GPT-6 Astra scored a top-tier 67 points in AI coding performance, demonstrating outstanding token efficiency, but still fell short of Fable 5.1.



Artificial Analysis, which independently evaluates the performance of AI models, has released benchmark results for OpenAI's '

GPT-6 Astra .' In the ' Artificial Analysis Coding Agent Index ,' which measures performance as a coding agent, GPT-6 Astra scored 67 points, reaching a level almost on par with Claude Opus 5, Fable 5, and Muse Spark 1.3, but falling short of the top score of 70 points achieved by Fable 5.1.

Benchmarking GPT-6 Astra | Artificial Analysis
https://artificialanalysis.ai/articles/benchmarking-gpt-6-astra



On September 3, 2026, Artificial Analysis released the results for the Artificial Analysis Coding Agent Index, which assesses the GPT-6 Astra's capabilities as a coding agent, and the Artificial Analysis Intelligence Index, which assesses its broader cognitive abilities. The two indices revealed different trends for the GPT-6 Astra; while it showed high cost efficiency in coding, the high cost of the overall cognitive ability assessment worsened its cost efficiency.

In the Artificial Analysis Coding Agent Index, GPT-6 Astra, running on Codex, scored 67 points. This is the same score as Fable 5 (67 points), and is close to Claude Opus 5 (68 points) and Muse Spark 1.3 (64 points). Meanwhile, Fable 5.1, running on Claude Code, recorded 70 points, taking the top spot.



One of the most significant improvements in GPT-6 Astra is token efficiency. When using Codex at maximum inference strength, GPT-6 Astra consumed approximately one-third the number of tokens as GPT-5.6 Sol and approximately one-seventh as Claude Opus 5. Artificial Analysis estimates that GPT-6 Astra uses approximately 70% less tokens compared to GPT-5.6 Sol.



This high token efficiency is also reflected in the cost per task. At maximum inference, GPT-6 Astra had almost the same cost per task as GPT-5.6 Sol, yet it scored 2 points higher in the Coding Agent Index. Furthermore, while achieving the same score of 67 as Fable 5, its cost per task was less than half that of Fable 5. While Gemini 3.8 Flash is cheaper than GPT-6 Astra when focusing solely on token price, GPT-6 Astra is notable for its lower cost per task.



Therefore, GPT-6 Astra is positioned at the 'Pareto frontier' when comparing its Coding Agent Index score with its cost per task. In other words, GPT-6 Astra is a cost-efficient model, having the highest score among models with comparable costs and the lowest cost among models with comparable scores.

On the other hand, the Artificial Analysis Intelligence Index shows a slightly different evaluation of GPT-6 Astra. GPT-6 Astra scored 61 points, the same as GPT-5.6 Sol and Muse Spark 1.3, and 5 points lower than Fable 5.1, which scored 66 points.



The Intelligence Index also shows an improvement in the token efficiency of GPT-6 Astra, with the number of output tokens at maximum inference being approximately 10% less than that of GPT-5.6 Sol.



However, the API fees for GPT-6 Astra are $10 (approximately 1,500 yen) per 1 million input tokens and $50 (approximately 7,500 yen) per 1 million output tokens, which is 2.5 times higher than GPT-5.6 Sol's $4 (approximately 600 yen) for input and $20 (approximately 3,000 yen) for output. As a result, the cost per task at maximum inference is 75% higher than GPT-5.6 Sol.



In other assessments, the 'AA-Omniscience' scale, which evaluates model knowledge and hallucination, showed a decrease in hallucination rates at peak inference from 92% for GPT-5.6 Sol to 51% for GPT-6 Astra. Simultaneously, accuracy also increased by 4 points, and Artificial Analysis explains that this reduction in hallucination was not simply due to refraining from answering questions.



In the 'AA-Briefcase,' which assesses complex knowledge work over long periods, the GPT-6 Astra improved by approximately 80 Elo points from the GPT-5.6 Sol. On the other hand, in the 'GDPval-AA v2,' which assesses economically valuable work across 44 occupations, it decreased by approximately 80 Elo points.



In addition, GPT-6 Astra also decreased by about 2-3 points from GPT-5.6 Sol in τ3 -Banking, which deals with customer support for banks; SciCode, which deals with Python problems in the scientific field; and AA-LCR, which is a long-context inference test that deals with a large amount of documents.



The results of the Artificial Analysis suggest that GPT-6 Astra is not so much a model that significantly surpasses its predecessor in all capabilities, but rather a model that greatly improves token efficiency, particularly as a coding agent. While it achieves top-class performance in the Coding Agent Index with a low cost per task, its Intelligence Index remains at 61 points, the same as GPT-5.6 Sol, and the benefits of improved token efficiency are offset by the 2.5 times higher token price.

Related Posts:

in AI, Posted by log1i_yk