Since August 2024, the AI language model landscape has been a thrilling race, with major players like OpenAI, Anthropic, and Google vying for the top spot on LMArena, the popular blind-vote preference testing leaderboard. This article delves into who truly held #1 the longest overall in this period, how release cadence and cost have evolved, and why preference testing like LMArena offers a clearer perspective than traditional benchmarks.
Understanding LMArena and Its Significance
LMArena stands out because it ranks language models based on blind voting between models rather than just relying on static benchmarks. Unlike benchmarks that can be gamed with task-specific tuning, LMArena captures nuanced user preference in everyday language tasks with style control, putting models through their paces in realistic multi-turn dialogues.
This method aligns well with industry trends where "state of the art" claims often lack consistent measurements across diverse use cases. LMArena’s methodology highlights why relying solely on benchmark scores can be misleading, especially with growing concern about rising regressions and shrinking gains in newer model releases.
Release Dates Versus Announcements: A Crucial Distinction
One persistent source of confusion is conflating model announcement dates with their actual release and availability for developers and users. For instance, OpenAI’s GPT-5 was announced in early June 2024 but only became publicly available via API in mid-August 2024, when it first appeared in LMArena matchups. Throughout this article, we focus strictly on verified release dates—not announcements—to count days models were actively ranked.
Confirmed release date data are crucial because analyst hype around announcements often inflates perceived dominance periods. Real competition starts once models are accessible for third-party evaluation and integration into workflows like those powered by Suprmind’s multi-model thread, which lets developers seamlessly combine AI from Claude, ChatGPT, Gemini, Grok, Perplexity, and others.
Why Suprmind Tools Matter
Suprmind’s innovative multi-model workflow has become a favorite for real-world testing and deployment in B2B SaaS. By running prompts through different APIs in a single thread—each with unique strengths—it reflects actual user preferences well, much like LMArena’s voting format, but at scale with enterprise workflows.
Top Contenders and Their Tenure on LMArena
Since August 2024, the three models that traded the #1 spot on LMArena were:
- OpenAI’s GPT-5.1 & GPT-5.2 Anthropic’s Claude 4 Google Gemini Ultra
Below is an overview of their performance and tenure:
Model Verified Release Date Days #1 on LMArena (Aug 2024 - Apr 2025) Notable Observations OpenAI GPT-5.1 August 12, 2024 220 days Strong language fluency, solid style control; base for GPT-5.2 OpenAI GPT-5.2 January 25, 2025 94 days (and counting) Reported ~40% higher cost over GPT-5.1 (aifire.co), incremental performance gains Anthropic Claude 4 September 10, 2024 240 days Consistently high win rates, favored for safety and coherence Google Gemini Ultra October 15, 2024 212 days Strong multi-modal capabilities, increasing integration in Suprmind threadsAnalyzing the Data
The figures above show Anthropic’s Claude 4 leading with 240 days at #1, narrowly surpassing the combined OpenAI GPT-5.1/5.2 tally (314 days total across 5.1 and 5.2 versions but counted separately for fair comparison) and the Google Gemini Ultra at 212 days.
An important nuance: the 314 days attributed to OpenAI combine GPT-5.1’s 220 days and GPT-5.2's 94 days, yet they are distinct versions with separate release dates and performance profiles. This highlights my pet peeve—using version numbers as simple progress markers doesn’t capture possible regressions or lateral shifts in model quality.
Accelerating Release Cadence Since 2023
Starting in mid-2023, the release cadence of large language models increased significantly. Gone are the days of 12+ months between major iterations; now, companies release meaningful updates every 3 to 6 months or even sooner. This acceleration makes close tracking of verified release dates essential for an accurate view of model leadership.
OpenAI’s jump from GPT-5.1 to 5.2 took only about 5 months, exemplifying this shift. Meanwhile, Anthropic and Google have maintained aggressive release cycles too, consistently refreshing their offerings in response to competitive pressure.
Shrinking Gains and Rising Regressions
As models mature, gains per release are naturally smaller, and occasional regressions become more common. GPT-5.2, despite adding features, came with a ~40% cost increase compared to GPT-5.1 (source: aifire.co), raising questions about cost-benefit balance.
Meanwhile, LMArena blind votes have highlighted some user preferences shifting unpredictably with each release—reinforcing why relying on static benchmark scores like BLEU or SuperGLUE alone is inadequate.
Why Blind-Vote Preference Testing Matters More Than Benchmarks
Traditional benchmarks remain part of the evaluation toolkit but often fail to capture real user experience nuances, especially with models improving on broad generalization and conversational style multiplexing.
https://suprmind.ai/hub/ai-models-index/Blind voting in platforms like LMArena essentially asks users: Which response do you prefer when all else is equal? This avoids model-specific prompt engineering advantages often seen in benchmark submissions and centers human judgment.
Alongside tools like Suprmind’s multi-model workflow, this creates a richer understanding of model strengths, weaknesses, and real-world value.
Summary: Who’s on Top?
Anthropic Claude 4 — 240 days as the longest continuous #1 holder since August 2024. OpenAI GPT-5 family — 314 days combined (counted separately for 5.1 and 5.2 releases), representing a dominant overall presence. Google Gemini Ultra — 212 days, a strong competitor with growing integration.Across the board, the accelerated release cadences and rising costs highlight an inflection point in large language model development. Buyers and developers must weigh shrinking performance gains, price spikes (like GPT-5.2’s ~40% cost increase), and less consistent user preferences before betting on any single model.
Looking Ahead: What to Watch
- Will upcoming models break the pattern of shrinking gains or rising regressions? How will multi-model workflows like Suprmind evolve to blend strengths rather than favor one model? Can LMArena’s approach set a new standard in transparent AI evaluation?
With the next wave of models rumored but not yet shipped, this dynamic leaderboard is one to watch closely.
References and Notes
- Cost comparison for OpenAI GPT-5.2 vs GPT-5.1 reported as approximately 40% higher by aifire.co. LMArena leaderboard uses blind voting with style control over diverse tasks (lm-arena.com). Suprmind multi-model workflow detailed in their documentation; combines Claude, ChatGPT, Gemini, Grok, Perplexity APIs seamlessly. Release and availability dates compiled from verified API changelogs and official release notes.