OpenLuxAI API gateway pricing: how token, request and plan costs work
An AI gateway bill has two layers: the model usage consumed by each request and the gateway or platform fee charged on top. OpenLux currently publishes a USD catalogue rate for each of its 19 models, meters usage against that rate and does not add an undisclosed gateway multiplier. Compare a worked monthly example before choosing a plan.
Reviewed by OpenLux Engineering Team · Last verified
Which billing model does your workload use?
Map the unit you consume to the invoice before comparing vendors. The lowest per-token price is not automatically the lowest monthly bill.
| Billing unit | Typical workload | What drives the bill | What to verify |
|---|---|---|---|
| Input and output tokens | Chat, reasoning, coding agents, classification and retrieval generation | Prompt size, response length, context reuse and model rate | Whether input, output and cached tokens are priced separately |
| Per request | Image generation, video generation and music generation | Number of calls, output resolution, duration or quality | Whether a request can consume multiple internal units |
| Per 1K characters | Text-to-speech and voice generation | Characters synthesised and requested voice or format | Minimum charge, character rounding and audio format |
| Subscription plus allowance | Predictable monthly product usage | Plan fee, included credits, overage and commitment length | Whether unused allowance expires and how overage is billed |
| Gateway fee on top of provider cost | Bring-your-own-key or BYOK gateway products | Provider usage plus percentage, per-request or platform fee | Whether observability, guardrails or support are charged separately |
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.
How we evaluated this
The calculator uses the current single published catalogue rate for each OpenLux model. It is an estimate, not a quote: model billing can distinguish input, output and cached tokens, and the console bill is the authoritative record. Taxes, promotions and plan credits are excluded.
Competitor pricing is described by model type rather than reproduced as a list of volatile numbers. Every provider changes packaging and rates frequently; follow the official pricing links below on the day you buy.
How do you compare AI API prices without being misled?
Model the bill at your real volume. A representative calculation catches more problems than a rate-card screenshot.
- Use your actual 30-day input, output, image, video and speech volumes rather than a generic token estimate.
- Separate input, output, cached and reasoning token rates when the provider publishes them.
- Include the gateway fee, plan fee, payment currency and any minimum commitment.
- Check whether unused subscription credits expire and whether overage has a different rate.
- Estimate retries, failed requests and evaluation traffic, which still consume resources.
- Add finance and engineering time if you self-host routing, logging or billing.
- Request an invoice example showing every field that will appear on the monthly bill.
- Run a small real workload and reconcile the console usage record with the invoice.
What makes up an AI API gateway bill?
The first layer is model consumption. Text models usually charge by input and output tokens, image and video models by request or generated output, and speech models by characters or audio duration. The second layer is the access provider: a hosted gateway may add a platform fee, plan fee or markup, while a self-hosted gateway moves that cost into infrastructure and engineering.
A useful quote separates these layers. If a vendor advertises a low model rate but adds a percentage platform fee, a minimum monthly spend or an observability charge, the effective cost at your volume can be higher than a simpler published rate.
How does token pricing work?
A token is a unit of model input or output. A long prompt, retrieved context, tool schema and conversation history all contribute input tokens; the generated answer contributes output tokens. Input and output can have different prices, and some providers offer lower-cost cached input when the same prefix is reused.
Calculate a monthly estimate as separate input and output lines whenever possible. If the provider publishes one blended catalogue rate, state that limitation clearly and use the console data from a pilot to refine the estimate.
How does per-request pricing work?
Image, video and music models often bill per successful request. The unit can hide significant differences: one image request may vary by resolution, quality or editing operation; one video request may vary by duration and output format. Read the model-specific notes, not only the headline request price.
If your workflow retries failed generations or creates several candidates before selecting one, multiply the request estimate by the expected attempts and selection rate. A cheap test call can become an expensive production habit.
Is pay as you go cheaper than a monthly plan?
Pay as you go is cheaper when usage is spiky, experimental or genuinely low. You pay for what you consume and avoid committing to capacity that may go unused. A plan is useful when your baseline usage is predictable and the included allowance would be consumed anyway.
Compare the plan against the forecast, not against a single busy day. Include unused-credit rules, rollover, overage, cancellation and the value of any additional support or concurrency. A plan should beat a straightforward usage estimate on total value, not merely look cheaper than the largest possible monthly bill.
What hidden costs should you check?
Look for percentage gateway fees, minimum monthly spend, top-up expiry, payment-method fees, currency conversion, support tiers, audit logs, data-retention add-ons and charges for features that appear in the product tour. A self-hosted gateway also has compute, storage, monitoring, upgrades and on-call time.
The most expensive hidden cost is often integration rework. A gateway that does not preserve streaming, tool-call or multimodal behaviour forces engineering time even if its token price is attractive. Treat compatibility as part of the total cost.
