OpenAI API Pricing (September 2026): Every Model's Per-Token Price, and What a Small Business Actually Spends

The short answer

As of September 2026, OpenAI API pricing per million tokens (input / cached input / output) is: GPT-6 Astra $10 / $1 / $50, GPT-5.6 Sol $4 / $0.40 / $20, GPT-5.6 Terra $2 / $0.20 / $12, and GPT-5.6 Luna $0.20 / $0.02 / $1.20. Cached input bills at 10% of the input price, the Batch API halves everything, and there is no monthly platform fee. Those numbers come from OpenAI's API pricing page and its developer pricing docs, both read on September 15, 2026. The Sol price is promotional, down from $5 / $30, and OpenAI says it holds at least through November 21, 2026. If you searched "ChatGPT API pricing," this is the same page: the API is what software calls, ChatGPT is the app, and a ChatGPT subscription does not include API access. What decides your bill is which model each job runs on, whether the prompt is cached, and whether anyone is waiting for the answer. I install these workflows for owner-led businesses, so the back half of this page is what three real workloads cost.

OpenAI API pricing per million tokens (September 2026)

The current family first. Standard processing, contexts under 272K tokens.

Older models OpenAI still lists, same units (input / cached / output): GPT-5.5 $5 / $0.50 / $30, GPT-5.4 $2.50 / $0.25 / $15, GPT-5.4 mini $0.75 / $0.075 / $4.50, GPT-5.2 $1.75 / $0.175 / $14, GPT-5.1 and GPT-5 $1.25 / $0.125 / $10, GPT-5 mini $0.25 / $0.025 / $2, GPT-5 nano $0.05 / $0.005 / $0.40, GPT-4.1 $2 / $0.50 / $8, GPT-4o $2.50 / $1.25 / $10, GPT-4o mini $0.15 / $0.075 / $0.60. If you are still on GPT-4o, you are paying more than Terra for a two-year-old model. If you are on GPT-5 mini, Luna is cheaper and newer.

The multipliers, all from the same page:

• Long context. Over 272K input tokens, input and cached input bill at 2x and output at 1.5x. Astra becomes $20 / $2 / $75.
• Batch API. 50% off input and output for asynchronous jobs that complete within 24 hours.
• Flex processing. The Batch rates, synchronous, "in exchange for slower response times and occasional resource unavailability."
• Fast mode. Double the standard rate: Astra $20 / $100, Sol $8 / $40.
• Data residency. A 10% uplift on models released on or after March 5, 2026 when you route through regional endpoints.

Update (September 27, 2026): GPT-6 Sol and GPT-6 Luna, and the Agents API beta

A week after this page went up, OpenAI filled out the GPT-6 family. Its API changelog for September 22, 2026 lists GPT-6 Sol at $2 input, $0.20 cached input and $10 output per million tokens, and GPT-6 Luna at $0.10 / $0.01 / $0.50, for prompts up to 272K tokens. Both are reasoning models that take text and images and return text through the Responses and Chat Completions APIs. OpenAI's model pages describe Sol as built "to power complex coding and agentic workflows" and Luna as its "most efficient model for focused, high-volume tasks." The same 272K rule applies (Sol goes to $4 / $15 over the line, Luna to $0.20 / $0.75), and Batch halves them to $1 / $5 and $0.05 / $0.25.

What that does to the lineup above: GPT-6 Sol costs half of GPT-5.6 Sol and matches GPT-5.6 Terra on input while charging $2 less per million on output. GPT-6 Luna is half of GPT-5.6 Luna on both sides. GPT-5.6 is still listed at the prices in the table, including the Sol promo OpenAI says holds at least through November 21, 2026, so nothing already running breaks. But for new work, the GPT-6 pair is the cheaper place to start. Rerun the support-bot example below with the prompt cached: GPT-6 Sol comes to about $66 a month (against $74 on Terra) and GPT-6 Luna to about $3.30 (against $7.40 on GPT-5.6 Luna). Arithmetic on OpenAI's rates, same token counts. Test on your own tickets before you switch a live workflow.

