- In Teradata's own SWE-bench Pro test against Claude Code, both on Opus 5, its Tera agent platform used 73% fewer tokens, finished 42% faster and cost 58% less.
- Outside analysts quoted by InfoWorld say these vendor benchmark results are not total cost of ownership, and a step that Tera's Harness wrongly discards could give cheaper but weaker output.
- Before relying on the numbers, have engineers rerun the test on your own workflows, have finance track full cost per finished task including reruns and review time, and ask about switching costs.
What Teradata is adding to Tera
Teradata is extending Tera, the AI workspace it launched in May inside its Autonomous Knowledge Platform, with a Context Engine, an execution layer named Tera Harness, and reusable agent skills. The commercial pitch is cost containment. As enterprises run more agents, the platform should spend fewer calls on work that does not help, without losing the business context attached to each task. Harness drafts a plan before anything goes to a large language model. Teradata says it then groups tasks that do not depend on one another, discards model or tool calls that add nothing, and caps how many steps a workflow may run according to its progress. Teradata plans to make the context engine, execution layer and agent skills generally available by December.
The efficiency numbers, and what they are measured against
The headline figures come from Teradata's own tests on the SWE-bench Pro benchmark, where Tera was set against Anthropic's Claude Code, both using the Opus 5 model. Teradata reported 73% fewer tokens, 42% faster completion and 58% lower total cost, with higher task completion rates. Tokens and total cost are separate measures. All three figures come from benchmark tasks run by the vendor, not from an enterprise's own workflows.
Outside views on cost control and budgeting
The analysts InfoWorld quotes treat this as a budget question. They say the economics of agent execution could become increasingly important for CIOs scaling agents within allocated AI budgets. Ashish Chaturvedi of HFS Research says the problem enterprises are only now discovering is agents that keep looping without making progress, and that shaving calls from each workflow adds up at volume. Stephanie Walter of HyperFrame Research argues that cutting how much inference an agent needs may be a more practical lever than switching to a cheaper model, because it reduces the number of inferences rather than the price of each one.
Advait Patel, a senior site reliability engineer at Broadcom, connected this to forecasting. He said the call count for an identical task can run from five one day to fifty the next, and that planning up front should narrow that spread and simplify budgeting and scaling decisions. Teradata's published figures cover tokens, speed and total cost. As reported, none of them measures that day-to-day variability.
The caveats a budget-holder should weigh
Patel also named the downside of pruning. Harness decides which steps are dispensable, so a step wrongly discarded produces a cheaper but weaker output. Results, he said, need close verification.
Robert Kramer, managing partner at KramerERP, warned against treating the benchmark as a cost figure: "Teradata's benchmark results should not be treated as equivalent to enterprise total cost of ownership," he said. Kramer's yardstick is the all-in price of one finished business task, from model fees and data compute through to reruns and the reviewer time spent checking what the agent produced.
Walter raised portability. Leaning on Tera for context, execution and reusable skills could embed those functions in workflows and make a later move to another platform harder. She expects the first buyers to be existing Teradata customers with complex, governed data estates, for whom Tera is "a logical expansion." Organisations standardised on Snowflake or Databricks are a tougher sell, in her view, since agents alone are not reason enough to migrate.
What to ask your team
Ask engineering to reproduce the comparison on your own workflows and data volumes rather than relying on SWE-bench Pro tasks. Ask finance to track the cost of each completed business task, including retries and human review time, as Kramer describes, rather than treating per-token figures as the whole picture. Ask what the review process is when Harness prunes a step that turns out to have mattered, and who signs off on the result. Ask how much agent logic, skills and context configuration would need rebuilding if you later moved off Tera, which is the portability question Walter raised.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: InfoWorld.