How are OpenLux model prices organised?
The OpenLux catalogue publishes a USD rate and billing unit for each model. Text, embedding and rerank models are listed per million tokens, image/video/music models per request, and speech per thousand characters. The same catalogue is the reference for metered account usage.
The calculator above converts expected monthly units into a list-rate estimate. Use it for budgeting and model selection, then use the console’s per-request records to reconcile actual usage. See the pricing page for plan fees and any current promotions.
What does a 10M-token month cost?
Using the published gpt-5.6-sol catalogue rate of $0.1838 per million tokens, 10 million units is approximately $1.84 before taxes or plan credits. The same volume on a premium model costs more because the published model rate is different; the workload is constant while the model choice changes the price.
This example deliberately uses a blended catalogue unit. If a provider distinguishes input and output pricing, split the 10 million into your expected prompt and response mix before comparing quotes.
What does a 100M-token month cost?
At $0.1838 per million tokens, 100 million units is approximately $18.38 at the same published rate. At a premium rate such as $0.4412 per million tokens, the estimate rises to $44.12. The difference is model selection and workload shape, not a hidden gateway multiplier.
These figures are intentionally simple. Real billed usage can be higher because prompts contain conversation history, retrieved documents, tool definitions and retries. Forecast from measured pilot data once you have one representative week of logs.
How should a team build its first budget?
Start with the workload, not the model. Estimate calls per month, average prompt tokens, average output tokens, retry rate and media volume. Multiply by the published unit rates, add a 20% learning allowance for evaluation and failed attempts, and run the total through both a cheap and premium model.
Then compare the forecast with a monthly allowance. If the allowance is smaller than the forecast, model the overage. If it is much larger, leave the unused credit out of the comparison and ask whether the plan still provides value through support, concurrency or other documented benefits.
Models that change the cost equation most
| Model | Vendor | Price | Vs. official |
|---|---|---|---|
| gemini-3-flash-preview | $0.0368 / 1M tokens | -92% | |
| deepseek-v4 | DeepSeek | $0.0613 / 1M tokens | -90% |
| gpt-5.6-sol | OpenAI | $0.1838 / 1M tokens | -96% |
| claude-opus-5 | Anthropic | $0.4412 / 1M tokens | -91% |
Prices are metered in USD; the full list lives in the model catalogue.
Key takeaways
- Separate model usage from gateway/platform fees before comparing providers.
- Estimate text workloads with input, output and retry volume, not a single monthly token guess.
- Per-request media products need quality, resolution, duration and retry assumptions.
- Pay as you go fits spiky demand; a plan should beat the forecast on total value, not just headline allowance.
- Use the calculator for a first estimate and the console logs for the real monthly forecast.
Frequently asked questions
Are AI API gateway prices per token?
It depends on the model and provider. Text, embedding and rerank models commonly use token units; image, video and music models often use per-request units; speech may use characters or duration. Always read the specific unit beside the model price.
Does OpenLux add a hidden markup?
OpenLux publishes the catalogue rate applied to metered usage. Read the current catalogue and pricing page for the exact unit and any plan terms. The console is the authoritative usage record, and enterprise terms should be confirmed in writing.
Should I choose a plan or pay as you go?
Use pay as you go when usage is uncertain or spiky. Evaluate a plan when the baseline forecast is predictable and its allowance, overage, cancellation and support terms produce a better total outcome.
Why is my invoice different from the calculator?
The calculator uses a single published unit and excludes taxes, promotions, plan credits, retries, input/output differences and any other model-specific billing detail. Reconcile the estimate with console usage and the billing statement.
How much should I budget for evaluation traffic?
A practical starting allowance is 10% to 30% above forecast production usage, depending on how often your team tests prompts, models and failure cases. Measure the first month rather than leaving the allowance permanently vague.
Is a cheaper token model always the cheapest option?
No. A cheaper model may need longer prompts, more retries or stronger escalation to a premium model. Compare cost per successful task, not cost per token alone.
Sources and official documentation
- OpenLux pricingCurrent plans and billing information.
- OpenLux model cataloguePublished USD rate and unit for every model.
- OpenRouter pricingOfficial pricing and plan information for comparison.
- Portkey pricingOfficial gateway plan information.
- LiteLLMSelf-hosted gateway and pricing considerations.
Keep reading
AI API gateway guide
Choose the right operating model before comparing prices.
OpenLux plans
See the current subscription and usage options.
Full model catalogue
Compare every published model rate.
Enterprise procurement
Contract, invoice and support questions.
Gateway security
The controls that belong beside the price comparison.
OpenAI-compatible API
Avoid migration rework when switching gateways.
Turn your forecast into a real API test
Run representative traffic through OpenLux, compare the console estimate with the invoice and then choose pay as you go or a plan.