Snowflake Opens Decision Model Preview for Bounded AI Tasks
Snowflake Decision adds typed classification, scoring and routing inside Cortex AI Functions, with private-preview access and only vendor-run benchmark evidence so far.
Edited by Tyronne Panaino
Snowflake introduced Snowflake Decision in private preview on October 8, adding a purpose-built model for bounded decisions to Cortex AI Functions. The change matters to data and operations teams that use model calls across large tables but need a category, score or probability rather than a free-form answer.
The model is available through Snowflake's AI_COMPLETE function for approved preview users. Snowflake says calls stay inside Cortex AI Functions, where the platform's existing governance, access-control and audit policies apply. That makes the release primarily a workflow and control-plane change, not a general replacement for a large language model.
What Snowflake Decision is designed to do
Snowflake frames the model around classification, scoring, filtering, routing and triage. These are tasks where the possible outputs can be defined before the call. A support operation, for example, can ask the model to choose a destination team, score the tone of a message and estimate whether it is urgent.
The documented interface supports three kinds of questions: choosing from a fixed list, assigning a score against an ordered rubric, and returning a probability for a yes-or-no statement. Results come back in a structured answer map. Choice and score outputs include a confidence value, and the response can include probabilities for the available options.
This contract is the main difference from asking a general-purpose model for prose and then parsing the result. A typed answer can feed a database or workflow directly, while a bounded option set gives an application a clearer rule for accepting, rejecting or escalating the output. Snowflake explicitly positions general-purpose models as the better choice when the answer space is open, the task needs creative generation or the request requires multi-step reasoning.
Where it fits in an enterprise pipeline
The initial use cases are repetitive decisions made across many rows: assigning incoming tickets, grouping customer reviews, checking whether assistant answers follow a policy and deciding which requests should be sent to a larger reasoning model. The final example makes Snowflake Decision a possible routing layer inside a multi-model pipeline.
For existing Snowflake users, the practical appeal is proximity to governed data and SQL-based processing. A team can define the questions in a call to AI_COMPLETE and keep the result alongside the source records. Snowflake's announcement says the same governance, access and audit rules already used with Cortex AI Functions apply to the decision call. The source does not independently test those controls or describe every preview limitation, so organizations still need to validate their own policy and data-handling requirements.
Snowflake's benchmark needs careful reading
Snowflake reports that Decision scored 57.63 and ranked first among 72 models in its evaluation. The company ran 29 benchmarks from Jev Decision Index 0.2.1, covering 88,837 test requests. Snowflake also says these 29 tests were only a subset of the broader benchmark because the current interface, including a maximum of 32 options, could not run the full set.
Those details are useful because they define the comparison instead of presenting the ranking without context. They do not turn the result into independent evidence. Snowflake selected and ran the evaluation, and this run did not fetch a third-party reproduction. The announcement also does not provide public preview pricing, latency distributions, model architecture details or a production reliability record.
Availability and next checkpoints
Snowflake Decision is not described as generally available. Customers must contact their Snowflake account team to request access, after which the company will provide preview documentation and next steps. That access path limits how broadly the release can be tested today.
The next verifiable checkpoints are wider availability, transparent pricing, published latency and throughput data, documentation of preview constraints and independent comparisons on the same bounded-decision tasks. It will also matter whether Snowflake exposes enough calibration information for teams to choose useful probability thresholds and monitor changes over time.
Status
Confirmed. Snowflake has opened the private preview and documented the interface, use cases and its own evaluation. Internal confidence is medium because the only fetched evidence is Snowflake's first-party article and the benchmark has not been independently reproduced here.
Sources
Update note: Last reviewed 2026-10-13. We will revise this post if Snowflake publishes general-availability terms, pricing, fuller evaluation data or material preview changes.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.