GPT-6.1 Terra vs. GPT-6 Astra: One sacrifices accuracy for lower cost and higher speed at high effort

On October 26, OpenAI unveiled GPT-6.1 Terra, positioning it as a cost-effective, speedy model for routine business tasks that costs just a tenth of its flagship GPT-6 Astra. Listed as a mid-tier option in OpenAI's GPT-6 lineup, Terra commands $1 per million input tokens and $5 per million output tokens, representing just a tenth of Astra's $10 and $50 rates.

According to the company, Terra completes tasks more than three times faster than Astra in the Agents’ Last Exam benchmark. However, early benchmarks revealed both advantages and limitations. Plotly noted Terra correctly identified three out of five hidden patterns in its analytics test and successfully answered 40 of 43 questions.

Artificial Analysis disclosed that Terra only provided an answer to 38% of questions it couldn't solve, down from 88% for GPT-5.6 Terra in the same benchmark. In an accounting benchmark, Terra scored 78.2% accuracy. Amid the ongoing AI competition, each new release must be positioned as the best to avoid being overlooked. This can result in a confusing experience for users when reality doesn't align with the hype.

To explore the trade-offs between Terra and Astra, I conducted three synthetic tests comparing their speed, estimated token costs, and accuracy. These results reflect single runs on synthetic tasks, not indicative of overall performance or repeatability. Despite setting the reasoning effort to high, I did not impose the same reasoning-token budget on both models and capped output tokens, as each model has built-in limits.

To ensure a comprehensive evaluation, I validated results against keys I created and verified after developing three versions of each test. These tests focus on real-world scenarios rather than purely computational benchmarks. Note that matching effort settings does not guarantee identical reasoning token usage. Both models delivered identical scores on easy and medium-difficulty tests.

To uncover potential weaknesses, I subjected the models to hard tests against a weaker model, GPT-5 Mini, until Mini's score significantly decreased. Here are the final test results. Terra's SQL reasoning test revealed a significant gap between the two models. Astra achieved 48/50, while Terra only earned 26/50. Each of the 25 questions was worth two points, totaling 50 points.

Terra missed only one question, which involved calculating the total runtime for 151 software builds in a single region. Astra's response was 406 seconds slower than the accurate total of 31,428 seconds, amounting to about 1.3% below the correct answer. While Astra demonstrated higher accuracy in this instance, its incorrect total still required verification.

Terra generated 27,970 reasoning tokens and took 319.7 seconds to respond. In contrast, Astra produced 28,806 output tokens and took 319.7 seconds to respond. In the data validation test, Astra achieved a perfect score of 50, while Terra scored 45. Terra flagged 40 records that violated rules, earning one point per record, and earned an additional 5 points for 10 records that appeared incorrect but complied with all rules.

Conversely, Astra correctly flagged all 40 erroneous records. However, Terra incorrectly identified 10 valid records as incorrect, resulting in a loss of 10 points. Terra warned that it may miss some rules when examining all 4,950 records and might not identify all errors.