OpenAI API Pricing 2026: Cost per Million Tokens by Model

The OpenAI API costs $2 per million input tokens and $10 per million output tokens on gpt-6.1-sol, five times that on the flagship gpt-6-astra ($50 output), and $0.10 and $0.50 on gpt-6-luna, for prompts up to 272,000 tokens. Batch and Flex halve those rates, Fast mode doubles them, and cached input on gpt-6.1-sol is 95% off. Rates read on OpenAI's API pricing page on 19 October 2026.

Our take: for most teams the model choice is a 100x decision and everything else is a 2x decision. Luna to Astra is a hundredfold jump in price; caching, Batch and Fast mode each move the bill by a factor of about two. Route each request to the cheapest model that passes your evals first, then tune the processing tier.

Toolradar data: the OpenAI API profile on Toolradar, the software directory we run, ranks the API joint #2 among 483 API tools for how often other tools list it as an alternative, as of October 2026. Only one API in that category is cross-shopped more, which is why a price change here moves budgets across the whole stack.

How much does the OpenAI API cost per million tokens?

All prices are USD per million tokens on the Standard tier. Short context means up to 272,000 input tokens; above that, the long-context column applies to the whole request. OpenAI notes that gpt-5.6-sol's promotional pricing runs at least through 21 November 2026.

gpt-6.1-sol is the one to default to. It lists at the same input and output rates as gpt-6-sol but halves the cached-input rate, so any workload with a repeated system prompt pays less on the newer model for nothing. The 5.6 generation is now the expensive option at every tier: gpt-5.6-sol costs twice gpt-6.1-sol on input and output.

Which GPT model gives the best price per task?

The three current tiers sit on a clean ladder. gpt-6-astra costs five times gpt-6.1-sol, and gpt-6.1-sol costs twenty times gpt-6-luna, on both input and output. A request that luna handles correctly is a request you are overpaying for by 20x on sol and 100x on astra.

• gpt-6-luna fits classification, extraction, routing, tagging and high-volume chat where a wrong answer is cheap to catch.
• gpt-6.1-sol fits most production work: coding help, agents with short loops, RAG answers and summarization.
• gpt-6-astra fits the hardest reasoning and long agent runs, where one failed attempt costs more in engineering time than the token bill.

Legacy models deserve a second look in any old codebase. On the same pricing page, gpt-5.5 lists at $5 input and $30 output, gpt-4o at $2.50 and $10, and the pro models (gpt-5.5-pro and gpt-5.4-pro) at $30 and $180. A service still pinned to gpt-5.5 pays more than gpt-6.1-sol on every token, so migrating is a price cut before it is a quality question. Our best LLM APIs ranking compares OpenAI's ladder with other providers.

What do Batch, Flex, Fast and Ultrafast cost?

OpenAI sells the same models on five processing tiers. The tier multiplies the Standard rate:

Flex lists exactly the Batch rates for the current models, so either tier halves the bill for evals, backfills and nightly jobs that can wait. Fast and Ultrafast only make sense where latency is the product, such as a voice agent or an autocomplete box, because an Ultrafast astra call costs $300 per million output tokens.

Two uplifts apply on top of any tier. Regional processing (data residency) endpoints cost 10% more for models released on or after 5 March 2026, and FedRAMP endpoints carry the same 10% uplift. A team that needs EU data residency should add a tenth to every figure on this page.

How much does prompt caching save on OpenAI?

Cached input is 90% off the input rate on gpt-6-astra, gpt-6-sol and gpt-6-luna, and 95% off on gpt-6.1-sol. The current models also charge a cache write at 1.25 times the input rate when a prefix is first stored, so caching pays only when the same prefix is read more than once.

The math only needs your reuse count, which most teams can read from their logs. A 6,000-token system prompt read 1,000 times on gpt-6.1-sol costs about $12 uncached and about $0.60 in cache reads, plus one write. Put long, stable instructions, tool definitions and few-shot examples at the start of the prompt and the variable part at the end, because caching works on the shared prefix.

When does long context double the price?

Above 272,000 input tokens, the whole request moves to long-context rates: input doubles and output rises by half on every current model. A single 400,000-token document review on gpt-6.1-sol costs about $1.60 in input at the long-context rate, where the same tokens would cost $0.80 if they fitted under the threshold.

Chunking is therefore a pricing tool, not only a quality one. Two 200,000-token requests cost less than one 400,000-token request whenever the task allows the split, and RAG that retrieves the relevant 20,000 tokens beats both. Anthropic prices long prompts differently, which our Claude API pricing breakdown covers.

What a real workload costs on each model

Devshot data: we priced one workload on OpenAI's published rates on 19 October 2026. The job is a code-review bot that handles 30,000 requests a month, each with 8,000 input tokens (6,000 of them a shared, cached prefix) and 800 output tokens. That is 180 million cached input tokens, 60 million uncached input tokens and 24 million output tokens a month. Cache writes are left out because the shared prefix is written rarely next to 30,000 reads.

