OpenAI releases GPT-6.1 Terra: half the price and nearly on par with its senior models

OpenAI has released GPT-6.1 Terra, a mid-tier model that costs half as much as its predecessor and closes in on the flagships. A million input tokens cost $1, and a million output tokens cost $5. GPT-5.6 Terra was priced at $2 and $12, and Claude Sonnet 5.5 at $2 and $10. According to Artificial Analysis, the average cost per task at maximum reasoning effort fell from $1.40 to $0.40, and the Intelligence Index rose by 6 points. The model is available to Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and Codex, as well as via API.

Token economics: where exactly the price was cut

OpenAI positions Terra as the closest rival to Anthropic's Sonnet line, at half the price. Output tokens dropped from $12 to $5, a 2.4x cut. For agentic workloads, this matters more than the input discount: agents write a lot, and output tokens make up most of the bill. Cached input costs $0.05 per million, a quarter of the previous rate. Long system prompts and repeated context barely affect the budget anymore.

Artificial Analysis calls the model a new benchmark for "intelligence per dollar." At low reasoning levels, the Intelligence Index gain ranges from 7 to 13 points, and Terra sits on the Pareto frontier there: price and quality can no longer both be improved at once. Terra's own index score is 48. GPT-6.1 Sol scores 52, and GPT-5.6 Terra scores 42.

Computer use and coding

The main bet is automated software control. On Agents' Last Exam, Terra matches GPT-6 Astra at one-tenth the token price. Compared with Claude Opus 5.5 at maximum effort, the model uses a third as many tokens at a quarter of the price, which works out to roughly a nine-fold difference in task cost. According to the company, Terra completes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol.

The jump in coding is just as sharp. On Terminal-Bench 4.0, the model solves 50.3% of tasks, against 21.5% for GPT-5.6 Terra. Artificial Analysis says the 20-point gain puts it in the same class as GPT-6.1 Sol. On Terminal-Bench-Science 0.1, Terra beats Sol at a lower cost and trails Claude Opus 5.5 by five points at less than a fifth of the cost. For the heaviest tasks, OpenAI still recommends GPT-6.1 Sol and GPT-6 Astra.

Release numbers

• API prices: $1 input, $5 output, $0.05 cached input per million tokens • Intelligence Index: Terra 48, Sol 52, GPT-5.6 Terra 42 • Cost per task at maximum reasoning: $0.40 vs $1.40 • Coding Agent Index v2.0: Terra 48, Sol 54, GPT-6 Luna 31 • AA-Omniscience: guessing when it doesn't know 38% vs 88%, accuracy 2 points higher • Terminal-Bench 4.0: 50.3% vs 21.5% • MMMU-Pro: above its predecessor at all reasoning levels, almost on par with Astra and Sol • OSWorld 2.1: at medium effort, beats GPT-5.6 Terra's best result at a fraction of the price • Safety: protection against indirect prompt injections 98.8% vs 96.2%

The drop in hallucinations stands out more than the other metrics. The model invents an answer less often when it doesn't know one, and accuracy still rises. For automating forms, reading technical drawings and preparing proposals, that is a practical argument.

Safety and who should adopt Terra

According to the system card, GPT-6.1 Terra meets the "Critical" cybersecurity threshold under the Preparedness Framework and inherits the safeguards of GPT-6 Astra. The model more often stays within set boundaries and tells the user about its limitations. Against direct attacks on the instruction hierarchy, Terra holds up about as well as Astra and Sol, which score 99.99%. Against indirect injections, where a malicious command is hidden in third-party content, protection rose to 98.8%. For agents that read emails, web pages and documents, this metric is critical.

The practical takeaway for teams: Terra works as a workhorse for high-volume agentic and office scenarios. Teams that previously deployed the senior models for speed and price should move their tests there. Sol and Astra are better kept for complex tasks. Before migrating, it makes sense to run your own tasks at low and medium reasoning levels: that is where, according to Artificial Analysis, the gain in price-to-quality ratio is greatest.