Recombee vs Amazon Personalize: Hosted Recommender Wars

TL;DR.

• Recombee vs Amazon Personalize is rarely a pure feature fight. It is a procurement question: how much AWS lock-in are you willing to absorb to skip a separate vendor review?
• Recombee charges flat monthly tiers (free up to ~100k interactions/mo; $99 to $2,499/mo above that), and exposes ReQL for in-request business rules.
• Amazon Personalize charges pay-as-you-go ($0.05/GB ingestion, $0.002 per 1k interactions trained on v2 recipes, $0.15 per 1k v2 real-time recommendations) with a recipe catalog you pick from.
• Recombee retrains incrementally in real time. Personalize v1 forced manual retraining; v2 recipes train continuously, but the recipe-per-use-case model still constrains how you mix signals.
• Pick Amazon Personalize if you are already deep in AWS and value IAM, VPC, and procurement parity. Pick Recombee if you want one neutral API, faster integration, and ReQL filters without writing a re-ranking service.

If you are evaluating Recombee vs Amazon Personalize, the honest answer is that both will ship recommendations into your product within a week. The interesting question is which one you will hate in twelve months. We have integrated both for production traffic. This post walks through pricing, recipes vs ReQL, lock-in, and the failure modes that don't show up in the marketing pages. The framing applies whether you are also looking for a Recombee alternative or shopping for an Amazon Personalize alternative after a bill that surprised you.

Recombee vs Amazon Personalize: the two products in plain English

Amazon Personalize is an AWS service that takes your interaction data, trains one of about a dozen named "recipes" on it, and exposes a GetRecommendations API behind IAM. It is the productized descendant of the personalization stack Amazon uses internally. You upload datasets to S3, define schemas, train a solution, and deploy a campaign. Once live, you call it through the AWS SDK or the public API gateway.

Recombee is a Czech-built hosted recommender that takes your interaction data through a thin REST API, trains a hybrid model in the background (collaborative + content-based + deep learning), and exposes recommendation endpoints behind an API token. Their differentiator is ReQL, a query language that lets you push filters and boosters into the recommendation request itself.

Both are "hosted recommenders". Both promise that you do not have to run an inference cluster. The difference is whether you want a recipe catalog or a query language sitting between you and the model. If that distinction sounds abstract, it stops being abstract the first time a product manager asks for "trending items, but only from the user's preferred brand, boosted if they are on sale". On Recombee you write that as a ReQL clause. On Personalize you typically combine a recipe call with a filter expression and a re-ranking pass — or you train a second solution.

Pricing reality in Recombee vs Amazon Personalize: flat tiers vs pay-as-you-go

Recombee publishes a tier ladder. The free tier covers up to roughly 100,000 interactions per month, 100,000 recommendation requests, and 20,000 active users and 20,000 catalog items, per their pricing page. Standard is $99/mo for the same headline numbers with a $0.50 per-extra-unit overage. Plus is $899/mo for 2M interactions and 2M requests. Pro is $1,499/mo for 5M of each. Premium is $2,499/mo for 10M of each. Platinum is custom.

Amazon Personalize uses pay-as-you-go with no minimum fees. Per their pricing page, data ingestion is $0.05 per GB. For the v2 recipes (User-Personalization-v2 and Personalized-Ranking-v2), training costs $0.002 per 1,000 interactions ingested for training, and real-time recommendations cost $0.15 per 1,000 requests. The legacy v1 recipes use a different model: $0.24 per training hour, with real-time inference tiered from $0.0556 per 1,000 requests down to $0.0139 per 1,000 requests at scale. There is also a 1 TPS minimum charge on v2 inference even at zero traffic, which is not obvious until you read the small print.

The honest comparison:

• Small project (<100k interactions/mo, <100k requests/mo). Recombee is free. Personalize will cost you maybe a few dollars on v2 plus the 1 TPS minimum. Both are fine. Recombee wins on simplicity.
• Mid-sized SaaS (~2M interactions/mo, ~2M requests/mo). Recombee Plus at $899/mo flat. Personalize v2 at this volume runs roughly $4 for training + $300 for inference + ingestion = much cheaper than Plus on inference alone, but you will pay engineering time for IAM, schemas, and Avro.
• High-volume marketplace (~50M requests/mo). Personalize v2 inference is $0.15 per 1k regardless of scale; that is $7,500/mo on inference alone, plus training, plus ingestion. Recombee Premium at $2,499/mo (10M requests) is cheaper if you can fit; Platinum custom-pricing is the lever above that.

The bill that surprises people on Personalize is almost always the same: 1 TPS minimum on a campaign you forgot to delete, or a v1 solution that trained 200 hours over a debugging week. Recombee surprises are rarer because the tiers are headline numbers, but the overage rate climbs on Standard.

