GPT-6.1 Terra shows higher efficiency than Claude Opus 5.5 in tests

Testing the new frontier

OpenAI recently released GPT-6.1 Terra, a model designed to handle complex reasoning and multi-step tasks. To see how it stacks up against the competition, independent testing compared it against Anthropic's Claude Opus 5.5. The evaluation focused on tasks requiring high effort, such as using a browser to fix software bugs and tracking biological cell divisions.

The tests were designed to be difficult enough to differentiate the models. Because both models performed similarly on easier tasks, the researcher increased the difficulty of the prompts until distinct differences emerged. These final tests included a release triage, a cell lineage tracking task, and a large-scale CI history audit.

As thenewstack.io reported, the results were gathered by running each model through its respective API at the maximum effort setting. This means the models were given the highest possible reasoning budget allowed by the providers. The goal was to see which model could navigate complex instructions with the best balance of accuracy and resource usage.

Efficiency in browser tasks

One of the primary tests involved a fake GitHub-style site where the models had to fix a failed software release. The task required the models to navigate a browser interface, clicking through pages and typing URLs to identify six different root causes. Both models successfully identified the issues, but they took very different paths to reach the solution.

Opus 5.5 took a more exhaustive approach, typing URLs 102 times to move between pages. In contrast, Terra was more efficient, only typing URLs 23 times and primarily using clicks to navigate the site. This difference in behavior resulted in Terra using about 17.1% of the estimated uncached token cost compared to Opus 5.5 for the same task.

The data shows that Terra produced about five times fewer output tokens than Opus 5.5 during this specific browser navigation. While both models reached the same score on the 50-point scale, Terra demonstrated a much leaner way of interacting with web elements. This suggests a potential advantage for developers building autonomous agents that need to navigate websites frequently.

Scientific coding and speed

The cell lineage test provided a closer look at scientific coding capabilities. In this scenario, the models were given a noisy 72-frame microscope movie of dividing cells. They had to write and execute Python code to track every division and trace each cell back to its parent. This type of task requires both spatial reasoning and precise code generation.

Terra outperformed Opus 5.5 in this category, scoring 48 out of 50 points. It correctly matched all 20 divisions and answered all summary questions accurately. Opus 5.5 scored 43.5 out of 50, missing two divisions and reporting one that never actually happened. Additionally, Opus 5.5 used significantly more turns and code executions to reach its result.

Speed was another major differentiator in this test. Terra finished the cell lineage task in about 15 minutes, while Opus 5.5 took approximately 36 minutes. This indicates that Terra is not only more accurate in this specific scientific context but also significantly faster at executing the necessary code to reach a conclusion.

Handling large datasets

The final test involved a CI history audit, which required the models to analyze 1,200 rows of test results. The data was intentionally messy, featuring three different time zones, various duration formats, and duplicate rows. The models had to answer 50 specific questions about the data without using any external tools.

Terra successfully completed the task with a score of 47 out of 50. It missed only a few status counts and one question it should have marked as unanswerable. It managed to process the entire dataset and provide answers for every prompt within the 128,000-token output limit per response.

Opus 5.5, however, failed to produce a final answer before hitting its output limit. This means it could not complete the task at the maximum effort setting provided. While this doesn't mean Opus 5.5 is incapable of the task under different settings, it highlights a limitation in how it handles large-scale data analysis compared to Terra.

What the results mean

The overall findings suggest that GPT-6.1 Terra is a highly efficient model for complex, multi-step workflows. Across the three final tests, Terra's estimated uncached token cost was about 12.1% of the cost of Opus 5.5. It also completed the tasks significantly faster, showing a total time difference of about 25.7 minutes over the course of the three tests.

However, the researcher noted that Terra still struggles with certain types of numerical accuracy. In previous tests involving large datasets, Terra left some totals uncalculated and provided incorrect answers. Users should remain cautious when asking the model to perform precise arithmetic on large tables or complex database records.

For most browser-based and code-execution tasks, the data suggests Terra is a strong candidate for production. Developers might want to run trials at the max effort setting to see if the quality gains justify the cost. Comparing the specific accuracy, completion rates, and time requirements for each model remains the best way to choose the right tool for a specific job.