OpenAI API Pricing Explained (2026 Guide)

GPT-6 Astra now leads OpenAI's frontier lineup for the hardest end-to-end and agentic work. The GPT-5.6 family remains available in three cost and capability tiers: Sol for flagship professional work, Terra for balance, and Luna for cost-sensitive, high-volume workloads.

At Standard short-context rates, Astra costs $10.00 input and $50.00 output per 1M tokens; Sol costs $4.00 input and $20.00 output; Terra costs $2.00 input and $12.00 output; Luna costs $0.20 input and $1.20 output.

Astra supports reasoning effort from low through max, while the GPT-5.6 models also support none. The visual comparison above is a price snapshot checked on September 9, 2026, not a live billing feed. See our GPT-6 Astra API guide for long-context pricing and migration details.

OpenAI charges per token, not per request. A token is roughly 4 characters of English text, or about 3/4 of a word. A 1,000-word English article is approximately 1,300 tokens. The complete OpenAI guide provides broader context on how pricing fits into the overall API workflow.

Every API call has two token counts: input tokens (the prompt you send) and output tokens (the model's response). These are billed at different rates — output tokens are typically 3-4x more expensive than input tokens.

Chinese text uses more tokens per character than English because tokenizers are optimized for Latin scripts. A Chinese character typically costs 1-2 tokens, so factor this in when estimating costs for Chinese-language applications.

GPT-4o is OpenAI's flagship multimodal model, supporting text, image, and audio inputs. GPT-4o-mini is the cost-optimized version, recommended for most general-purpose tasks.

The o-series models (o1, o3, o4-mini) are designed for complex reasoning tasks like math, coding, and scientific analysis. They use internal reasoning tokens that are billed as output tokens, so workload cost depends on more than the visible answer length.

Because reasoning models generate extensive internal chain-of-thought before producing the final answer, the actual output token count can be much higher than the visible response length. Always test with real workloads to understand the actual cost.

OpenAI offers prompt caching for repeated prefixes. If your API requests share the same leading prompt (e.g., system instructions, few-shot examples), the cached portion is billed at a 50% discount for GPT-4o models and 75% for o-series models.

Caching is automatic — you don't need to configure anything. The cache typically persists for 5-10 minutes of inactivity, and the minimum cacheable prefix is 1,024 tokens. For applications with long system prompts, caching can significantly reduce costs.

If your use case doesn't require real-time responses, the Batch API offers a 50% discount on all models. You submit a batch of requests (up to 50,000 per file), and OpenAI processes them within 24 hours (typically much faster).

Batch API is ideal for bulk processing tasks like content classification, document summarization, data enrichment, and evaluation runs. The request format is identical to the standard API — you only need to wrap your requests in a JSONL file and submit it via the batch endpoint. If batch requests fail with 429 errors, troubleshooting rate limit errors can help you identify whether the issue is concurrency or quota.

Processing tier changes both price and delivery characteristics. Standard is the baseline for normal synchronous requests. Batch and Flex use discounted rates for workloads that can accept delayed or variable processing. Fast mode charges a premium for prioritized processing.

For GPT-5.6 Sol, Terra, and Luna, Batch and Flex short-context prices are 50% of Standard. Fast mode is 2x Standard for these three models. Select a tier based on latency requirements rather than treating the discounted tiers as temporary promotions.

Not every OpenAI service is billed as text tokens. Realtime audio can have separate audio, text, and image token rates; transcription is often estimated per minute; image generation uses image and text token rates; and Sora video generation is priced per second by resolution.

Built-in tools can add charges on top of model tokens. Web search is priced per call plus applicable search-content tokens, file search combines storage and tool-call charges, and hosted containers are billed by memory size and session duration. Include both model and tool usage in production budgets.

To estimate your monthly API cost, multiply your expected input tokens per request by the input price, plus the expected output tokens by the output price, then multiply by the number of requests per month.

For example, using GPT-4o-mini with an average of 500 input tokens and 200 output tokens per call, at 10,000 calls per month: (500 × $0.15 + 200 × $0.60) / 1,000,000 × 10,000 = $1.95/month. This makes GPT-4o-mini extremely affordable for most applications.

For reasoning models, be extra careful with estimates. A complex coding question on o3 might generate 5,000+ hidden reasoning tokens, making a single call cost $0.20+ instead of the $0.01 you might expect from the visible output alone.

Credits, payment requirements, and account eligibility can vary. Check the Platform billing page for the terms attached to your own account rather than assuming a public trial amount. If you have not created an OpenAI API key yet, our step-by-step guide covers registration and billing setup.

Set project budgets and alerts, restrict production keys, and monitor usage by model and service tier. Review configuring your key securely before deploying. Prices on this page are a dated editorial snapshot; the official pricing and billing pages control your final charge.