GPT-6.1 Terra Officially Released: Agentic Tasks on Par with Flagship Astra, 3x Faster at One-Tenth the Price

On October 26, 2026, OpenAI officially released GPT-6.1 Terra, with API pricing at $1 per million input tokens and $5 per million output tokens. According to an official AWS announcement, the model scores 15 percentage points higher than GPT-5.6 Terra on OpenAI's "Agents' Last Exam" agentic benchmark, on par with the flagship GPT-6 Astra, while completing tasks more than 3x faster than both GPT-6 Astra and GPT-6.1 Sol. The model became generally available on Amazon Bedrock the same day.

The specific numbers show this is a substantial update. According to BleepingComputer, at low reasoning effort, GPT-6.1 Terra's error rate dropped from 53.6% for GPT-5.6 Terra to 16.7%; on the coding benchmark Terminal-Bench 4.0, the solve rate at maximum reasoning effort rose from 21.5% to 50.3%, more than doubling. These figures are OpenAI's own reported numbers rather than independent third-party test results. Data from third-party evaluation platform BenchLM.ai as of October 26 shows GPT-6.1 Terra with a composite capability score of 75.9/100, ranking 10th among 229 indexed models, including 4th of 58 in multimodal capability (95th percentile, score 91.6), 8th of 133 in agentic capability (95th percentile, score 72.4), and 12th of 188 in knowledge capability (94th percentile, score 74.9)—though BenchLM.ai also explicitly notes that five categories—coding, reasoning, multilingual, instruction following, and math—remain "not yet measured."

A Middle Tier Forced to "Return"

Understanding this release requires reckoning with the naming confusion left behind by OpenAI's current generation of models. OpenAI's design logic runs from small to large: Luna (small) → Terra (medium) → Sol (large) → Astra (larger). Yet during the GPT-6.1 cycle, Terra was quietly discontinued, while Sol-6.1 was positioned at the former Terra-6.0 price point and in some scenarios already outperformed Astra, leaving the middle tier of the product line effectively vacant for a period—a brief market window in which "Sol-6.1 ≈ Astra performance at old Terra pricing" existed.

Terra now returns with version 6.1, filling that gap. In a Hacker News discussion, an engineer claiming to work at OpenAI (username tedsanders) explained that GPT-6.1 Terra outperforms Astra and Sol in technical communication, matches both in visual understanding and computer use, but is much faster; on "most other tasks" it falls slightly short. He advised developers to view the model as "a mid-range option between Sol (too expensive) and Luna (too dumb), especially for vision, computer use, and communication tasks." In the same thread, another user described the degree of naming confusion by asking whether "Microsoft's chief brand officer jumped ship to OpenAI"—the OpenAI engineer's reply acknowledged that "we don't always manage clarity and concision."

What the Cost-Performance Formula Actually Means

Terra-6.1's pricing is identical to its predecessor Terra-6.0, but performance has improved substantially. This is a classic price-anchoring maintenance strategy: significantly raising the performance ceiling of the old price band without disturbing users' price expectations. According to BleepingComputer, compared with GPT-6 Astra, Terra's per-token pricing is about one-tenth; according to the AWS announcement, compared with GPT-6.1 Sol, both input and output prices are more than 50% lower. Both sets of numbers point to the same conclusion: for agentic task scenarios, Terra is currently the option with the highest marginal cost-performance ratio in OpenAI's product line.

For agent systems, the weight of the speed factor is often underestimated. Batch tasks run by automated workflows are extremely sensitive to throughput—3x speed means 3x the task volume on the same budget, or latency compressed to one-third. The same-day general availability on AWS Bedrock further lowers enterprises' migration costs: teams already running services in the AWS ecosystem can access Terra without switching infrastructure, and Bedrock's security controls and access auditing directly cover enterprise compliance needs.

This is not the first time an OpenAI model has landed through Bedrock. Previously, both GPT-6 Astra and GPT-6.1 Sol were listed on Bedrock; Terra's addition reflects a systematic effort to fill gaps in the pricing ladder rather than a one-off collaboration. The distribution relationship between OpenAI and AWS is forming a stable channel model, which for enterprise users who rely on cloud infrastructure rather than calling the OpenAI API directly means that model selection is becoming more practically actionable.

A New Variable in the Competitive Landscape

At the broader market level, GPT-6.1 Terra's release continues an industry-wide trend: mid-sized models surpassing larger models on specific tasks is no longer an anomaly but is becoming the mainstream narrative. In the Hacker News discussion, some cited Sonnet 5.5 outperforming the larger Fable 5.1 in certain dimensions and Gemini 3.8 Flash surpassing Gemini 3.1 Pro in some scenarios as examples, showing that the phenomenon spans different vendors' product lines. OpenAI's repositioning of Terra is essentially an acknowledgment that in the specific application scenario of agent tasks, the competitive weight of speed and cost has surpassed parameter scale.

Terra's vision and agentic capabilities both sit at the 95th percentile on BenchLM.ai, and these two capabilities map precisely to the core use cases described in the AWS announcement: drafting proposals, interpreting technical drawings, codebase investigation, and agentic operations across business systems. This capability combination points to a specific market segment—enterprise-grade workflows that don't need the strongest model but require stable visual understanding and fast task execution. This overlaps heavily with the "engineers' daily workhorse" positioning long held by Anthropic's Sonnet series, and the direct pricing pressure is real.