StepFun · open weights on Arena · first on Arena 14 August 2025

step-3

151 days for open weights to match gpt-4.5-preview-2025-02-27

This open-weights model matched the score of gpt-4.5-preview-2025-02-27 in Vision 151 days after gpt-4.5-preview-2025-02-27 took first place.

Embed

Who held first place

<iframe src="https://jesseahawkins.com/frontier/embed/" title="Frontier: who held first place on each Arena leaderboard, and for how long" width="100%" height="560" style="border:0;max-width:100%" loading="lazy"></iframe>

Credits LMArena.

The open-weights catch-up

<iframe src="https://jesseahawkins.com/frontier/embed/catch-up/" title="Frontier: the days until open weights matched each closed leader" width="100%" height="900" style="border:0;max-width:100%" loading="lazy"></iframe>

Credits LMArena.

The saturation clock

<iframe src="https://jesseahawkins.com/frontier/embed/clock/" title="Frontier: the days from each benchmark’s release to a model at 90%" width="100%" height="640" style="border:0;max-width:100%" loading="lazy"></iframe>

Credits Epoch AI.

Each frame links back to its page. Preview it.

Cite

This page

Jesse Hawkins, “step-3 · Frontier”, jesseahawkins.com, data retrieved 30 Oct 2026. https://jesseahawkins.com/frontier/step-3/. Data: LMArena, Arena leaderboard dataset, revision 4491b06, CC BY 4.0; Epoch AI, Capabilities & benchmarking, CC BY 4.0.

BibTeX

@misc{hawkins2026frontierstep3,
  author = {Hawkins, Jesse},
  title = {{step-3 · Frontier}},
  year = {2026},
  howpublished = {\url{https://jesseahawkins.com/frontier/step-3/}},
  note = {Data: LMArena, Arena leaderboard dataset, revision 4491b06, CC BY 4.0; Epoch AI, Capabilities \& benchmarking, CC BY 4.0, retrieved 2026-10-30}
}

Arena · LMArena

LMArena, "Arena leaderboard dataset", Hugging Face, revision 4491b06. https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset. CC BY 4.0.

Benchmarks · Epoch AI

Epoch AI, 'Capabilities & benchmarking'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/benchmarks' [online resource]. CC BY 4.0.

Download

Post on X

Text: best place 32

place 211 of 419 on 28 October 2026

1st10th100th

On the board: 14 August 2025 to 28 October 2026, the newest snapshot
Score: 1350 (1343–1358) on 28 October 2026 · 6,291 votes

Vision: best place 11

place 106 of 167 on 19 October 2026

1st10th100th

On the board: 14 August 2025 to 19 October 2026, the newest snapshot
Score: 1176 (1165–1188) on 19 October 2026 · 3,193 votes
Matched: gpt-4.5-preview-2025-02-27 on 18 August 2025, 151 days after it took first place

Not in Epoch AI’s benchmark data

No model in Epoch AI’s list is the same release as this Arena name.

Sources

Arena scores come from people voting between two anonymous answers. First place is the highest score; the top two are often within each other’s margin of error. LMArena, Arena leaderboard dataset, revision 4491b06, retrieved 30 October 2026; Epoch AI, Capabilities & benchmarking, retrieved 30 October 2026. Both CC BY 4.0.