How Long Does the #1 AI Model Stay on Top?
A frontier model now holds Arena's #1 spot for about 3 weeks. A lab that keeps shipping can hold it for months.
The signal
- LaunchesThree flagships in three days: Claude Sonnet 5.5 (28 Sep), GPT-6.1 Sol (29 Sep), Gemini 4 Argon (30 Sep).
- ArenaThe 2 Oct 2026 snapshot puts gemini-4-argon-high at #1. It ends a 236-day Anthropic run.
Since Jan 2025, Arena's #1 spot has changed hands 16 times. Half of the finished reigns lasted 23 days or less.
Days each reign lasted, by the top-rated model on the Arena text board (overall), Jan 2025 – 2 Oct 2026. One bar per reign, in date order.
#1 = highest point rating on each snapshot; confidence intervals often overlap at the top. A reign runs from the first snapshot at #1 to the first snapshot with a new #1. Snapshots are irregular, so durations are ±a few days. The current reign is still running and is excluded from the median.
Show the data (16 rows)
| Model | Days at #1 |
|---|---|
| deepseek-r1 | 38 |
| grok-3-preview-02-24 | 23 |
| chatgpt-4o-latest-20250326 | 28 |
| o3-2025-04-16 | 62 |
| gemini-2.5-pro | 145 |
| gemini-3-pro | 82 |
| claude-opus-4-6-thinking | 13 |
| claude-opus-4-6 | 27 |
| claude-opus-4-6-thinking | 130 |
| claude-opus-5-high | 1 |
| claude-opus-5-max | 23 |
| claude-opus-5-high | 8 |
| claude-opus-5-max | 6 |
| claude-fable-5.1-max | 23 |
| claude-opus-5.5-high | 5 |
| gemini-4-argon-high (still #1) | 2 |
The crown changes hands
One pedestal per reign at #1 since Jan 2025, in date order; height = days at #1. Press a pedestal or use the arrow keys.
The Mode read
Between 28 and 30 September, Anthropic, OpenAI and Google each shipped a flagship model. By the 2 Oct 2026 Arena snapshot, gemini-4-argon-high was #1. That ended a 236-day run in which Anthropic held the top spot with 6 different models.
Single models do not hold the lead for long. Of the 15 reigns at #1 that started and ended since January 2025, half lasted 23 days or less, and 5 lasted under two weeks. 13 different models took turns. A model lead is now an asset that lasts one launch cycle.
This is the parity clock. If a frontier lab cannot hold a lead for a month, a startup whose pitch is "we use the best model" cannot hold one either. The durable assets are a fast shipping cadence and a customer relationship strong enough that you can change the model under the product and nobody leaves.
A model lead lasts weeks. A customer relationship lasts years.
What we would back: Products that route across models and own the workflow, not the weights.