Teradata rebuilds Tera around agents and claims lower token costs
Teradata’s own benchmarks put Tera well ahead of Claude Code on cost. The new components arrive in Q4 2026, giving partners time to test the claims
- Announced: 22 September 2026, San Diego
- Availability: Q4 2026
- Customers choose their models and where workloads run, across cloud, on-premises and sovereign environments, Teradata says
- Teradata says Tera supported 512 concurrent agents on a single virtual machine with eight virtual processors
Existing Teradata customers have reason to trial Tera once the components ship, but the cost claims are vendor benchmarks and should be treated as a hypothesis. Ask for the published method, then rerun it on a representative pipeline with the model already in use before any saving goes into a budget. The vendor-neutral context layer is the part to test hardest: if it really works across other platforms’ data without moving it, it could matter more than the token savings.
“Most enterprises are not starting from scratch with AI. They are dealing with tools that do not work together and a skills gap that makes those tools hard to use at scale.”Sumeet Arora, chief product officer, Teradata
Teradata said on 22 September that it is rebuilding Tera, its AI tool for data teams, into what it calls an agentic coworker for data engineering, analysis and data science. Three new components – Tera Context Engine, Tera Harness and Agent Skills – are due in Q4 2026.
Teradata describes Context Engine as a vendor-neutral layer that links metadata, lineage and business definitions across databases, pipelines, catalogs and models, including data Teradata does not manage, without moving it. Harness routes each step of a task to a skill, tool, dataset or model and cuts model calls that don’t improve the result. Teradata says it starts agents from 84 pre-built execution patterns rather than having them reason from scratch.
The cost argument rests on Teradata’s own benchmarks. It says that on SWE-bench Pro, using the same Opus 5 model, Tera consumed 73% fewer tokens than Claude Code, finished 42% faster and cost 58% less. On data-eng-bench, a pipeline benchmark built by Snowflake Labs and Bespoke Labs, it claims 53% lower cost per reliably solved task than Snowflake Cortex Code running Opus 5, based on the benchmark’s published results. Teradata says it has published its method and materials for independent checking.
Sources
- Teradata, “Teradata Transforms Tera into an Agentic Coworker; Introduces Tera Context Engine and Tera Harness”, press release, 22 September 2026. https://www.teradata.com/press-releases/2026/tera-an-agentic-coworker
