The cheapest AI call is the one you structure well. Seven techniques, most of them documented by the providers themselves, let teams cut API spending without switching to a worse model. They matter more now: Ramp’s latest data, covered in our story on flat AI token spend, shows that sophisticated buyers are already using several of them.
Prices change often, so this list names mechanisms and the published discounts we could verify in the docs on November 7, 2026. Always check each provider’s pricing page before budgeting.
1. Cache the part of the prompt that never changes
Many requests start with the same instructions, tool definitions or reference documents. Prompt caching lets the API resume from that shared prefix instead of reprocessing it. OpenAI says caching is enabled by default on supported models and that cached input tokens are discounted by up to 95%, depending on the model. Anthropic describes explicit cache breakpoints: cache reads cost 0.1 times the base input price, while writes cost 1.25 times for the default 5-minute lifetime and 2 times for a 1-hour lifetime.
The practical rule is to put stable content first and variable content last, because a cache hit needs the whole prefix to match.
2. Send non-urgent work through a batch endpoint
If nobody is waiting for the answer, batch it. OpenAI’s Batch API offers a 50% cost discount versus synchronous calls, a separate pool of higher rate limits and completion within 24 hours. Anthropic’s batch processing works the same way: you submit many requests together, most finish within an hour, and results are available once all complete or after 24 hours. Good candidates are nightly classification, document summarization and evaluation runs.
3. Route simple tasks to cheaper models
Not every request needs the flagship model. Ramp’s economist told TechCrunch that advanced users “offload some of the spend from standard and light models to even cheaper light models via routers.” A simple router, or a first pass with a small model that escalates only when confidence is low, can move a large share of traffic to lower price tiers.
4. Cap output and reasoning effort
Output tokens usually cost more than input tokens, so set sensible max_tokens limits and ask for concise formats such as JSON with only the fields you need. Reasoning models add another dial: Anthropic’s docs list an effort setting, and OpenAI’s examples show reasoning effort set to none for simple jobs. Lower effort on easy tasks saves tokens that would otherwise be spent thinking.
5. Trim the context you send
Long conversations resend history every turn. Summarize older turns, drop tool outputs you no longer need and retrieve only the passages that matter instead of pasting whole documents. Fewer input tokens cost less and are also faster.
6. Set budgets and defaults
Ramp’s Ara Kharazian said mature AI users block expensive models by default for long-running tasks and impose per-employee budgets. Do the same: a default model, a spending cap per team and an alert before a runaway agent loop burns through a month of budget. Anthropic’s enterprise token spending reports show the kind of visibility that makes this possible.
7. Measure cost per outcome, not per token
A cheaper model that needs three retries can cost more than a stronger one that succeeds the first time. Track cost per resolved ticket, merged pull request or accepted answer. Benchmark sites increasingly publish cost per rollout next to accuracy, which is the right unit for this comparison.
Where to start: a sensible order of operations
Not every technique pays off equally for every team, so work from the cheapest change to the most invasive. Start with measurement: before changing anything, find out which features or teams generate most of your tokens, because a handful of workflows usually account for most of the bill. Then apply the changes that do not alter model behavior at all: caching stable prefixes and moving non-urgent jobs to a batch endpoint. These two keep outputs the same and only change how you pay.
Next come the changes that can affect quality: trimming context, capping output length, lowering reasoning effort and routing to smaller models. For each, run the same evaluation set before and after and compare success rate as well as cost. Finally, put guardrails in place, with budgets, alerts and defaults, so that savings are not erased by a runaway loop or a new feature that quietly switches everyone to the most expensive model. Revisit the list every quarter, because providers change cache lifetimes, discounts and model tiers regularly, and a setting that was optimal in the spring may not be in the winter.
What to watch next
- New cache lifetimes and discounts from the major providers.
- Further price cuts in the OpenAI and Anthropic price war.
- Wider adoption of routers and small specialist models.
- Provider dashboards that show cache hit rates and spend by team.
Key takeaways
- Stable prefixes plus caching are the highest-leverage change for repeated prompts.
- Batch endpoints halve cost for OpenAI jobs that can wait up to 24 hours.
- Routing, output caps and budgets stop waste before it starts.
- Judge models by cost per successful outcome.
- Verify current prices on official pages before committing to a budget.
Frequently asked questions
What is prompt caching?
Prompt caching reuses the processing of an identical prompt prefix across requests, so repeated instructions or documents are billed at a reduced cached-input rate and return faster.
How much does the OpenAI Batch API save?
OpenAI documents a 50% cost discount compared with synchronous APIs, with each batch completing within 24 hours.
Will cheaper settings lower answer quality?
They can, which is why you should test on your own tasks first. Caching and batching do not change the model’s output, while routing and lower effort might.
