OpenAI API Pricing in 2026: Per-Token Costs for GPT-5, GPT-4o, DALL-E, Sora and Whisper

GPT-5 on the API: when the flagship is worth 5x the price

GPT-5 at about 1.25 dollars per 1M input tokens and 10 dollars per 1M output tokens is the most capable model in the OpenAI API as of May 2026. The instinct for many developers is to make GPT-5 the default. In practice, that burns money on every routine call. The right move is to default to GPT-5 mini for most traffic and reserve GPT-5 for calls where the task is hard enough that mini misses. In our stack, only about 20 percent of calls route to GPT-5 directly, and another 10 percent are escalations from a mini call that failed a quality check. The rest stay on mini or nano.

GPT-5 mini at 0.25 dollars per 1M input: the real production workhorse

GPT-5 mini at 0.25 dollars per 1M input tokens and 2 dollars per 1M output tokens is the model most chatbots, copilots, and content tools should be running on. It is 5x cheaper than GPT-5 on input and 5x cheaper on output. Quality is close enough on routine work that most users cannot tell the difference. Pair it with prompt caching for a stable system prompt and you are paying half the input rate on most calls. Pair it with the Batch API for any non-realtime work and you are paying half again. Stacking those discounts moves the effective input cost into the cents-per-million range.

GPT-5 nano at 0.05 dollars per 1M input: when you only need a routing decision

GPT-5 nano at 0.05 dollars per 1M input tokens and 0.40 dollars per 1M output tokens is built for high-volume narrow tasks. Classification, intent routing, tag extraction, content moderation, simple summaries. At this price you can embed an LLM call in places where the unit economics would not have made sense at flagship pricing. A million classification calls on nano cost about 50 dollars total assuming short prompts. The same workload on GPT-5 is roughly 25 times more expensive.

GPT-4.1 and GPT-4o on the API: still relevant in 2026

GPT-4.1 at 2 dollars per 1M input and 8 dollars per 1M output is the prior flagship. Apps that validated their prompts and evals against GPT-4.1 often keep running on it because re-validation costs engineering time. GPT-4o at 2.50 dollars per 1M input and 10 dollars per 1M output is still the right answer when your app sends mixed text, image, and audio inputs to a single endpoint. Neither is a default pick for new builds in 2026. New apps should start on GPT-5 mini and selectively upgrade to GPT-5 where needed.

DALL-E 3, Sora 2, Whisper, and TTS: pricing in the multimodal lineup

Image, video, and audio models price differently from text. DALL-E 3 is about 0.04 dollars per standard image and 0.08 dollars per HD image. Sora 2 video is billed per second of output, with the per-second rate varying by resolution. A 10 second high-resolution clip can cost more than a hundred DALL-E images, so budget control for video features should track seconds generated, not raw API calls. Whisper at 0.006 dollars per minute of audio is the cheapest workhorse in the entire API: a 60 minute call transcribes for about 0.36 dollars. TTS at about 15 dollars per 1M characters lets you build a full speech-in, speech-out loop on OpenAI infrastructure.

Embeddings and the cost of running RAG at scale

The text-embedding-3-small model at 0.02 dollars per 1M tokens is roughly two orders of magnitude cheaper than chat output. Embedding an entire 1M token corpus costs about two cents. That is what makes RAG economically viable at scale: you can re-embed your knowledge base on a regular cadence without making a meaningful dent in your monthly bill. The larger and higher-dimensional embedding models cost more but follow the same pricing pattern. Plan your RAG budget around inference (chat completion) cost, not embedding cost.

Batch API vs realtime: a 50 percent discount for 24 hours of patience

The Batch API gives a flat 50 percent discount on both input and output tokens compared to realtime endpoints. The trade is processing time: jobs complete within a 24 hour window instead of in real time. Submit a JSONL file of requests, poll for completion, get the results back. Use cases that fit perfectly: nightly classification of yesterdays content, bulk summarization of a document backlog, weekly report generation, monthly embedding refreshes of a large corpus. The Batch API is one of the biggest cost levers most teams underuse. Walking through your workload and tagging anything that does not need to be realtime is usually a quick path to cutting the bill by 20 to 40 percent.

Prompt caching: the discount you get without writing any code

Prompt caching automatically detects when a long prefix of your prompt is the same as a recent call (within a 5 to 10 minute window) and bills that cached portion at 50 percent of the normal input rate. There is no API flag to enable, no code change required. It is auto-applied. The savings are largest when you have a long, fixed system prompt with variable user input appended at the end. To maximize it: put stable content first (system prompt, retrieved documents, examples) and variable content last (user message, current question). Many production chatbots see 30 to 60 percent reduction in input token cost from prompt caching alone.

Cached input savings math, worked out

Take a chatbot that sends a 3,000 token system prompt plus 200 tokens of user input per call on GPT-5 mini. Input cost without caching: 3,200 tokens times 0.25 dollars per 1M = 0.0008 dollars per call. With caching applied to the 3,000 token system prompt, the cached portion is billed at 0.125 dollars per 1M while the new 200 tokens are billed at the full 0.25 dollars per 1M. New per-call input cost: about 0.000425 dollars. That is a 47 percent reduction in input cost, achieved with zero code changes, just by keeping the system prompt stable across calls. At a million calls per month, that is roughly 375 dollars in input cost saved per month on just one model.

When self-hosting beats the API

Self-hosting open-weight models like Llama 4 or Qwen on your own GPUs only beats the OpenAI API at very specific scale and use-case profiles. The break-even is roughly: sustained traffic above 50 million output tokens per day with steady utilization, a narrow workload that a smaller open model handles well, and an engineering team that already owns the infra muscle. For everything below that bar, the OpenAI API is cheaper after factoring in GPU rental, ops staff time, model evaluation, and the lost capability of GPT-5. Most teams should stay on the API until they have both the volume and the workload pattern that makes self-hosting pay back. Verified May 2026.

What we actually pay for in our own stack

Across our own production stack, the routing mix is roughly 70 percent GPT-5 mini, 15 percent GPT-5 nano (for classification and routing layers), 10 percent GPT-5 (for the hardest reasoning calls), and 5 percent everything else (embeddings, Whisper, TTS, DALL-E). Prompt caching is auto-applied across all chat models. The Batch API handles our nightly summarization and weekly report generation. The combined effect is that our average effective input cost sits around 0.10 dollars per 1M tokens, well below the headline GPT-5 mini rate of 0.25 dollars. That is the kind of number you should be targeting once you have the discount stacking dialed in.

The verdict: which OpenAI API model should you pick

My verdict after building on this API for years: default new builds to GPT-5 mini, route classification and intent calls to GPT-5 nano, and reserve GPT-5 for the hardest reasoning. Always turn on prompt caching by structuring your prompts correctly, and route everything that does not need a realtime response through the Batch API for the 50 percent discount. Start on the starter credit, hit Tier 1 after spending 5 dollars, and let the tier ladder lift you naturally as production traffic accumulates. The single biggest mistake I see new developers make is calling GPT-5 by default for everything. The single biggest cost win is moving to mini and stacking the caching and batch discounts. Verified May 2026.