Recommendation API cost is the recurring expense of generating or ranking suggested items through a hosted service. Total cost includes any billed inference, training, data processing, infrastructure, and operational work.
Retrieval selects candidate items from a catalog. Ranking orders eligible candidates. Inference is the production calculation returning recommendations. An interaction is a recorded event, such as an impression, click, or purchase. Billing definitions vary by contract.
TL;DR: What should you budget?
- At 20 million monthly requests, every $0.05 per 1,000 requests adds $1,000 in monthly inference charges.
- At illustrative rates of $0.0556, $0.15, and $0.60 per thousand, monthly inference equals $1,112, $3,000, and $12,000 respectively.
- Training, interaction ingestion, minimum throughput, retrieval, and engineering ownership can change total costs substantially.
- A ranking-only API is not equivalent to a complete recommendation engine. Compare identical workloads and responsibilities.
- Evaluate cost per successful response and measured business outcome, not advertised request pricing alone.
How much do 20 million recommendation requests cost?
The answer is a formula, not a universal vendor price: multiply billable requests by the contract's unit rate, then add fixed fees and separately billed components. Publicly documented inference tariffs use different rates for different model classes and plans, so one figure cannot represent the entire market.
At 20 million requests, there are 20,000 units of one thousand calls. The table below is an arithmetic sensitivity analysis, not a list of vendor quotes or verified industry averages.
| Assumed price per 1,000 calls | Monthly inference cost |
|---|---|
| $0.025 | $500 |
| $0.05 | $1,000 |
| $0.10 | $2,000 |
| $0.15 | $3,000 |
| $0.30 | $6,000 |
| $0.60 | $12,000 |
For example, 20,000,000 / 1,000 x $0.15 = $3,000. If a plan includes one million calls, subtract that allowance before applying the relevant tier. If it has a monthly platform minimum, add or compare that commitment as the contract specifies. Never add both without checking whether usage is already included.
What does a recommendation API actually bill for?
A request may return ten items or one hundred items, yet some contracts bill per call. Others charge for recommendations generated, active users, item counts, or guaranteed throughput. Consequently, twenty million feed refreshes do not always equal twenty million billable transactions.
Inference: Real-time personalized responses, candidate scoring, or reranking. Separate billing can apply to batch jobs and online endpoints.
Training: Model creation and retraining, sometimes measured by computation hours or interactions processed. A rules-based ranker might have no training charge, while an actively trained recommender has recurring model expenses.
Data: Behavioral event ingestion, catalog updates, retained records, storage, and network transfer. A single feed request can produce multiple impression events plus later clicks.
Capacity: Reserved throughput, service minimums, regional deployment, and dedicated infrastructure. Billing floors can matter when traffic drops overnight.
Operations: Monitoring, fallback responses, integration maintenance, incident handling, and experimentation remain real costs even when absent from a vendor invoice.
A buyer should require a written definition of each billable unit and whether failed requests, retries, test environments, and metadata responses count.
Why can actual billed requests exceed 20 million?
Consider a thirty-day month: twenty million requests average approximately 7.72 per second. Traffic, however, arrives unevenly. A platform that serves bursts of 100 requests per second may need higher reserved capacity despite its modest average.
Now assume 5% of requests are retried once after a timeout. If each retry is billable, traffic increases from 20 million to 21 million calls. At $0.15 per thousand, inference rises from $3,000 to $3,150.
Some managed deployments also bill a minimum transaction rate while an endpoint remains active. An active one-request-per-second floor corresponds to 2,592,000 monthly transactions over thirty days, although whether this creates incremental charges depends on hourly traffic and the vendor's precise billing rules.
The correct forecast uses hourly request logs, not monthly totals alone. Include seasonality, experiments, duplicate requests, failure retries, and multiple recommendation surfaces.
How do training and interaction events change the invoice?
Assume a hypothetical service charges $0.05 per gigabyte ingested, $0.002 per thousand training interactions, and $0.15 per thousand inference requests. These are illustrative unit assumptions, not a consolidated vendor quote.
With 200 GB uploaded, ten million interactions considered during each of eight training runs, and twenty million inference requests, the arithmetic is:
- Data ingestion: 200 x $0.05 = $10.
- Model training: 80,000 units x $0.002 = $160.
