OpenAI released GPT-6.1 Terra on 27 October 2026, a mid-tier model that costs half as much as the one it replaces and is aimed at software that operates a computer the way an office worker does. The API list price is $1 for every million tokens read in and $5 for every million written out, with cached input at $0.05. OpenAI says that undercuts GPT-5.6 Terra and the bigger GPT-6.1 Sol by 50%.
The pitch is high-volume work. OpenAI’s announcement gives the example of reading incoming invoices and typing the details into accounting software, a job where a company runs thousands of repetitions and every cent per task adds up. Developers can call the model as gpt-6.1-terra. Subscribers on the Plus, Pro, Enterprise, and Edu plans get it in ChatGPT Work and Codex, where OpenAI says the same usage allowance now stretches over more tasks than on GPT-5.6 Terra, and well beyond what Sol or GPT-6 Astra allow.
The benchmark numbers are strong, and every one of them comes from OpenAI. On Agents’ Last Exam, a set of professional tasks such as converting building plans into a 3D model, Terra at medium reasoning effort scores 15 percentage points above its predecessor for about a third of the cost. OpenAI says it matches GPT-6 Astra, its most capable model, at a tenth of the price per token.
The comparison with a rival is also OpenAI’s own. The company claims that at max effort Terra is more than three times as efficient with tokens as Anthropic’s Claude Opus 5.5 and a quarter of the price per token, which it puts at roughly nine times lower cost per task. On a scientific-research test called Terminal-Bench-Science 0.1, OpenAI says Terra lands five points or fewer behind Opus 5.5 at under a fifth of the estimated cost. The announcement includes no independent results, and Anthropic is not quoted.
Accuracy is where the claimed gain is largest. On an internal OpenAI test built from prompts that tend to provoke mistakes, errors at low reasoning effort dropped from 53.6% on the older Terra to 16.7% on the new one. At maximum effort it dropped by 18 points to 4.3%. Even the improved low-effort figure works out to roughly one wrong answer in six on prompts designed to trip the model, so the setting you choose matters as much as the model.
Coding shows a similar jump. On Terminal-Bench 4.0, which has agents fix problems from the command line, Terra at maximum effort solves 50.3% of tasks against the 21.5% published for its predecessor. OpenAI itself says Sol and Astra remain the better choice for the most complex coding work, so Terra is a cheaper workhorse and not a replacement for the top tier.
On safety, OpenAI says Terra reaches the Critical level for cybersecurity in its Preparedness Framework and keeps the protections built for Astra. The company also reports that it is more likely than GPT-5.6 Terra to stay within scope and to admit limits and blockers, and that it resisted malicious instructions in prompt-injection tests. Those evaluations are OpenAI’s too, and the announcement does not spell out what the Critical rating requires in practice.
Sam Altman, OpenAI’s chief executive, wrote on X, as quoted in a Fox News live blog, that the company is especially excited for “this new era of computer use.” The price cut explains why: a model that clicks through forms and reads drawings is only a business when each run costs pennies.
For anyone paying Sol rates to push forms and invoices through an agent, the first test is cheap. Rerun a week of real jobs on Terra at medium effort, count the wrong entries, and compare that error rate with the 50% bill reduction.
Reported by OpenAI on 27 October 2026, with the Altman post quoted from a Fox News live blog of the same date.
