Teradata is adding a context engine, Tera Harness execution layer and reusable agent skills to Tera, with general availability planned by December.
The new stack aims to make multistep agent workflows cheaper and more predictable by planning execution before LLM calls, batching independent tasks and cutting loops or tool calls that do not advance work.
In Teradata’s SWE-bench Pro tests, Tera used 73% fewer tokens than Claude Code, finished tasks 42% faster and posted 58% lower total cost while using the same Opus 5 model.
Analysts said that could help CIOs budget agentic workloads more reliably, but they warned aggressive pruning can miss necessary steps and that benchmark savings may not match full enterprise ownership costs.
The launch extends Tera, introduced in May under Teradata’s Autonomous Knowledge Platform, and is expected to appeal first to existing Teradata customers rather than rivals’ installed bases.
Will Teradata's promise to slash agentic AI costs by 58% translate to real-world savings, or just shift expenses to heavy orchestration overhead?
By aggressively cutting AI reasoning steps to save money, is Teradata risking the very emergent intelligence that enterprises are actually paying for?
Can a vendor-neutral context engine truly govern enterprise AI without moving data, or will Snowflake and Databricks loyalists simply refuse to switch?