Recipes vs ReQL

Amazon Personalize ships about a dozen "recipes" — algorithms wired for specific use cases. The catalog, per the AWS recipe documentation:

• USER_PERSONALIZATION — User-Personalization-v2 (Transformer-based), User-Personalization (HRNN successor), and legacy HRNN variants.
• PERSONALIZED_RANKING — Personalized-Ranking-v2, Personalized-Ranking.
• POPULAR_ITEMS — Trending-Now (recent interaction velocity), Popularity-Count (lifetime popularity).
• RELATED_ITEMS — Similar-Items (uses item metadata), SIMS (collaborative-only), Semantic-Similarity.
• PERSONALIZED_ACTIONS — Next-Best-Action.
• USER_SEGMENTATION — Item-Affinity, Item-Attribute-Affinity.

If you want "users who liked X also liked", that is SIMS or Similar-Items. If you want a homepage feed, that is User-Personalization-v2. If you want a sort key for a search-results page, that is Personalized-Ranking. Each recipe is a separate solution, with its own training run, its own campaign, its own bill line item. Filters live in a separate filter-expression DSL evaluated post-recommendation. SIMS in particular only uses your Interaction dataset for similarity — not item metadata — which trips people up the first time they wonder why "color" or "brand" is being ignored, per the SIMS docs.

Recombee collapses most of that into one model surface and one query language. ReQL lets you write filters, boosters, and rerank logic inline:

'brand' == context("user_brand_preference")
boost: 'price' < 50 ? 0.3 : 0

You attach that to a recommendation call. The benefit is operational: you do not run a separate filter service or push business rules into your API gateway. The cost is that ReQL is a Recombee-specific language; you cannot port those rules to anything else without rewriting them. We have seen teams build hundreds of ReQL rules over a year. That is real lock-in, even though the engine itself is hosted neutrally.

If your product needs constantly evolving business rules and you want product managers to ship them, ReQL is the better surface. If your product needs a small number of well-defined recommendation contexts and you want each one served by a model purpose-built for it, recipes are the better surface.

Setup, SDKs, and developer experience

Recombee gives you a REST API and ten SDKs (per their comparison page), and you can have a working integration in an afternoon: create a database, send addItem and addInteraction calls, hit recommendItemsToUser. Schema is implicit; properties are key-value with optional types.

Amazon Personalize gives you the AWS SDK in every language AWS supports. The integration is heavier because of AWS conventions: you define a dataset group, define dataset schemas in Avro JSON, upload to S3, kick off an import job, wait for it to finish, create a solution, train a solution version (this can take hours), create a campaign, and then call GetRecommendations. None of those steps are hard. All of them are slow if you are not used to AWS plumbing.

A practical heuristic from our own integrations: a team comfortable in AWS hits production on Personalize in about a week. A team without AWS muscle memory hits production on Recombee in two or three days. If you are still picking your stack and weighing this against building from scratch, our five patterns for adding personalization covers the spectrum from "buy" to "build".

Real-time retraining vs scheduled retraining

Recombee positions itself as incrementally trained in real time, and in practice new interactions and new items are reflected in recommendations within seconds.

Amazon Personalize v1 required you to manually create a new solution version to incorporate new data (this is still true for SIMS, per the docs), and that is the historical complaint that the Recombee comparison page leans on. Personalize v2 recipes (User-Personalization-v2, Personalized-Ranking-v2) train continuously and use the Transformer-based architecture announced in 2024. They handle the "new items get cold-start treatment" case with a recency-aware sampling strategy.

If you are deciding Recombee vs Amazon Personalize today, the real-time gap has narrowed. The remaining difference is in cold-start: Recombee lets you push item metadata and use it from minute zero; Personalize v2 has improved cold-start but still benefits from a few hours of interaction data before it stops falling back to popularity. We covered the structural issues here in cold-start problem and day-zero personalization.

Lock-in cost: AWS lock-in vs Recombee lock-in

This is the part of Recombee vs Amazon Personalize that nobody runs through honestly.

Amazon Personalize lock-in. Your interaction data lives in S3 (portable), but your schemas, your recipe choices, your filter expressions, your campaign ARNs, and your IAM policies are all AWS-specific. Migrating off is a "change your data pipeline plus your inference path plus your auth model" project. The upside: if you are already on AWS, none of those are new things. You pay the cost you have already paid.

Recombee lock-in. Your interaction data lives in their Postgres-like store (exportable through their API), but every ReQL rule you have written is in Recombee's DSL. ReQL is more expressive than the Personalize filter language, which means a year of ReQL is a year of work to translate. The upside: you have a single neutral vendor with a small API surface; nothing in your data plumbing is Recombee-specific outside the ingest call.

Both vendors are lock-in. The choice is which kind of lock-in you can live with. If you have a CTO mandate to be cloud-portable, Personalize is worse. If you have a security review that requires every vendor to support PrivateLink and SCIM, Recombee is the harder sell. We discuss this trade in more depth in personalization platforms in 2026.

