OpenAI and AWS report 82% cost reduction per task with GPT-5.6 Terra

OpenAI Developers announced a collaboration with AWS to optimise the GPT-5.6 model in its Terra release and the Kiro development environment, with the stated goal of an improved price-performance ratio. The headline figure is a reduction of 82% in cost per successfully completed task, measured on Terminal-Bench 2.1.

Terminal-Bench 2.1 assesses coding capabilities and terminal-environment interaction; the selected metric, cost per successful task, reflects economic efficiency rather than raw accuracy. The run was performed inside Kiro, an AWS-managed specification-driven development environment, implying that the structured workflow enables the model to operate with fewer iterations and fewer tokens.

The rollout of the Terra version marks a clear direction: not merely a larger model but a cost-oriented engineering model. OpenAI did not release absolute accuracy scores, latency numbers, or a comparison with the prior model version, so the figure relates solely to the price-performance ratio in this particular setup.

The announcement lacks critical details: whether Terra is already accessible via API, what its context-window limits are, and how it behaves outside Kiro. Without additional data, the 82% claim remains a manufacturer statement under controlled conditions rather than an independent metric suitable for production use.

In a separate response, a request was made to the model Grok for security red-team timing request, including prompt injection and Zero-Trust principles, but this request was external and not part of the official OpenAI or AWS announcement.