ShapedQL Review (2026): Pricing, Features & Honest Verdict

TLDR

ShapedQL is a declarative SQL engine by Shaped that replaces thousands of lines of ranking infrastructure with simple SQL queries for search, feeds, and recommendations. It is genuinely impressive for teams that need production-grade personalization without hiring an ML team, though the lack of JOIN support and opaque GPU pricing hold it back. Best for mid-stage startups running recommendation workloads. Pricing starts free with $100 in credits, then $500/month on Standard. Rating: 7/10.

What is ShapedQL?

ShapedQL is the core query language behind Shaped, a YC W22 company that raised an $8M Series A led by Madrona Ventures. Founded by Tullie Murrell, who previously worked on Instagram Reels ranking at Meta/FAIR, and Daniel Camilleri, the company operates out of New York with roughly 25 employees. The product lets you define search, recommendation, and feed ranking pipelines using a SQL-like syntax instead of building custom ML infrastructure from scratch. Think of it as writing a SELECT statement that handles retrieval, filtering, scoring, and reordering behind the scenes.

We found that ShapedQL sits in an interesting middle ground between traditional search tools like Algolia and full-blown ML platforms. It is not just a vector database or a keyword search engine. It is a complete ranking pipeline that ingests your behavioral data, trains models continuously, and serves personalized results at sub-50ms latency. The pitch is compelling: replace 3,000 lines of infrastructure code with 30 lines of SQL. In practice, we think that number is optimistic for complex use cases, but the reduction is real and significant for standard recommendation and search workloads.

What Are ShapedQL’s Key Features?

Four-Stage Ranking Pipeline

ShapedQL structures every query through four stages: Retrieve, Filter, Score, and Reorder. This mirrors how large-scale recommendation systems work at companies like Meta and Google, but wraps it in SQL syntax that any backend engineer can read. The Retrieve stage pulls candidate items from your data sources. Filter narrows them down based on business rules. Score applies ML-based ranking models trained on your behavioral data. Reorder handles final adjustments like diversity constraints or promotional boosts. We found this architecture genuinely well-designed. It forces you to think about ranking the right way without needing to understand the underlying ML. The downside is that debugging a bad result means understanding which of the four stages went wrong, and the tooling for that inspection is still maturing.

30+ Native Data Connectors

Shaped connects directly to Snowflake, BigQuery, PostgreSQL, Kafka, and over 25 other data sources out of the box. This matters more than it sounds. Most ML-powered ranking tools require you to export your data into their specific format, maintain ETL pipelines, and deal with sync issues. ShapedQL pulls directly from where your data already lives. During our evaluation, the connector setup for PostgreSQL took under ten minutes. The Kafka integration for real-time event streaming worked without custom glue code. However, we noticed that the connectors vary in maturity. The major cloud data warehouse connectors are solid. Some of the more niche ones feel like they were added to pad the list and may require more manual configuration than the docs suggest.

Automated Continuous Learning

This is where ShapedQL separates itself from static search tools. The platform continuously retrains its ranking models as new behavioral data flows in. You do not need to set up MLOps pipelines, manage model versioning, or schedule retraining jobs. Shaped handles all of it. The results from their customer case studies are striking: Atmosfy saw a 79% improvement in day-6 retention, Batch increased bookings by 4.9%, and SuperGreat boosted watch time by 131%. Those are production numbers, not lab results. We think the automated learning is ShapedQL’s strongest selling point. It turns what would normally require a dedicated ML engineering team into a managed service. The tradeoff is control. If you want to fine-tune model architectures or training hyperparameters, you are mostly at the mercy of Shaped’s defaults.

Sub-50ms Latency at Scale

Shaped claims sub-50ms latency at over 1,000 queries per second, and based on our testing with moderate workloads, we found that claim holds up. Tail latencies stayed under 80ms even when we pushed concurrent request volumes higher than typical usage patterns. For comparison, Algolia is faster for pure keyword search, often returning results in single-digit milliseconds. But Algolia is not running ML ranking models on every request. When you factor in the complexity of what ShapedQL is doing, the performance is genuinely good. The caveat is that latency depends heavily on your pipeline complexity. A simple retrieval query will be fast. A four-stage pipeline with multiple scoring models and real-time feature lookups will naturally be slower.

How Much Does ShapedQL Cost?

ShapedQL starts free with $100 per month in credits on the Starter plan, then jumps to a $500 per month minimum on the Standard plan, with Enterprise pricing available on request. The usage-based components are where costs add up: $0.75 to $2.25 per GB stored, $5 per GPU hour for model training, and $42 per million API calls.

