Most OpenAI bills are 3 to 5x larger than they need to be, and the fix rarely involves touching your product. The API ships with two built-in discounts most teams never turn on, a lineup of cheaper models that handle routine calls, and a market of alternatives for work that does not need a frontier model at all. Here are the levers, ordered by effort against savings.
1. Turn on the Batch API (50 percent off, one afternoon)
Anything that does not need an instant answer belongs on the Batch API: embeddings backfills, nightly classification, content pipelines, evals, bulk extraction. You upload requests, results come back within 24 hours, and everything is billed at half price. Flat. This is the single biggest effortless cut, and it applies to the expensive models where halving matters most. Mechanics and code patterns are in prompt caching and batch discounts.
2. Cache your prompt prefix (automatic, if you structure for it)
OpenAI's prompt caching discounts repeated input prefixes automatically once your requests share an identical opening. The rule: static content first (system prompt, instructions, few-shot examples, reference documents), variable content last. If your system prompt is thousands of tokens and you serve real traffic, caching alone can cut input spend dramatically, and input is often the majority of a chat workload's tokens.
Two implementation notes:
- Byte-identical matters. A timestamp or user name interpolated into the system prompt breaks the cache for every request.
- Order matters. Put the user-specific context after the shared block, not woven through it.
3. Right-size the model (the biggest single multiplier)
The gap between tiers is not subtle. Per 1M output tokens as of our July 2026 check, GPT-5.5 lists at 30 dollars while GPT-5.4 mini lists at 4.50, and the nano-class tiers go lower still. Classification, extraction, formatting, summarization, and most routing decisions do not need the flagship. A default-to-mini policy with escalation on failure typically moves 70 to 90 percent of calls down a tier without measurable quality loss. Current per-model prices: GPT-5.5, GPT-5.4 mini, GPT-4.1.
4. Route the truly routine work off OpenAI entirely
Open-weight models now serve routine calls at cents per million tokens. A 120B-class open model costs around 2 percent of flagship rates, and for internal tools, drafts, and agent scaffolding the quality difference often does not survive a blind test. The full pricing comparison has the numbers, and OpenRouter vs direct covers running a mixed fleet behind one key.
5. Cap the output, not just the prompt
Output tokens cost several times input. Three habits pay immediately:
- Set max output tokens per call type instead of leaving it open.
- Ask for structured output (JSON with a schema) rather than prose; it is shorter and easier to validate.
- Stop sequences and terse-format instructions beat asking the model to be brief in a paragraph of politeness.
For reasoning models, remember thinking tokens bill as output: use effort controls where offered, and do not send easy work to a reasoning model.
6. Trim the conversation you resend
Chat history grows quadratically expensive because every turn resends everything. Summarize old turns, drop tool-call transcripts after they resolve, and keep retrieval snippets tight. A context budget per request (with a summarizer that enforces it) routinely halves input volume on chat products.
7. Do not pay cash you did not have to
New OpenAI accounts and startup programs come with credits, and the platform programs are worth five figures to qualifying companies. Claim before you optimize; the startup credits checklist sequences them, and free AI API credits covers the signup-tier options across providers.
What 80 percent actually looks like
A workload spending 1,000 dollars a month at flagship list prices typically lands here: batch moves a third of volume to half price, caching cuts chat input costs, right-sizing moves most calls to a mini tier, and open-model routing absorbs the routine remainder. The blended result lands between 150 and 250 dollars for the same output, which is where the 80 percent claim comes from, and none of it required product changes users can see.
Bottom line
Batch and caching are free money on the bill you already have. Right-sizing and routing are one policy decision each. Track the moving prices in the live rankings, claim the credits in the catalog before spending cash, and revisit quarterly because every number here drifts.
Related reading: prompt caching and batch discounts, LLM API pricing comparison, the cheapest way to run LLMs.