GPT-6.1 Terra Halves OpenAI API Prices

OpenAI launches GPT-6.1 Terra, an intermediate model sold at half the price of GPT-5.6 Terra via API. The company positions it for desktop tasks, software control, and coding, with fewer hallucinations according to cited tests.

OpenAI has unveiled GPT-6.1 Terra, an intermediate AI model that costs half as much as GPT-5.6 Terra on the API, according to the company. Available in ChatGPT Work, Codex, and via the API for Plus, Pro, Enterprise, and Edu plans, it targets office tasks and automated software control. The goal is clear: to offer better value for money without sacrificing performance, a challenge where costs quickly add up.

A model designed for everyday work, not just demonstrations

OpenAI describes GPT-6.1 Terra as a model that is both fast in visual reasoning and precise in computer control. It is intended to draft proposals, read technical diagrams, fill out forms, or manage actions in software, with writing considered more natural than its predecessor.

On pricing, the reduction is clear: 1 dollar per million input tokens and 5 dollars per million output tokens, compared to 2 and 12 dollars for GPT-5.6 Terra, and 2 and 10 dollars for Claude Sonnet 5.5 according to cited figures. Cached inputs drop to 0.05 dollars per million tokens, a quarter of the previous rate. The message to the market is simple: AI useful for daily tasks must also remain cost-effective.

Des gains mesurables sur le prix, le code et le contrôle machine

Artificial Analysis presents GPT-6.1 Terra as a new benchmark in intelligence per dollar. At maximum reasoning effort, the model gains 6 points on the Intelligence Index compared to GPT-5.6 Terra while costing less than a third per task, at 0.40 dollars versus 1.40 dollars. At lower effort, the gap widens further, with 7 to 13 points improvement according to the cited evaluation.

The model also improves in coding. On Terminal-Bench 4.0, it solves 50.3% of tasks at maximum effort, compared to 21.5% for GPT-5.6 Terra. In the OSWorld 2.1 office agent benchmark, OpenAI says it surpasses the best score of its predecessor at medium effort for a fraction of the cost. And on Agents' Last Exam, it reaches a level close to GPT-6 Astra for a tenth of the token price. Spreadsheets may suffer silently, but they remain the ultimate test for this kind of launch.

Moins d’hallucinations et un niveau de sécurité jugé critique

OpenAI also claims to have reduced hallucinations, those fabricated answers when the model does not know the response. On the private AA-Omniscience benchmark, GPT-6.1 Terra “guesses” in only 38% of cases, compared to 88% for GPT-5.6 Terra, with a two-point increase in accuracy. The improvement remains notable, though caution is advised [to be verified] outside the company’s and Artificial Analysis’s tests.

In security, GPT-6.1 Terra reaches the “Critical” threshold of OpenAI’s Preparedness Framework and inherits protections from GPT-6 Astra. According to the system card, it resists almost as well as GPT-6 Astra and GPT-6.1 Sol to direct attacks on instruction hierarchy, with 99.99% success on these tests. Against indirect prompt injection attacks, its defense rate rises to 98.8%, compared to 96.2% for GPT-5.6 Terra.

An intermediate positioning, but not the finish line

OpenAI places GPT-6.1 Terra between its most powerful models and its more expensive offerings. For the most demanding coding tasks, the company still recommends GPT-6.1 Sol and GPT-6 Astra. Terra remains a deliberate compromise: more affordable, faster, and sufficient for a wide range of professional uses.

This strategy says something about the current AI market: the battle is no longer only about raw scores, but about cost per task, execution speed, and operational reliability. It is often there that deployments are won, far from flashy but overpriced demonstrations.

By the numbers

  • 1 dollar / million de tokens — entry price of GPT-6.1 Terra, worldwide, OpenAI.
  • 5 dollars / million de tokens — output price of GPT-6.1 Terra, worldwide, OpenAI.
  • 50.3% — tasks solved on Terminal-Bench 4.0, at maximum effort.
  • 38% — guessing rate on AA-Omniscience, versus 88% for GPT-5.6 Terra.
  • 98.8% — defense rate against indirect injections, according to the system card.

To read