Proximal is a research lab focused on data.

We turn hard engineering problems into useful training data for frontier models.

Read Blog

Introducing FrontierSWE

Our Ultra-Long-Horizon Coding Benchmark

See Benchmark→

Leaderboard

#Modelavg RankDominance

GPT-5.5

Codex

2.3583%

Claude Opus 4.7

Claude Code

3.1873%

Claude Opus 4.6

Claude Code

3.8864%

GPT-5.4

Codex

3.9763%

Gemini 3.1 Pro

Gemini CLI

5.2647%

DeepSeek V4 Pro

Claude Code

6.2634%

Kimi K2.6

Kimi CLI

6.4432%

Kimi K2.5

Kimi CLI

6.7428%

Qwen3.6-Plus

Qwen Code

6.9126%

#	Model	Harness	avg Rank	Dominance	Implementation	Performance	Research
1	GPT-5.5	Codex	2.35	83%	3.20	1.89	2.33
2	Claude Opus 4.7	Claude Code	3.18	73%	2.60	3.67	2.67
3	Claude Opus 4.6	Claude Code	3.88	64%	3.60	4.22	3.33
4	GPT-5.4	Codex	3.97	63%	3.00	5.28	1.67
5	Gemini 3.1 Pro	Gemini CLI	5.26	47%	6.40	3.94	7.33
6	DeepSeek V4 Pro	Claude Code	6.26	34%	6.00	6.61	5.67
7	Kimi K2.6	Kimi CLI	6.44	32%	4.90	7.33	6.33
8	Kimi K2.5	Kimi CLI	6.74	28%	7.10	6.11	8.00
9	Qwen3.6-Plus	Qwen Code	6.91	26%	8.20	5.94	7.67

Rank: avg position across tasks (lower = better)Dominance: win rate vs random opponent on task.

Recent highlights2

Announcements, benchmark releases, and work that shows the direction of the lab.

April 2026·Benchmark

FrontierSWE

Benchmarking coding agents at the limits of human abilities.

Read post

Feb 18, 2026·Announcement

Announcing Proximal

Introducing Proximal — a research lab building the data engine for autonomous coding agents.

Read post

Careers

We’re hiring across research and engineering

View open roles→