- Inference: 20,000 units x $0.15 = $3,000.
- Total modeled service charges: $3,170 monthly.
This excludes retrieval, application hosting, support, storage beyond included allowances, and taxes. If the plan prices training by compute hours instead, the same calculation cannot be reused. Confirm whether historical events are charged again when retraining or only upon first ingestion.
Is ranking-only infrastructure cheaper than a complete recommender?
It can be cheaper for a platform that already retrieves suitable candidates, but it is not inherently cheaper for every application. Full recommendation systems often handle candidate discovery, feature processing, model training, ranking, and serving. A ranking-only API assumes upstream retrieval exists and bills for ordering a supplied shortlist.
This architectural separation matters because you should not purchase duplicate functionality. The Gortex recommendation engine architecture guide explains how retrieval, ranking, and serving work together.
For a social feed, existing follow-graph retrieval might already provide 200 eligible posts. The remaining problem is ordering those posts for a recipient. Conversely, a catalog application that needs to discover relevant products among millions still requires candidate generation. An ordering endpoint alone cannot solve that discovery problem.
Where does Gortex fit in this cost calculation?
Gortex's ranking API receives a recipient, context, and previously retrieved candidates through POST /v1/decide. It returns an ordered result, supported sponsored placements, and decision identifiers. Its current scope includes calibrated linear ranking and deterministic sponsored-slot filling.
Gortex is in private beta, and a public 20-million-request tariff is not verified. Therefore, assigning it a made-up dollar amount would be misleading. Its current ranking approach should not be presented as a trained personalized recommendation model, nor should its sponsored placement be confused with a live auction or campaign-management platform.
The relevant cost question is whether one decisioning integration can replace separate application-side ordering and sponsored-placement logic. The ranking API buyer's guide outlines latency, reliability, integration boundaries, and fallback considerations.
How do you compare API costs fairly?
Use an identical test contract. Specify monthly call volume, peak requests per second, candidates per call, returned items, active users, events ingested, retraining cadence, and required regions. Compare the same features and success criteria rather than treating nominal request rates as interchangeable.
Do not substitute monthly active users for requests without a measured calls-per-user ratio. For instance, 500,000 active users generating forty decisions each produce 20 million requests. A client that refreshes recommendations on every scroll may produce substantially more. Instrument distinct surfaces separately so product teams can see whether homepage carousels, search reranking, and notification feeds are responsible for additional billable traffic.
Then calculate three measures: monthly total cost, cost per thousand usable responses, and cost per meaningful outcome. A response is usable only if it satisfies availability, eligibility, freshness, and latency requirements.
For example, suppose a hypothetical service costs $4,000 monthly and produces usable responses for 95% of twenty million calls. Cost per thousand usable responses is $4,000 / 19,000,000 x 1,000, or approximately $0.211. The 95% figure is an assumption, not a reported reliability rate.
Evaluate outcomes through controlled online tests. Research by Hidasi and Czapp on offline recommender evaluation warns that flawed historical tests can mislead comparisons. Offline ranking scores alone do not establish incremental conversions or retention.
What hidden costs should engineering teams investigate?
Feature storage, network egress, event pipelines, and search infrastructure can remain outside the recommendation invoice. Engineering effort also includes schema integration, privacy reviews, observability, incident response, and contract migration.
Ask how the system behaves when a dependency is slow. A cached recommendation or deterministic eligible fallback may preserve the experience without requiring a second paid call. Gortex's feed ranking systems guide explains why batching, caching, and graceful degradation belong in the serving design.
If a feed includes sponsored placements, include monetization-related implementation costs separately. The Gortex marketplace advertising guide explains placement rules and measurement without equating advertising revenue to recommendation quality.
What should a buyer choose at 20 million requests?
Choose the smallest architecture that satisfies actual product requirements and constraints. If discovery and retrieval are already reliable, assess ranking-only decisioning. If candidates must be found and personalized across a large catalog, budget for retrieval and training as well.
Request itemized quotes under three traffic scenarios, then measure relevance, p99 latency, failures, and total ownership cost. The defensible answer is not a single market-average price: at twenty million calls, every $0.10 per thousand changes inference spending by $2,000 monthly, before the rest of the system is considered.
Top comments (0)