The other September change is the Agents API, in public beta since September 10, 2026. OpenAI describes it as agents built on "a managed Codex harness" where OpenAI handles session orchestration, context compaction and recovery. There is no separate agent fee: per OpenAI's docs, model usage bills at the chosen model's API rates, tools at their standard rates, and OpenAI-hosted sandboxes at the container rates in the next table. So the cost of an agent is the tokens it burns plus the sandbox time, which is exactly why the cheaper GPT-6 models matter. What agents are actually shipping for owner-led businesses right now, from OpenAI and the others, is on AI agents for small business.

Update (October 4, 2026): GPT-6.1 Sol. A week later OpenAI replaced GPT-6 Sol on its flagship rate card. The API changelog for September 29 lists GPT-6.1 Sol at $2 input, $0.10 cached input, $2.50 cache write and $10 output per million tokens up to 272K, so the only price that moved is cached input, which halved. Batch and Flex stay at $1 / $5. On the cached support-bot example below that takes Sol from about $66 to about $63 a month (10M × $2 + 30M × $0.10 + 4M × $10). Small, but it widens the gap over GPT-5.6 Sol, whose promotional $4 / $20 still runs at least through November 21, 2026.

Tools, images, voice, and the other line items

Token prices are the headline; these are the rows that show up on a real invoice.

Images are priced in tokens, not per image, so "one image" costs whatever its size and quality consume; OpenAI's pricing pages print no per-image figure. And fine-tuning is winding down, closed to new users per OpenAI's docs.

Cached input and Batch: the two discounts that decide the bill

Every OpenAI pricing roundup copies the input and output columns and skips the cached column, which is where the money is. On the current models a cache hit bills at 10% of input: Terra's $2 becomes $0.20, Sol's $4 becomes $0.40. OpenAI's developer table lists a cache write at 1.25x input on the same models, a small premium to store the prefix once so every later request reads it at a tenth of the price. If your requests share a system prompt, a knowledge base, or a set of examples and you are not caching, you are leaving most of the discount on the table.

Batch is the other half. Anything with no human waiting, such as overnight document processing, weekly reports, and backfills, runs at half price with a 24-hour completion window. The two stack: Terra cached input on Batch is $0.10 per million, one twentieth of the standard input rate. For a nightly job, Batch. For an internal tool where a retry is fine, Flex. For a customer waiting in a chat window, standard.

Which model for which job

OpenAI's own guidance, from its Codex pricing page, is that Sol is "built for the hardest work," Terra is "the everyday workhorse," and Luna is "optimized for fast, high-volume work." Here is how that maps onto the jobs I see in small companies.

The rule I give every client: Luna for volume, Terra as the default, Sol for judgment, Astra when the task earns it. The mistake I fix most often is a support bot on Sol because someone picked the top of the list. Since September 22, the same rule reads cheaper with the GPT-6 names: GPT-6 Luna ($0.10 / $0.50) for volume, GPT-6 Sol ($2 / $10) as the default and for judgment calls, Astra when the task earns it. The table above keeps the GPT-5.6 rows because they are still what most running setups are on.

What three owner-sized workloads actually cost

Arithmetic on the September 2026 rates. The token counts are assumptions from real installs; swap in yours.

1. A support-drafting bot. It reads the incoming ticket plus your policies and drafts a reply for a human to approve. 10,000 drafts a month, 4,000 tokens in (3,000 of it a stable system prompt and knowledge base), 400 tokens out. That is 40M input, 4M output.

• Terra, no caching: 40M × $2 + 4M × $12 = $128 a month
• Terra, system prompt cached: 10M × $2 + 30M × $0.20 + 4M × $12 = about $74
• Luna, cached: 10M × $0.20 + 30M × $0.02 + 4M × $1.20 = about $7.40
• Sol, no caching: $240. Astra, no caching: $600. Neither drafts a noticeably better "your order shipped Tuesday."

Run a 200-ticket sample on Luna and Terra, have the approver score them, and let the numbers pick.

2. A document pipeline. 5,000 invoices, contracts, or intake forms a month, 8,000 tokens in, 500 out, and nobody waiting on the result. 40M input, 2.5M output, which means Batch.

