Top tips for reducing API costs
We run a daily autonomous research pipeline, spawning parallel AI subagents, for exactly $0/month in inference, and buy paid models only when a task genuinely needs the quality jump. These are the tips that did it, in the order that matters.
1. Rotate free models instead of depending on one
Free tiers are real: OpenRouter alone has offered dozens of free models, from Google Gemma to DeepSeek variants. The problem is not availability, it is reliability: every model has its own rate-limit bucket, and a burst exhausts one fast. Round-robin your calls across free models with a health probe, and reassign work when one returns 429s. Rotation, not capacity, is what makes free tier production-grade for batch work. Full method in our zero-cost AI agent toolkit.
2. Route by task, not by habit
Sending every call to your best model is like taking a taxi to the mailbox. In our own benchmarking, cheap models matched or beat expensive ones on summarization and structured extraction, while code generation favored the premium tier. Split your workload: drafting, extraction, and classification go to cheap or free models; the premium model gets code and final-pass judgment only.
3. Use an independent judge instead of a bigger model everywhere
A common cost trap is upgrading the whole pipeline to fix quality. Cheaper: keep workers cheap, add one stronger model as a judge that grades outputs and sends failures back. One judge call per batch beats paying premium rates on every call.
4. Cache and deduplicate aggressively
Prompt-caching discounts and provider-side context caching can cut input costs dramatically on repeated system prompts. Outside the API, the cheapest token is the one you never send: hash and reuse results for identical queries, and trim transcripts before resending them.
5. Batch and schedule around pricing
Some providers discount off-peak usage substantially; DeepSeek, for example, has offered reduced rates during off-peak hours. If your workload tolerates scheduling, nightly batch runs are cheaper than real-time calls, and they are usually all an autonomous pipeline needs.
6. Set a budget and a freeze rule
- Know your rate-limit mechanics: many free tiers are limited per account, not per key, so extra keys do not help.
- Track usage where the provider lets you; where it does not (free tiers often hide usage from APIs), your 429 rate is the meter.
- Have a freeze rule: decide in advance which part of your stack gets cut when a free model dies or turns paid, and rehearse the respawn-with-cooldown path instead of retry storms.
What not to do
- Do not create extra accounts to dodge quotas. It violates provider terms and does not raise per-account limits.
- Do not build latency-sensitive products on free tier. Free hosting churns; models disappear mid-run. It is for batch research, not production SLAs.
- Do not optimize before measuring. Log tokens and cost per task for a week; the biggest line item is usually a surprise.