When each one wins in Recombee vs Amazon Personalize

We try to reduce the Recombee vs Amazon Personalize decision to four questions.

1. Are you on AWS already? If yes, Personalize is a one-click procurement and a familiar bill line. If no, Recombee is faster to ship and easier to budget.
2. How many recommendation contexts do you serve? One or two? Either works. Five or more, each with custom filters? Recombee's ReQL pays off because you do not run a separate filter service. Personalize requires either a filter expression per context or a re-ranker.
3. What does your team look like? ML engineers love Personalize because they can tune recipes, override hyperparameters, and run HPO. Generalist engineers prefer Recombee because the API surface is small.
4. What is the worst case? If the worst case is "we get acquired by an AWS-native company", Personalize. If the worst case is "we get acquired by a GCP-native company", Recombee.

Neither product is right for everyone. They are both correct, opinionated answers to the same question. The pitch you should reject is the one that says either of them is "the only managed recommender that scales".

For the structural decision of what a recommender even is, we wrote recommendation engine vs personalization layer and it covers why a recommender alone is rarely the right top-level abstraction.

How ×marble fits in

×marble is a knowledge-graph personalization product. Hosted recommenders like Recombee and Amazon Personalize are great at "given user u and context c, return items ranked by a learned interaction model". That is a real, hard problem and they solve it well. They are weaker at "explain why this item was ranked here", "compose recommendations across products" (Vivo for daily briefings, Music for tracks, Video for YouTube), and "share a single entity graph across surfaces". That is the gap ×marble was built to fill. If you are running Vivo for daily AI briefings or marblexmusic for music personalization, the same graph backs both. You can pair ×marble with a hosted recommender (we have customers running Personalize for the ranking model and ×marble for the explanation and cross-product graph), or use it as the spine of your personalization layer instead of either.

FAQ

Is Recombee a good alternative to Amazon Personalize?

Recombee is a good alternative if you want faster integration, flat-fee pricing, and a single API surface with ReQL for business rules. The trade-off is vendor diversity: you are leaving an AWS service for a smaller, focused company. For teams not already deep in AWS, Recombee is usually the easier integration. For teams who care about procurement parity with their cloud provider, Personalize wins.

What is the difference between Recombee vs Amazon Personalize?

Recombee is a hosted recommender with a REST API, ReQL query language for inline business rules, and flat monthly tiers from free to $2,499/mo. Amazon Personalize is an AWS service with about a dozen named "recipes" (User-Personalization, SIMS, Trending-Now, etc.), AWS-native auth and bill, and pay-as-you-go pricing per GB ingested and per 1,000 requests. Recombee retrains incrementally in real time; Personalize v2 recipes also retrain continuously but the recipe-per-use-case model is structurally different from a single ReQL surface.

Which Amazon Personalize recipe should I use?

For a homepage feed of personalized items, use User-Personalization-v2. For a "users who liked X" related-items widget, use Similar-Items if you have item metadata or SIMS if you only have interactions. For a "trending" feed driven by recent interaction velocity, use Trending-Now. For ranking a curated list of items per user, use Personalized-Ranking-v2. AWS recommends starting with the v2 variants whenever your data fits; v1 recipes are now mainly for cases v2 does not cover yet, like SIMS-style only-interactions similarity.

How much does Amazon Personalize cost per month?

Amazon Personalize charges $0.05 per GB of data ingested, $0.002 per 1,000 interactions for v2 recipe training, and $0.15 per 1,000 v2 real-time recommendation requests. A small workload (1 GB/mo, 100k interactions, 100k requests) costs single-digit dollars per month plus the 1 TPS minimum charge on active campaigns. A 2M-request workload runs around $300/mo on inference plus training and ingestion. There is a free tier covering 20 GB, 5M v2 training interactions, and 50,000 v2 requests for the first two months.

When should I not use a hosted recommendation engine?

Hosted recommenders are a poor fit when your business rules are the actual product (Spotify-style discovery), when your model needs to share state with non-recommendation models (a search relevance model that also learns from clicks), or when you cannot send user data to a third party for compliance reasons. They are also overkill when your catalog is small enough that hand-tuned rules outperform a learned model. For everyone else — most product teams — they are the correct first choice. We covered the broader stack picture in the marketing engineer's personalization stack.

Further reading

• The marketing engineer's personalization stack — where hosted recommenders sit in the broader stack.
• Recommendation engine vs personalization layer — why a recommender is rarely the whole answer.
• Personalization platforms in 2026 — the wider buy-vs-build landscape.
• Cold-start problem and day-zero personalization — how both vendors handle new items and users.
• Amazon Personalize pricing — primary source for AWS cost components.
• Recombee pricing — primary source for tier ladder and overage rates.
• Amazon Personalize recipe catalog — full recipe list with use-case mapping.