at the long rate, so a 280K-token prompt costs about twice a 270K one. Trimming retrieved context under 272K is the single cheapest optimization for RAG over large corpora.

Regional processing (data residency) endpoints and FedRAMP endpoints add a 10% uplift for models released on or after 2026-03-05. gpt-6.1-sol in an EU region is $2.20 / $11.

How are reasoning tokens billed?

Reasoning tokens are billed as output tokens. They are not returned by the API but they occupy the context window and appear in usage.output_tokens_details.reasoning_tokens, per the OpenAI reasoning guide. A reply with 1,000 visible tokens and 4,000 reasoning tokens bills 5,000 output tokens: $0.05 on gpt-6.1-sol, not $0.01. OpenAI notes reasoning can run from a few hundred to tens of thousands of tokens per request. Lower reasoning.effort is the direct cost lever; measure tokens per task before and after.

What does an OpenAI API task actually cost?

Cost per task is the number that matters, not price per token. Assumptions (edit them for your workload):

Derivation for gpt-6.1-sol cached: 2,000 x $2 / 1M = $0.004, plus 8,000 x $0.10 / 1M = $0.0008, plus 1,000 x $10 / 1M = $0.01, total $0.0148 per task. The cached column assumes the prefix is already warm; the first call pays a write (8,000 x $2.50 / 1M = $0.02 instead of $0.016). Run your own token counts through the LLM cost calculator.

Moving from gpt-4o to gpt-6.1-sol cuts this task 14% uncached ($0.035 to $0.030) and 41% cached ($0.025 to $0.0148), because gpt-4o's cache read is 50% of input while gpt-6.1-sol's is 5%.

What other charges appear on an OpenAI API bill?

Tool calls and storage are billed on top of tokens:

• Web search: $10.00 per 1,000 calls, plus search content tokens at model rates.
• File search: $2.50 per 1,000 calls, storage $0.10 per GB per day after 1 GB free.
• Hosted Shell and Code Interpreter containers: $0.03 per 20-minute session at 1 GB, up to $1.92 at 64 GB.
• Embeddings: text-embedding-3-small $0.02 and text-embedding-3-large $0.13 per 1M tokens.
• Moderation: omni-moderation-latest is free.
• Fine-tuning: OpenAI is winding down the fine-tuning platform; it is closed to new users.

An agent that makes 3 web searches per task adds $0.03, the same as the entire gpt-6.1-sol token cost above. We audited this in built-in agent tools, the second bill.

Is OpenAI on Azure or Bedrock priced the same?

OpenAI models on Microsoft Azure and Amazon Bedrock are billed by those clouds under their own price lists, not by OpenAI. Check the Azure OpenAI pricing page and Amazon Bedrock pricing for the model and region you use. Going through a cloud makes sense when committed spend, private networking or procurement rules matter more than the per-token rate.

Is ChatGPT Plus the same as API access?

ChatGPT Plus and other ChatGPT plans are subscriptions for the ChatGPT apps; they do not include API credits. The API is pay as you go per token, billed separately on the OpenAI platform. Usage caps on the API are rate limits and spend limits set per project, not a monthly plan.

When is the expensive model worth it?

gpt-6-astra at $50 per million output tokens is worth it when a cheaper model needs retries or extra agent turns to finish. If gpt-6.1-sol needs more than 5x the tokens of gpt-6-astra to complete the same task, Astra is cheaper per completed task. That happens on long-horizon coding and research agents; it rarely happens on extraction, classification or summarization, where gpt-6-luna at 1/100th of Astra's price is the right default. Don't route by sticker price alone: measure tokens and success rate per task on your own eval set first.

Where Vynaris fits

Vynaris is an OpenAI-compatible API that routes each request to the cheapest model that can handle it, across OpenAI and other providers. If your traffic mixes easy and hard requests, routing the easy share to a Luna-class model is where most savings come from. See Vynaris pricing. If all your traffic is hard reasoning or all trivial extraction, you don't need a router; pick one model.

FAQ

How much does the OpenAI API cost per 1,000 tokens?

Divide the per-million price by 1,000. gpt-6.1-sol costs $0.002 per 1,000 input tokens and $0.01 per 1,000 output tokens. gpt-6-luna costs $0.0001 and $0.0005.

What is the cheapest OpenAI API model?

For text, gpt-5-nano at $0.05 input / $0.40 output per 1M tokens. Among current-generation models, gpt-6-luna at $0.10 / $0.50. Batch or Flex halves either.

Is the OpenAI API free?

No. API usage is billed per token from the first request. The moderation model omni-moderation-latest is the one free endpoint.

How is OpenAI API usage calculated?

Each response returns a usage object with input tokens, cached tokens and output tokens (including reasoning tokens). Multiply each by its per-1M rate and divide by 1,000,000.

Does the API charge for the system prompt and chat history?

Yes. Everything you send is input tokens, every turn. Long chats resend history, which is why caching matters; see the full-conversation token tax.