• Luna standard: $8 + $3 = $11; on Batch, $5.50
• Terra standard: $80 + $30 = $110; on Batch, $55
• Sol standard: $160 + $50 = $210; on Batch, $105

This is the workload where model choice is a real cost decision, and the one where accuracy is measurable because the ground truth is in the accounting system. Measure the error rate, price the errors, then pick.

3. A nightly report. One run a night that pulls 60,000 tokens of exported data (yesterday's orders, tickets, pipeline changes) and writes a 2,000-token summary for the owner's inbox. 30 runs a month, 1.8M input, 60K output.

• Terra: 1.8M × $2 + 0.06M × $12 = $4.32 a month
• Sol: $7.20 + $1.20 = $8.40
• Astra: $18 + $3 = $21

At this volume the model is a rounding error. Use Sol or Astra; the report is read by the highest-paid person in the company and it needs to be right. If it also searches the web, add $10 per 1,000 calls, which at 30 calls a month is 30 cents.

Across all three: a small business running a support bot, a document pipeline, and a nightly report on sensible models spends between $20 and $200 a month on the OpenAI API. The people spending $2,000 are running everything on Sol with no cache and no Batch. The full cost picture for a business, seats and setup included, is on its own page.

ChatGPT vs the API: when a $20 seat beats it

OpenAI's pricing FAQ says it plainly: API usage is billed separately from ChatGPT Plus, Business, Enterprise, and Edu. So the question is not "which is included" but "which meter fits the job." A ChatGPT seat is $20 a month for Plus, and $20 a user billed annually or $25 monthly for a Business Standard seat, with Premium seats at $100 carrying 5x the usage; the seat math is on ChatGPT Business pricing.

Here is the comparison that matters. A person who uses ChatGPT hard, say 50 messages a day for 22 working days with about 2,000 tokens of context in and 500 out per message, moves roughly 2.2M input and 0.55M output tokens a month. On Sol at API rates that is $8.80 plus $11, about $20. The same money as the seat, and the seat also brings connectors, memory, file uploads, projects, image generation, the admin console, and the no-training default, none of which you have to build. For a human using AI, the seat wins, and it wins harder the more they use it, because the seat's everyday chat is unlimited and the API meter never stops.

The API wins when there is no human in the loop. A bot that answers tickets at 2am, a pipeline that processes every invoice, a report that writes itself: none of those can hold a ChatGPT seat, and at Luna or Terra prices the API costs a fraction of one. The install I do for most owner-led businesses is seats for the team plus one API key for the two or three workflows that run on their own, with a monthly budget set in OpenAI's billing settings so the meter has a ceiling.

OpenAI API pricing vs Anthropic API pricing

Tier by tier, as of late September 2026, the two vendors are close, and both moved on September 22. GPT-6 Sol at $2 / $10 is now priced exactly like Claude Sonnet 5 ($2 / $10), with GPT-5.6 Terra at $2 / $12 beside them. GPT-5.6 Sol at $4 / $20 matches Claude Opus 5.5, which Anthropic launched the same day at $4 / $20 (Opus 5 stays listed at $5 / $25). GPT-6 Astra and Claude Fable 5.1 are both $10 / $50. At the cheap end OpenAI wins on sticker: GPT-6 Luna at $0.10 / $0.50 is a tenth of Claude Haiku 4.5 at $1 / $5. Both discount Batch by 50% and bill most cache hits at 10% of input (Anthropic charges 5% on Opus 5.5 and 2.5% on Fable 5.1), so the arithmetic above transfers. The Claude side is on Anthropic API pricing; the head-to-head on tool-call reliability, long context, and the routing pattern most production teams end up with is Claude API vs OpenAI API. Short version: the vendor is a smaller lever than the model tier, the cache, and the queue.

The business-owner answer

The OpenAI API is cheap enough now that "can we afford it" is the wrong question for almost every workflow a small business would run. The right questions are which model, cached how, batched when, and with what ceiling. Get those right and the three workloads above run for less than one software seat. Get them wrong and you pay Astra prices for Luna work. If you want the workflow-by-workflow model and cost plan written down for your business, that is the AI Readiness Assessment: $2,500 flat, two weeks, credited in full against a build within 90 days. If the real question is which of these your company should be running at all, that conversation is free on a strategy call.