Output tokens are the real cost driver. On gpt-6.1-sol, the 24 million output tokens cost $240 of the $378 total, even though they are a tenth of the token volume. Capping response length, asking for diffs instead of rewritten files, and keeping reasoning effort low where it does not help will cut more than any input trick.

The same table shows the legacy penalty. Moving this bot from gpt-5.5 to gpt-6.1-sol saves $732 a month at identical traffic, and moving it to Flex on top halves what remains. Adding one web search per request would add $300 a month in tool fees alone, before the search content tokens, so tool calls belong in the cost model too.

What do tools, embeddings and audio cost?

Built-in tools carry their own fees on top of the model's token rates:

• Web search: $10 per 1,000 calls, plus search content tokens billed at the model's rates; the preview version for non-reasoning models is $25 per 1,000 calls with free content tokens.
• File search: $2.50 per 1,000 tool calls, and storage at $0.10 per GB per day after the first free GB.
• Hosted Shell and Code Interpreter: from $0.03 per 20-minute session for a 1 GB container to $1.92 for 64 GB.
• Embeddings: text-embedding-3-small at $0.02 and text-embedding-3-large at $0.13 per million tokens.
• Transcription: gpt-transcribe at $0.0045 per minute and Whisper at $0.006 per minute.
• Realtime voice: gpt-realtime-2.1 audio at $32 input and $64 output per million tokens, and the mini at under a third of those rates.

OpenAI is also winding down its fine-tuning platform: new users can no longer start, and fine-tuned models stay available until their base models are deprecated. A team planning a fine-tune in 2026 should plan on prompting, caching and routing instead. For a voice product, compare these audio rates with the specialist vendors in our best text-to-speech APIs list.

How does OpenAI's pricing compare with alternatives?

On short prompts, OpenAI's current ladder and Anthropic's line up almost rung for rung, so the real differences are long-context rules, caching rates and tokenizers; the side-by-side lives in our Claude API pricing page. A team that wants one bill across several providers can route through an aggregator, compared in our OpenRouter alternatives roundup, and a team that wants to stop paying per token can self-host an open model, as covered in our Ollama review and cheapest GPU cloud guide. The free LLM APIs guide lists providers with free tiers for prototypes.

For coding work specifically, the API is one route among several. Our AI coding assistant pricing comparison prices the subscriptions, the OpenAI Codex review covers OpenAI's own agent, and Claude Code vs Codex sets the two agents against each other.

How we checked these prices

We read OpenAI's API pricing page, including its Standard, Batch, Flex, Fast and Ultrafast tables and the tools, embeddings, audio and fine-tuning sections, on 19 October 2026. Every figure on this page is from that page; the workload table is our arithmetic on those rates. We did not run the workload against the API. Token counts vary by tokenizer and prompt, so treat the scenario as a model of your bill, not a quote. We re-check the page when OpenAI changes a model or rate. Louis Corneloup, who founded Toolradar and Dupple, edited it.

FAQ

How much does the OpenAI API cost per million tokens in 2026?

On the Standard tier for prompts up to 272,000 tokens, gpt-6-astra costs $50 per million output tokens, gpt-6.1-sol a fifth of that and gpt-6-luna $0.50, with input priced at a fifth of output on all three. Rates are from OpenAI's API pricing page, read 19 October 2026.

What is the cheapest OpenAI model?

gpt-6-luna is the cheapest current chat model at $0.10 input and $0.50 output per million tokens, and half that on Batch or Flex. Older models such as gpt-5-nano list lower still, but they are a previous generation and a new project should start on luna.

How much does the OpenAI Batch API save?

Batch halves the Standard rate on every listed model, so gpt-6.1-sol drops to $1 input and $5 output per million tokens. Flex charges the same rates for the current models, which suits evals, backfills and nightly jobs that can wait.

How much does cached input cost on the OpenAI API?

Cached input is 90% off the input rate on most current models and 95% off on gpt-6.1-sol, where it costs a twentieth of the uncached rate. Current models also charge a one-time cache write at 1.25 times the input rate when a prefix is first stored.

Why did my OpenAI bill double on long prompts?

Requests over 272,000 input tokens are billed at long-context rates, which double input and raise output by half for the whole request. Splitting documents or retrieving only the relevant chunks keeps requests under the threshold.

Is gpt-6.1-sol cheaper than gpt-5.5?

Yes. gpt-5.5 lists at two and a half times gpt-6.1-sol's input rate and three times its output rate, so our code-review workload costs $1,110 a month on gpt-5.5 and $378 on gpt-6.1-sol with caching.

Cite this: Devshot, "OpenAI API Pricing 2026: Cost per Million Tokens by Model", October 2026.

Some offers on this page may be paid placements or contain affiliate links.