Guide

SAP-RPT-1 consultant: scoping predictions on SAP's relational foundation model

As of 2026-10-09

Edited by Cédric Mary, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

A SAP-RPT-1 consultant scopes and delivers predictions on business tables with SAP's relational foundation model: choosing the use case, building the context rows the model learns from at each call, deploying it in SAP AI Core, checking accuracy against the incumbent approach, and sizing the cost. The name in the search is already dated. SAP-RPT-1 has been generally available since 8 December 2025, but SAP has announced its retirement by 31 December 2026, and the current release is SAP-RPT-1.6, generally available since 7 September 2026. In the corpus generated on 8 October 2026, one of 1,698 live postings names SAP-RPT, for an engineering role on the model family itself.

What the model does, and why the consultant's job is the context

SAP-RPT predicts a value in a business table, a class such as a cost centre or a payment-delay bucket, or a number such as days late, without training a model for the use case. Each request carries context rows whose target is known and query rows whose target cell holds a [PREDICT] placeholder. The weights are fixed; the context rows are the training set for that call. This is in-context learning applied to rows and columns, and SAP's research basis, the ConTextTab paper of June 2025, adds language-model embeddings of column names and values and pretraining on real-world tables.

The consequence shapes the role. Change the context rows and the answer can change, so most accuracy problems are context problems: a missing class, stale rows, or a column that leaks the target, which produces near-perfect pilot accuracy that collapses in production. Up to ten target columns are predicted jointly, so correlated targets belong together and independent ones in separate calls.

Versions and status, by date

SAP-RPT-1 was announced at SAP TechEd in November 2025 and has been generally available in the generative AI hub since 8 December 2025, in a small and a large variant. SAP-RPT-1.5 followed on 27 July 2026 with explainability, and SAP-RPT-1.6 on 7 September 2026, adding a context mode parameter on the large variant. SAP-RPT-1-OSS on Hugging Face is a research and non-commercial release, not the commercial model. SAP's release notes mark the models as unavailable in sovereign cloud offerings.

For a client, the first deliverable is often a migration line: any pilot or proposal still pointing at the RPT-1 model names needs moving to 1.6 or 1.5, with accuracy re-validated after the switch.

Access, call pattern and limits

Evaluation can start for free in the RPT Playground, limited to 25 rows, 4 columns and 50 target classes; a public Playground API allows up to 1,000 requests per hour. Productive use goes through the generative AI hub in SAP AI Core, which requires the extended plan: a configuration with scenario foundation-models, executable aicore-sap and the model name, deployed in a resource group, then called on the deployment's predict endpoint with context rows placed before query rows.

For RPT-1.6, the standard model accepts up to 2,048 context rows and 100 columns, the large model 65,536 rows and 256 columns; both accept up to 512 query rows and 10 target columns per request. SAP recommends 500 to 2,000 context rows on the standard model and 4,000 to 36,000 on the large one, and deep context mode only on the large model when more than 8,000 context rows exist. The upgrade path is standard first, then large, then deep mode.

Tabular orchestration and the second model family

Tabular orchestration removes the custom pipeline around the model. At design time a data destination in HANA Data Lake Files, a tabular artefact and a scenario are registered; at runtime the caller sends only query rows, and the workflow selects context rows with a random or a heuristic strategy. As at 7 September 2026 it supported models up to RPT-1.5, with 1.6 planned for the end of that month.

Since SAP completed its acquisition of Prior Labs on 17 July 2026, the hub also hosts TabPFN-3.5 Plus, generally available in AI Core since 15 September 2026. The architecture question is no longer whether to use RPT but which tabular model for which table shape, benchmarked on the customer's data. SAP's claim of up to twice the prediction quality of narrow models and 3.5 times that of LLMs is a vendor claim from the TechEd 2025 announcement, not an independent benchmark.

Cost, governance and evaluation

RPT consumption is metered in GenAI tokens converted into capacity units, at per-model rates published in SAP Note 3437766; deep context mode costs more by design. Joule agents are metered differently, in AI Units per action, so a mixed estate keeps the two meters apart in its cost model. We hold no list price, so this guide gives no figure per prediction.

There is no training step to audit, so review moves to the context: version it, log which rows each production prediction used, and re-run known-answer rows from the context as queries whenever the business population shifts. A context row is as sensitive as a document chunk sent to an LLM, so data masking applies to tabular pipelines too.

Decision table: when SAP-RPT is the right tool

New or frequently changing classification or regression on business tables, no data-science team per use case — Route: SAP-RPT-1.6 in the generative AI hub · Check: extended plan, sovereign-cloud constraint · Risk: building new work on RPT-1.

Stable, high-volume scoring with a validated gradient-boosted model — Route: keep the incumbent · Check: cost per prediction · Risk: replacing a cheaper model for novelty.

Large reference data, several consumers — Route: tabular orchestration · Check: supported model version · Risk: assuming the plain endpoint retrieves its own history.

Free-text reasoning or dedicated time-series forecasting — Route: an LLM or a forecasting method · Risk: forcing a tabular model onto the wrong problem.

Accuracy insufficient on the standard model — Route: large model, then deep mode above 8,000 context rows · Check: latency and cost.

Who writes this guide

Cédric Mary, editor of Analytics Legends, is a freelance SAP AI & Analytics architect whose delivery focus is the data layer these services run on (Datasphere, Business Data Cloud, Databricks). This guide is written from that vantage point and from the corpus cards listed below, not from a SAP-RPT project of his own.

What we cannot assert

No source we hold publishes a price per prediction, a GenAI-token rate we could quote without SAP Note 3437766, or a day rate for the role. Quality figures for SAP-RPT are SAP's own claims and benchmarks. We hold no confirmation that tabular orchestration supports RPT-1.6 or TabPFN in production beyond SAP's plan as at September 2026.

Frequently asked

Is SAP-RPT-1 still worth learning?

The mechanism is, the version is not: SAP has announced RPT-1's retirement by 31 December 2026. Learn the call pattern on RPT-1.6, which shares it.

Do I need SAP AI Core to use SAP-RPT?

For productive use, yes, on the extended plan. The free Playground and its API are for evaluation.

Can the open-source SAP-RPT-1-OSS model go into production?

No. It is released for research and non-commercial use; productive scenarios use the commercial models in AI Core.

Engage the editor

This guide is written by Cédric Mary, SAP AI & Analytics Architect & Team Lead. 27 years of SAP — Datasphere, Business Data Cloud, SAC Planning, Joule. Freelance or permanent · Hybrid / remote · Anywhere in EMEA. Available from 19 October 2026.

Profile and contact

What this page is built on

SAP-RPT (RPT-1 → 1.5 → 1.6) — Using SAP's Tabular Foundation Model
SAP-RPT — How the Relational Pretrained Transformer Works (Architecture)
SAP AI Core
Analytics Legends live contract corpus, 8 October 2026