OpenAI API Pricing Explained (2026): Cost per Token & Real Examples

Executive Summary

OpenAI's ecosystem has two distinct billing worlds: ChatGPT (subscriptions) and the OpenAI API (usage-based). ChatGPT plans are priced per user/month (Plus, Pro, Business), while the API is metered primarily by tokens (input, cached input, output) and by non-token units when you use tools (web search, file search), storage (vector stores), or sandboxed execution (containers / code interpreter).

For typical production workloads, the largest cost levers are: output tokens (including reasoning tokens for applicable models); model choice and processing tier (Standard vs Batch vs Flex — Batch is 50% cheaper with 24h turnaround); tool calls and storage (web search, file search, containers); and very long context (pricing multipliers for 1M-context models).

For the sample chatbot workload (100k messages/month, 1k tokens/message, 75% input / 25% output), estimated monthly spend is ~$13.75 (GPT-5 Nano), ~$68.75 (GPT-5 Mini), ~$343.75 (GPT-5), or ~$562.50 (GPT-5.4) at Standard pricing.

How OpenAI API Pricing Works

OpenAI bills for input and output tokens. Cost modifiers include:

• Prompt caching: Cached input is ~10x cheaper than fresh input for most models
• Batch API: 50% discount on input and output; 24h completion window
• Flex processing: Batch rates plus prompt caching discounts; slower, may return "resource unavailable"
• Priority processing: Paid tier for lower, more consistent latency

API accounts use prepaid billing: minimum $5 purchase, credits expire after 1 year (non-refundable).

Token Pricing Table (Standard)

Output pricing explicitly includes reasoning tokens where applicable — "thinking longer" can materially increase spend even when visible text output is short.

Batch API: 50% Cost Reduction

Batch API is designed for non-real-time jobs. You save 50% on both input and output tokens compared to real-time API calls. Ideal for summarisation, classification, extraction, and embeddings. Jobs complete within 24 hours.

Embeddings

Embeddings are priced per input token (no output).

Cost per 1,000 vectors (512 tokens/vector): text-embedding-3-small $0.01024 (Standard) / $0.00512 (Batch); text-embedding-3-large $0.06656 / $0.03328.

Tools, Storage, and Containers

Hidden cost centers:

• Web search: $10/1K calls + search content tokens at model rates
• File search: Vector storage $0.10/GB/day; tool calls $2.50/1K
• Containers: From Mar 31, 2026: $0.03/container and $0.03/session (20-min session units)

Real Cost Examples

Chatbot: 100k messages/month, 1k token/msg, 75% input / 25% output

75M input + 25M output tokens monthly.

Embedding pipeline: 1M documents × 512 tokens

512M tokens. text-embedding-3-small: $10.24 (Standard) / $5.12 (Batch). text-embedding-3-large: $66.56 / $33.28.

Multimodal: 10k images/month (1024×1024, detail high)

50 text input + 200 output per image. Image tokenization varies by model (tile-based vs patch-based). GPT-5: ~$28.50/month. GPT-4o Mini can be expensive due to high image token footprint — use detail: "low" when possible. Use POST /v1/responses/input_tokens to get exact counts.

Sensitivity: Token Volume and Output Share

At 100k messages/month, cost varies with tokens per message and output share:

Shifting output from 15% to 40% increases cost significantly.

Hidden Costs and Constraints

Data residency: 10% uplift for gpt-5.4 and gpt-5.4-pro when using data residency endpoints (e.g. eu.api.openai.com).

Long-context multipliers: For 1M-context models, prompts above published token thresholds can incur higher input/output multipliers.

ChatGPT subscriptions (separate from API): Plus $20/month; Pro $200/month; Business $25–30/seat/month.

How to Reduce OpenAI API Costs

• Model routing: Default to GPT-5 Nano or GPT-5 Mini; escalate to GPT-5/GPT-5.4 only for complex reasoning or high-value requests.
• Constrain output: Cap max_tokens; control reasoning effort where available (reasoning tokens billed as output).
• Prompt caching: Cache system prompts, RAG chunks, repeated context. Cached input is ~10x cheaper.
• Batch or Flex for offline work: Batch gives 50% discount; Flex uses Batch rates plus caching.
• Optimise vision: Use detail: "low" when fine-grained vision isn't needed; resize images; use /v1/responses/input_tokens to validate.
• Minimise tool calls: Web search and file search add per-call fees. Gate tools behind intent classification.

Related

For a side-by-side comparison with Claude and Gemini, see AI API Pricing Comparison (2026).