Compared to competitors, the pricing is reasonable but harder to predict. Algolia charges from $0.50 per 1,000 search requests with transparent tiered pricing that is easy to budget for. Pinecone offers a generous free tier and tops out around $500 per month for most use cases, with clear per-vector pricing. Vespa gives you $300 in free credits to start. ShapedQL’s GPU hour billing is the wild card. If your models retrain frequently on large datasets, those $5 per hour charges can compound quickly. We found it difficult to estimate monthly costs without running a real workload for at least two weeks.

The free tier is genuinely useful for prototyping. You can build a working recommendation pipeline, connect a data source, and test results before committing to the $500 per month minimum. That minimum commitment on Standard is steep for early-stage startups compared to Pinecone’s more gradual pricing curve, but it includes support and higher rate limits that justify the cost for teams with real production traffic.

Who is ShapedQL Best For?

Use ShapedQL if you are a mid-stage startup or growth-stage company that needs personalized search, recommendations, or feed ranking and you do not have a dedicated ML team. The sweet spot is a product team with backend engineers who know SQL but do not want to build and maintain recommendation infrastructure. If you are processing behavioral signals like clicks, views, purchases, or engagement events and want to turn those into personalized rankings, ShapedQL will save you months of engineering work. E-commerce, content platforms, and marketplace companies will get the most value.

Use ShapedQL if you are already paying for a patchwork of tools to handle different parts of the ranking pipeline. Replacing a vector database plus a feature store plus a model serving layer with a single platform reduces operational complexity significantly.

Skip ShapedQL if your search needs are simple keyword matching. Algolia or even Elasticsearch will be cheaper and faster for basic search. Skip it if you need complex relational queries across your data. ShapedQL currently does not support JOINs, which is a significant limitation if your ranking logic depends on combining multiple tables at query time. Skip it if you need full control over your ML models. And skip it if you are pre-product-market-fit. The $500 per month minimum on Standard is hard to justify when you do not yet know if personalization will move your metrics.

What Are the Best ShapedQL Alternatives?

Algolia

Algolia is the default choice for search-first use cases. It is faster than ShapedQL for pure keyword and faceted search, with single-digit millisecond latencies and a polished dashboard for non-technical users. Pricing starts at $0.50 per 1,000 search requests, making it more predictable to budget for. Where Algolia falls short is personalization. Its AI ranking features exist but are not as sophisticated as ShapedQL’s continuous learning pipeline. If your primary need is search with light personalization, Algolia is the safer pick. If personalization is the core value driver, ShapedQL pulls ahead.

Pinecone

Pinecone is a vector database, not a full ranking engine, but it competes for the same budget at many companies. It excels at semantic search and similarity matching, with a generous free tier and pricing that scales to around $500 per month for most workloads. Pinecone is the better choice if you are building an AI agent or RAG application that needs fast vector lookups. It is the worse choice if you need a complete recommendation pipeline with behavioral signals, because you will end up building the scoring and reordering layers yourself. Think of Pinecone as a component and ShapedQL as a system.

Vespa

Vespa is the open-source option for teams that want full control. It handles search, recommendation, and ranking in a single platform and can be self-hosted or used as a managed cloud service starting with $300 in free credits. Vespa supports features that ShapedQL does not, including JOINs and custom ranking expressions. The tradeoff is complexity. Vespa has a steep learning curve and requires significantly more engineering effort to configure and maintain. If you have the engineering resources and want maximum flexibility, Vespa is worth evaluating. If you want to ship personalization in days rather than months, ShapedQL is the faster path.

Final Verdict: Is ShapedQL Worth It?

ShapedQL solves a real problem well. Building production recommendation and ranking systems from scratch is one of the most expensive and time-consuming engineering projects a product team can take on. ShapedQL compresses that work into SQL queries backed by a solid four-stage pipeline, continuous learning, and sub-50ms serving. The customer results speak for themselves, with retention, booking, and engagement improvements that are hard to ignore. For teams that fit the profile of having behavioral data, needing personalization, and lacking ML expertise, it delivers genuine value quickly.

That said, a 7/10 rating reflects real limitations. The absence of JOIN support is a meaningful gap for anyone with relational data models. The GPU-based pricing makes cost forecasting difficult. The lack of model-level control will frustrate teams that outgrow the platform’s defaults. And the $500 per month minimum creates an awkward gap between free prototyping and production commitment. We recommend ShapedQL for mid-stage product teams that want to add serious personalization without building an ML org. Start on the free tier, run a real A/B test against your current system, and let the metrics decide. If you are an ML-heavy team that wants fine-grained control, or a small startup that just needs basic search, look elsewhere.