How GPT-6 Became No. 1: Artificial Analysis CEO Explains The Much-Debated Index Updates
Set Trending Topics as a preferred source on Google.
Two updates in a matter of days, and a new model at the top: Shortly after GPT-6 Astra launched, Artificial Analysis revised its closely watched Intelligence Index twice. OpenAI’s new model climbed from fifth place to a shared lead with Anthropic’s Claude Fable 5.1, both posting a displayed score of 53. The rapid shift raises a question: Why did one of the most influential AI benchmarking platforms change its yardstick at that particular moment?
Micah Hill-Smith, co-founder and CEO of Artificial Analysis, said in response to questions of Trending Topics that the previous index had failed to adequately capture the new models’ capabilities. He rejected the suggestion that the changes were designed to favor a particular model. “I can assure you that we do not make changes to intentionally benefit any particular model,” Hill-Smith said. The goal, he added, is to give developers the most accurate and useful picture possible when choosing models.
From Fifth Place to a Shared Lead
The results over the past several days show how substantially methodology can affect a ranking. Under the previous index, GPT-6 Astra scored 61 points, level with its predecessor GPT-5.6 Sol and behind Fable 5.1 at 66. Version 4.2 narrowed the gap between the two new flagship models to 55 against 57. Version 4.3 then brought both to 53, as Trending Topics reported.
Those scores are meaningfully comparable only within the same version of the index: Each revision changes the basis of assessment. GPT-6’s climb illustrates how strongly the selection and weighting of tests can influence the standings.
According to Hill-Smith, that basis of assessment had become a problem. “Fable 5.1 and GPT-6 Astra demonstrated capabilities that we believed Index v4.1 wasn’t adequately capturing, in part because earlier models had already saturated several of the benchmarks used in Index v4.1.” When many capable models approach a test’s maximum score, it becomes harder for that test to distinguish between them.
Internal Results Prompted the Updates
The decision to accelerate the updates was based on measurements the company already had. “We had internal results last week that materially differed from the picture presented by Index v4.1,” Hill-Smith said.
The team decided that updating the index and website quickly would serve users better than waiting for the larger, previously planned upgrade. Artificial Analysis therefore brought forward evaluations that had already been built or implemented for version 5.
The CEO’s explanation confirms that the performance of Fable 5.1 and GPT-6 Astra influenced the timing of the changes. In the company’s view, the existing mix of tests was no longer keeping pace with advances among leading models.
More Knowledge Work, Harder Tasks and Private Test Data
The changes are documented in the company’s announcements. Version 4.2 added AA-Briefcase, which evaluates complex knowledge work, and GDP.pdf, which tests reasoning across extensive documents. GPQA Diamond, a scientific reasoning benchmark, was removed because it had become largely saturated.
The share of the index’s weighting assigned to evaluations with private tasks or answers also rose to 40 percent. Keeping that material private is intended to make it harder for model developers to optimize their systems specifically for known test questions.
Version 4.3 then replaced Terminal-Bench 2.1 with the more demanding version 4.0. AutomationBench-AA also replaced the previous banking evaluation with broader business workflows spanning multiple applications. The weighting assigned to private test components rose to 45 percent, while the overall category weights remained unchanged from version 4.2. Together, the changes shifted the assessment further toward complex tasks that AI agents are expected to complete autonomously.
How Reliable Is a Ranking That Changes So Quickly?
The pace of those revisions still presents a challenge for interpreting the leaderboard. Someone checking the index shortly after a model launches may find a substantially different ranking a few days later. Comparisons over time require attention to the underlying test conditions.
Hill-Smith forcefully defended the platform’s ambition: “I feel comfortable asserting that it has consistently been, and will continue to be, the best single metric for model intelligence.”
At the same time, he pointed to uneven performance profiles across models. Different models win different evaluations, he explained. Artificial Analysis therefore publishes every component result used to calculate the index, alongside additional results excluded from the overall score. For developers, the individual tests most relevant to their work can offer a more precise guide to a model’s suitability than its overall position alone.
Adam D’Angelo’s Ties to Artificial Analysis and OpenAI
Hill-Smith also addressed questions about potential influence. Adam D’Angelo, the co-founder and CEO of the question-and-answer platform Quora and a former chief technology officer at Facebook, is an angel investor in Artificial Analysis.
D’Angelo also sits on the board of directors of the OpenAI Foundation. The foundation controls OpenAI and can appoint and remove every director of the commercial OpenAI Group, according to the company’s explanation of its governance structure. His roles therefore create a connection between the benchmarking platform and the company behind GPT-6.
Hill-Smith explicitly denied that D’Angelo had influenced either update. “Adam is an angel investor in Artificial Analysis. He had no influence on these updates.” Investors are excluded from methodology decisions, he added: “No investor has ever influenced any benchmarking methodology decision, and we do not involve investors in these decisions in any way.”
“OpenAI Gave No Feedback” Related to the Updates
Hill-Smith was equally specific when asked about feedback from OpenAI. “I can specifically confirm OpenAI gave no feedback that related to the v4.2 or v4.3 upgrades.” The decisions, he said, were made entirely at Artificial Analysis’s discretion.
The platform regularly gathers input from companies building and buying AI technology. It assesses that feedback on its technical merits and makes its own methodology decisions. The published documentation makes the changes to the index traceable. The account of how those decisions were made internally remains the company’s own.
Version 5 Is Targeted for Late October
Another major revision is already planned. Artificial Analysis is targeting late October for Index v5, although the schedule could change following the two accelerated updates.
Hill-Smith declined to predict which model would lead: “At this time I have no idea whether GPT-6 will be in first place.” Internal, unpublished results showed impressive behavior from both Fable 5.1 and GPT-6, he said. The company would continue developing benchmarks, including with partners, to test capabilities it considers important to developers and users.
His personal assessment of the competition was clear: “The AI race is more competitive than ever, and I personally am primarily using both GPT-6 and Fable 5.1 this week.”
