Data notebook and app builder with an SQL and Python agent, YAML semantic models, and a YAML eval harness that grades the agent's answers.
Analyst and data-science teams who want agent-assisted work they can read as code rather than trust blind: the agent runs inside a notebook where every SQL cell, Python cell and output stays inspectable. It suits a team that will actually build an eval suite, because Hex ships one - YAML cases with an LLM judge plus a numeric_value rubric that compares the number in the answer against a literal target or a SQL result within a tolerance. Also a fit if you already maintain definitions in dbt MetricFlow, Cube or Snowflake Semantic Views, since Semantic Model Sync pulls those in rather than making you re-author them.
Hex publishes no accuracy figure for its agent - the evals documentation states the only numbers on the page are configuration limits (100 cases per suite, 15 rubrics per case, 3 concurrent runs per workspace), not performance. The eval harness that would give you a number is gated to Team at $75 per Editor per month and above, and is run from the Hex CLI, so a non-technical buyer cannot produce an accuracy baseline on Professional at $36. Costs stack in three places rather than one: per-Editor seats, AI credits that are granted monthly per seat and do not roll over, and compute billed per minute with published pay-as-you-go rates from $0.32/hr for Large to $2.58/hr for 4XL and $4.06/hr for an A10G GPU. Semantic models are optional, so nothing stops the agent writing free-form SQL against a connection with no governed layer behind it.
Pros
Cons
No. Nothing in the pricing pages or the evals documentation gives an accuracy percentage or benchmark score for the Hex agent. The evals page is explicit that its only numbers are configuration limits - 100 cases per suite, 15 rubrics per case, 3 concurrent runs per workspace - and illustrative example values. If you want a number, you have to build the eval suite and generate it yourself.
Two documented mechanisms. Semantic models, authored in YAML or synced from dbt MetricFlow, Cube or Snowflake Semantic Views, give the agent shared metric and dimension definitions. Separately, eval suites let you pin expected answers: the numeric_value rubric compares the number the agent returns against a literal target or a live SQL result within a tolerance you set, and suites can run on a schedule to catch regressions.
Three meters run at once. Editor seats are $36 per month on Professional and $75 on Team. AI features consume credits granted monthly per seat, which refresh monthly and do not roll over, with add-on credits purchasable. Compute is billed per minute: Medium is included on paid plans, while pay-as-you-go runs from Large at $0.32 per hour to 4XL at $2.58 and A10G GPU at $4.06.
Hex documents 16 connection types, including Snowflake, Databricks, BigQuery, Redshift, Athena, Postgres, Trino, Starburst and MotherDuck, so queries run against your own warehouse. For AI inference, Hex states it sends database metadata such as schemas plus project code and output content as model context, and that its LLM providers do not train on customer data.
Facts sourced from the vendor’s own site, last verified 2026-10-10.
The Weekly Source
One email a week. New listings, repriced tools, and the ones we would skip - including this one if its pricing changes.
