On October 26, 2026, OpenAI released GPT-6.1 Terra — a fast mid-tier model designed for agentic scenarios, computer use, visual reasoning, and everyday coding. Standard API prices are half those of its predecessor GPT-5.6 Terra and its sibling GPT-6.1 Sol, while cached input costs twenty times less than full input. According to vendor tests, the model outperforms GPT-5.6 Terra on agentic benchmarks and significantly reduces the cost of a solved task; independent confirmation of these figures is not yet available. Terra is already available to developers via API and to subscribers of paid ChatGPT plans.

image
image

What happened

The GPT-6.1 Terra model appeared in the OpenAI API on October 26, 2026, under the identifier gpt-6.1-terra and was immediately included in ChatGPT Work and Codex on Plus, Pro, Enterprise, and Edu plans. Standard prices are half those of GPT-5.6 Terra and GPT-6.1 Sol: one dollar per million input tokens, five cents for cached input, and five dollars per million output. According to OpenAI, on the Agents' Last Exam benchmark, Terra outperforms GPT-5.6 Terra by 15 percentage points at a medium reasoning level while being approximately three times cheaper per solved task. On Terminal-Bench 4.0 at maximum reasoning level, Terra solves 50.3% of tasks compared to 21.5% for GPT-5.6 Terra. In OpenAI's internal test on provocative prompts, the frequency of factual errors decreased from 53.6% to 16.7% at a low reasoning level and to 4.3% at maximum. The pricing schedule, model identifier, and list of plans are verified against model documentation, while quality figures currently exist only in the vendor's presentation.

Context

Terra occupies the mid-tier segment in OpenAI's lineup: according to the company, on Agents' Last Exam it is comparable to the flagship GPT-6 Astra at a price ten times lower — an argument that in the industry is usually associated with flagship distillation and computation compression. The pricing structure matches the architecture of agentic pipelines: a typical agent performs dozens of steps, each time resending nearly unchanged context, so recurring input at a symbolic price rebuilds the budget for such systems entirely. The main metric becomes economics: since agentic models can in principle solve routine tasks, the choice between them is determined by the cost per solved task and stability, not by a difference of a few points in quality. Another part of the background is the saturation of agentic benchmarks: when results reach half of solved tasks on Terminal-Bench 4.0, further points are gained through reproducibility and cost, not abstract quality.

Why this matters for the industry

For the industry, the release changes the margin arithmetic: existing agentic products can move high-volume routine tasks — form filling, reading bills, working with internal application screens — to Terra without changing vendors and significantly reduce the cost of a single task. Pipelines where input is mostly cached notice the effect on the day of migration, as cached input practically removes the fee for recurring context. This makes mass automation of routine tasks profitable in standard operation, not just in premium pilots, and creates direct price pressure on comparable mid-tier models. If the price-to-quality ratio is confirmed by early adopters, waves of migration from flagship models to the mid-tier are likely, along with its gradual transformation into a commodity layer of agentic automation. Then competition will finally shift to cost per solved task, reliability, data, integrations, and orchestration quality, including fallback, human-in-the-loop, and agent action auditing.

Why this matters for users

Plus, Pro, Enterprise, and Edu subscribers get Terra in ChatGPT Work and Codex right now, and the same usage limit covers more tasks than before: typical agentic scenarios, application interface analysis, and everyday coding consume quota more slowly. The agent is less likely to give confidently wrong answers to tricky formulations, most noticeably at high reasoning levels, so the volume of manual result verification decreases. Developers can connect gpt-6.1-terra to their applications immediately and test the vendor's statistics with their own runs on real tasks — this is the fastest way to find out the real price and quality for their case. For those who do not need reasoning in the pipeline, the reasoning.effort mode with a value of none is available, opening up classic sampling parameters temperature and top_p.

What is still unknown / limitations

All quality claims rely exclusively on runs by OpenAI itself: neither Agents' Last Exam, nor Terminal-Bench 4.0, nor the internal test on provocative prompts have been published with methodology or replicated by third-party teams. The comparison with GPT-5.6 Terra on Terminal-Bench 4.0 was conducted at maximum reasoning level on a single vendor benchmark without a single harness, fixed scaffold, or metrics like pass@k, so the correct presentation is a "vendor estimate," not an established fact. The claimed increase on Agents' Last Exam, combined with the promise of significantly lower task cost, mixes a quality metric and a cost metric: the composition of "task cost," including the accounting of cached tokens and the number of agent steps, has not been disclosed. For the internal test on provocative prompts, neither the dataset nor the protocol has been published, so the reduction in the frequency of factual errors remains a signal of a positive trend in calibration, not a reproducible result. Comparability with the flagship GPT-6 Astra is claimed at a price an order of magnitude lower and is verified only by an independent run of both models on identical infrastructure, and at the time of publication, such runs do not exist.

Sources

Author

Look at AI, editorial team