Cut Your OpenAI API Bill: Practical Cost Optimisation Guide

The single biggest cost lever is choosing the right model. GPT-5-mini costs $0.25 per million input and $2.00 per million output, compared to $1.25 and $10.00 for GPT-5. That is an 80% reduction in both input and output costs. For most standard tasks (summarisation, classification, simple Q&A, content generation), GPT-5-mini produces comparable quality. Reserve GPT-5 for complex reasoning, nuanced analysis, and tasks where quality degradation is unacceptable. Many production systems use GPT-5-mini for 80-90% of requests.

The Batch API offers a flat 50% discount on all token costs in exchange for asynchronous processing within a 24-hour window. Any workload that does not require instant responses should use batch processing. Content generation pipelines, nightly data processing, email summarisation queues, and classification tasks are all excellent candidates. The output quality is identical to real-time requests. A $2,000/month real-time workload becomes $1,000/month on batch with zero quality trade-off.

When your requests share a common prefix (system prompt, few-shot examples, RAG context), OpenAI automatically caches the processed tokens and charges 50% less for cached input tokens. This is especially impactful for chatbots with long system prompts, RAG applications, and any workflow where the beginning of each prompt is consistent. No configuration needed - caching happens automatically. For applications with 2,000-token system prompts, this can save 30-50% on input costs.

Every token in your prompt costs money. Techniques to reduce prompt size include removing unnecessary instructions, using concise phrasing, eliminating redundant examples, and structuring prompts efficiently. A prompt that says 'Please provide a brief summary of the following text, keeping your response to approximately 3 sentences' can be shortened to 'Summarize in 3 sentences:' with identical results. Across millions of requests, these savings compound significantly.

Beyond OpenAI's built-in prompt caching, implement application-level semantic caching. When a user asks a question similar to one already answered, return the cached response instead of making a new API call. Tools like GPTCache, Redis with vector search, or custom embedding-based caches can achieve 20-40% cache hit rates for common question patterns. This eliminates API calls entirely for repeated or similar queries.

Output tokens are 2-8x more expensive than input tokens. Use the max_tokens parameter to limit response length, and instruct the model to be concise in your system prompt. A response limited to 200 tokens costs significantly less than one that runs to 2,000 tokens. For many tasks (classification, extraction, short answers), responses of 50-200 tokens are sufficient. Setting appropriate limits prevents the model from generating unnecessarily verbose responses.

Fine-tuning a model on your specific task allows you to use shorter prompts (since the model already knows your domain) and get better results with a smaller model. A fine-tuned GPT-4o-mini can often match or exceed a standard GPT-5 on domain-specific tasks. The upfront cost of fine-tuning is offset by dramatic per-request savings when you eliminate lengthy system prompts and few-shot examples from every API call.

When you need JSON or structured data from the API, use structured outputs (json_schema) to guarantee valid output format. Without structured outputs, a significant percentage of responses may have formatting errors that require retries, wasting tokens. Structured outputs eliminate this waste by ensuring every response matches your schema on the first attempt. This is particularly impactful for data extraction and classification pipelines.

Not every request needs the same model. Implement a routing layer that analyses incoming requests and assigns them to the most cost-effective model that meets quality requirements. Simple tasks (classification, extraction, yes/no questions) go to GPT-4o-mini. Standard tasks go to GPT-5-mini. Only complex reasoning and analysis tasks go to GPT-5. Many production systems route 60-80% of traffic to the cheapest adequate model.

Set up usage monitoring and spending alerts through the OpenAI dashboard. Configure daily and monthly spending limits. Track cost per request, cost per user, and cost per feature in your application. Anomaly detection can catch bugs that cause excessive API calls (infinite loops, retry storms, prompt injection attacks that inflate token counts). A single day of runaway costs can exceed your entire monthly budget if not caught quickly.