Part of: this research — view all 4 rows
Historical Baseline The forecast depends on the organization-level point gap between the #1 and #2 organizations on the Artificial Analysis Intelligence Index on December 31, 2028, divided by the 2024–2026 average gap. Reconstructing the historical baseline yields a mean gap of roughly 2.5 to 3.0 points. While today's gap sits at a temporarily elevated 4.6 points following Claude Fable 5.2's quick overtake of Gemini 4 Carbon Gemini 4 Carbon is the new #1 on the Artificial Analysis ..., historical gaps have usually ranged between 1 and 4 points. For example, Gemini 3.1 Pro led Claude Opus 4.6 by 4 points in early 2026 Google's Gemini 3.1 Pro Preview tops Artificial Analysis ... Gemini 3.1 Pro Preview: The new leader in AI, Grok 4 led by 3 points in mid-2025 Artificial Analysis on X: "xAI gave us early access to Grok 4, and the gap narrowed to just 1 point around the launch of index version 5 Announcing Artificial Analysis Intelligence Index v5 Announcing Artificial Analysis Intelligence Index v5.
Benchmark Mechanics and Saturation The most critical confound for this ratio is the mechanical lifecycle of the index itself. As an evaluation suite ages, models saturate the benchmarks, artificially compressing the spread between the frontier competitors. Conversely, major index rescalings—such as the shift to version 5, which introduced harder, largely private evaluations and widened the open-weights gap from 3 to 12 points Announcing Artificial Analysis Intelligence Index v5—mechanically stretch out the gaps. Because older scores are not directly comparable to new ones Announcing Artificial Analysis Intelligence Index v5 Artificial Analysis Intelligence Benchmarking Methodology and the continuous Progress Index is not the target resolution surface https://artificialanalysis.ai/methodology/progress-index, the numerator will be heavily dictated by whether December 2028 coincides with a newly launched evaluation suite or a saturating, late-life benchmark.
Competitive Dynamics The original premise for an accelerating capability gap rests on models learning on the job. However, as of late 2026, no frontier lab is updating model weights from live deployment sessions. While Anthropic has established a formidable revenue and compute lead, and hyperscalers dominate the infrastructure, frontier algorithmic improvements (advancing roughly 3x a year) tend to diffuse quickly across the top players. Consequently, a persistent, compounding monopoly from deployment data has not yet materialized. Instead, the landscape remains a tight three-way race of leapfrogging between Anthropic, Google, and OpenAI. Set against related questions, this supports a stable expectation that the top two labs will maintain historically normal margins rather than collapse into a dead heat: intense competition need not eliminate the point gap at any particular snapshot.
Tail Risks and Final Distribution The median expectation of 1.0 reflects a continuation of this intense competitive leapfrogging, with the 2028 snapshot landing at the historical mean. The 25th and 75th percentiles of 0.55 and 1.7 allow substantial variation around historically normal margins without making a near-dead heat the central expectation. The lower tail, with a 10th percentile of 0.25, accounts for a scenario where two labs ship near-identical frontier models late in the year, or where the active index version is heavily saturated and mechanically compresses scores. The right tail, with a 90th percentile of 2.8, leaves room for a scenario where continuous deployment learning begins to operate and grants one lab a breakaway advantage, or where Artificial Analysis introduces a significantly harder index revision right before the resolution date, temporarily blowing out the point spread.
Set against related questions, this was adjusted slightly upward to reflect a stable expectation that the top two labs will maintain historically normal margins without collapsing into a dead heat.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited