Who leads AI?

The Arena snapshot used here is dated September 2, before Astra’s September 3 analysis; it cannot rank Astra. The launch stays visible as a challenger, with no inferred preference lead. Artificial Analysis scores Astra at 61 on its Intelligence Index, five points behind Fable 5.1. In its Coding Agent Index, Astra scores 67 versus Fable 5.1 at 70. These results do not establish a new overall benchmark leader. Separately, ARC Prize reports state-of-the-art ARC-AGI-3 Semi-Private results: 62.7% with its Standard harness and 99.9% with a Provider Adapter. Those results measure different model-and-harness configurations, not the overall index.

This milestone marks the September 1 release and benchmark result. In the September 2 Arena snapshot, Fable 5.1 (max) is third by raw score at 1504 ±11, with a rank spread of 1–14. Fable 5 leads at 1507 ±5. Overlapping intervals do not establish a unique winner. Fable 5.1 remains visible here, but does not create an Arena lab handoff. Fable 5.1 scored 66, above Opus 5 at 63 and GPT-5.6 Sol and Grok 4.6 at 61 on the same index version. Testing used the default server-side fallback.

This benchmark evidence is retained for context. It does not establish a human-preference lead on this date. Opus 5 (max) was narrowly first at 61 on that version of the index, effectively tied with Fable 5 at 60. This is a model update within Anthropic, not a lab handoff.

This benchmark evidence is retained for context. It does not establish a human-preference lead on this date. Opus 4.8 scored 61.4, 1.2 points above the previous leader GPT-5.5. Anthropic therefore had a documented index lead before Fable 5 launched in June.

This benchmark evidence is retained for context. It does not establish a human-preference lead on this date. GPT-5.5 broke the three-way tie with a three-point overall advantage six days later. The leading configuration was xhigh reasoning effort.

This benchmark evidence is retained for context. It does not establish a human-preference lead on this date. The three models scored 57.3, 56.8, and 57.2. With a ±1-point confidence interval, Artificial Analysis explicitly recommended treating them as tied.

This benchmark evidence is retained for context. It does not establish a human-preference lead on this date. Thirteen days later, Gemini 3.1 Pro Preview led the index by four points over Opus 4.6 and led six of its ten evaluations.

This benchmark evidence is retained for context. It does not establish a human-preference lead on this date. The retrospective identifies OpenAI as the leader at the end of 2025 and shows GPT-5.2 (xhigh) at the frontier. This milestone records the month, without assuming the launch date was the independently confirmed handoff.

This benchmark evidence is retained for context. It does not establish a human-preference lead on this date. The report places Grok 4 ahead of o3-pro and describes this as OpenAI’s first loss of the frontier in its index. July is the reporting window; no exact handoff day is inferred.

OpenAI released o3-pro on June 10. The Artificial Analysis Q2 report later places Grok 4 ahead of o3-pro and describes that as OpenAI’s first loss of its index frontier. This entry marks o3-pro’s arrival as the prior frontier, not an independently dated handoff day.

Anthropic reported Claude Opus 4 and Sonnet 4 as coding and agentic highs, including SWE-bench and Terminal-bench. Contemporary coverage placed them in a WebDev Arena tie with Gemini 2.5 Pro. Gemini 2.5 Pro still led the overall text Arena. This is a capability milestone, not an overall leadership handoff.

OpenAI released o3 as the successor to o1. Contemporary Arena snapshots still had Gemini 2.5 Pro first on the overall text board. This restores the reasoning-model launch to the record without treating it as an overall preference or index handoff.

The original January 22 paper reports DeepSeek-R1 performing comparably to OpenAI o1-1217 on reasoning tasks and releases model weights plus distilled models. This restores the open-weight reasoning challenge to the historical record. A provider-reported comparison is not an independently measured overall Intelligence Index or Arena lead. The date is the paper’s publication, not a claim about when a reign began.

This benchmark evidence is retained for context. It does not establish a human-preference lead on this date. OpenAI released o1-preview and described gains in difficult reasoning tasks. Some headline results in the announcement apply to a forthcoming o1 version rather than the released preview. This entry marks the reasoning-model launch, not an independently established overall index reign.

LMSYS placed Claude 3.5 Sonnet first on coding and hard prompts, and second overall behind GPT-4o. That is a category high, not an overall Arena handoff. The June 20 date is the model launch; the Arena reading was reported days later.

ChatGPT launched as a free research preview on November 30, 2022. It belongs at the beginning of this history. Its launch date is not a measured Arena or benchmark lead, so it is excluded from the duration calculations.