OpenLuxHow to find the cheapest LLM API for your workload
The cheapest LLM API is the model-gateway combination with the lowest delivered cost for your actual token mix, not the service with the smallest headline number. Compare input and output token rates, image or video request pricing, retries, plans and operational overhead. OpenLux publishes its USD rate for every catalogued model, so you can estimate a workload from observable inputs.
Reviewed by OpenLux Engineering Team · Last verified
What makes one LLM API cheaper than another?
Five variables explain most of the difference between a low estimate and the final invoice.
| Cost driver | What to measure | Why it changes the answer |
|---|---|---|
| Model choice | Quality threshold and accepted model | A cheaper model with more retries or manual review can cost more in total |
| Token direction | Monthly input and output tokens | Input and output prices can differ; averaging them hides the workload shape |
| Request pricing | Image, video, audio or agent calls | Media models often bill per request or unit rather than per token |
| Gateway terms | Published rate, markup, plan allowance and fees | The provider list price is not the account price after all gateway terms |
| Operational cost | Retries, latency, engineering and reconciliation | A low unit price loses if it creates failures or manual operations |
How do you calculate monthly LLM API cost?
Use this formula: monthly model cost = input tokens × input rate + output tokens × output rate + request-based model cost. Then add retries and failed calls, subscription fees or minimum commitments, taxes and the labour required to integrate and reconcile the service.
Measure one week of representative prompts, then extrapolate by workload rather than by user count. A customer-support summariser and a coding agent can send the same number of requests with radically different context and output lengths.
Why is output price often more important than input price?
Many text-model catalogues price input and output separately, and long responses can shift the blended cost materially. Systems that read large documents but return short classifications are input-heavy; agents that plan, call tools and explain steps are output-heavy.
Log both directions from day one. If the provider or gateway exposes only a single blended number, build a representative benchmark that preserves the same prompt and response distribution before comparing alternatives.
When are request-based models cheaper?
Image, video, music and speech models often use request-based units. One image request can be cheaper and safer than routing a vision task through a long chain of text calls, while a video request can dominate an otherwise small monthly bill.
Price each modality separately. Use the OpenLux calculator below for an illustrative text workload, then combine it with the published request rate for media models in the catalogue. Never fold unlike units into one misleading average.
How do monthly plans and pay-as-you-go compare?
Pay as you go is usually better for spiky, experimental or hard-to-forecast usage because it avoids a minimum commitment. Monthly and yearly plans make sense when spend is predictable and the included allowance plus priority routing is worth the commitment.
OpenLux publishes both pay-as-you-go rates and monthly/yearly plans. The catalogue currently lists 19 models; the correct plan threshold depends on your blended model mix, not on the price of one flagship.
What hidden costs should you include?
Include failed-request volume, provider retry costs, observability, security review, idle capacity, support plans, reconciliation and the engineering time spent maintaining multiple accounts. Also verify how a gateway rounds token counts and whether caching changes the published rate.
A single invoice can be cheaper than three vendor invoices even when the unit model rate is similar. Finance effort and payment terms are legitimate parts of the comparison, especially for teams buying from multiple regions.
How can you reduce API cost without hurting quality?
Route by task: use smaller models for extraction and classification, reserve reasoning models for decisions, and cache deterministic answers where appropriate. Shorten prompts, remove duplicate context and measure whether extra tool steps improve outcomes enough to justify their cost.
Do not optimise only for cost per million tokens. Optimise cost per successful task. Track quality acceptance, retry rate and human review alongside spend so a cheap route does not silently move cost into operations.
Which option fits?
Use pay as you go when
- Traffic is new, spiky or difficult to forecast.
- You want to compare several models before committing.
- There is no operational benefit to a monthly allowance yet.
Compare a plan when
- Monthly spend is stable and close to the plan value.
- The included allowance and routing benefits are usable.
- Yearly billing savings justify the commitment.
Calculate a monthly model bill
Pick a catalogue model and enter the units you expect to consume. The result uses the published USD rate and excludes taxes, discounts or plan credits.
Related models and live rates
| Model | Vendor | Price | Vs. official |
|---|---|---|---|
| gemini-3-flash-preview | $0.0368 / 1M tokens | -92% | |
| deepseek-v4 | DeepSeek | $0.0613 / 1M tokens | -90% |
| qwen3-max | Qwen | $0.0784 / 1M tokens | -88% |
| gpt-5.6-sol | OpenAI | $0.1838 / 1M tokens | -96% |
Key takeaways
- Compare cost per successful task, not cost per million tokens in isolation.
- Measure input and output tokens separately and keep media request costs separate.
- Use pay as you go while demand is uncertain, then move to a plan only when the economics are proven.
Frequently asked questions
What is the cheapest LLM API right now?
No single answer applies to every workload. The lowest delivered cost depends on the model, input/output mix, media requests, retries, plan and operations. Compare candidates with the same workload sample.
Is a cheaper model always less expensive to run?
No. A cheaper model can require more retries, longer prompts or manual review. Measure cost per successful task, not just the price per token.
How do I estimate image and video API costs?
Use the request-based rates in the model catalogue. Estimate the number of successful outputs, add retry allowance and keep media costs separate from token-based text costs.
Does OpenLux charge per API key?
Review the current pricing page for plan and key terms. Calls are metered against the selected catalogue model, and monthly plans may include token-equivalent allowance.
When should I use a monthly plan?
Use a plan when stable monthly spend and included allowance make it more economical than pay as you go. Keep usage visibility so the plan does not hide a rising cost per task.
Sources and methodology
Competitor claims were checked against the linked public pages on the verification date. Pricing and product terms can change; confirm them before purchase.
- OpenLux model catalogue and ratesChecked alongside the 19-model catalogue on 2026-10-30.
- OpenLux pricingPay-as-you-go metering and subscription terms.
- OpenAI API pricingOfficial upstream pricing reference.
- Anthropic Claude pricingOfficial upstream pricing reference.
- Gemini API pricingOfficial upstream pricing reference.
- DeepSeek API pricingOfficial upstream pricing reference.
Related comparisons and guides
Estimate your own monthly workload
Use the calculator, then check the per-model catalogue before choosing a plan.