Scroll
The frontier AI race, from public evidence
One road, two lanes, 9.4 points apart.
The West leads the measured frontier. China leads the open one. Walk the road so far, milestone by milestone.
0 / 10 milestones reached
WASD walk · Shift run · QE or drag to orbit · click to goTap to walk · swipe sideways to orbit
-
2017Context
The Transformer
“Attention Is All You Need” introduces the architecture nearly every frontier model is built on.
-
2020Context
Scale starts to speak
GPT-3, at 175 billion parameters, shows that a large enough language model can learn new tasks from a few examples.
-
2022Context
The public arrives
ChatGPT launches on 30 November 2022, and language models leave the lab.
-
2023Tracked data
Both lanes on the scale
The first year both ecosystems have a scored model on Epoch’s Capabilities Index.
- GPT-4 (Mar 2023)
- 125.9
- Yi-34B
- 117.4
- Gap
- 8.5
-
2024Tracked data
Both lanes climb 15 points
o1 leads the West; DeepSeek-V3 leads China. The West adds 16 ECI in the year, China 15.
- o1
- 141.8
- DeepSeek-V3
- 132.3
- Gap
- 9.5
-
2025Tracked data
The widest gap yet
GPT-5.2 Pro leads the West; DeepSeek-V3.2 leads China. The West adds 14 ECI in the year, China 14.
- GPT-5.2 Pro
- 156.0
- DeepSeek-V3.2
- 146.1
- Gap
- 9.9
-
2026Tracked data
Today’s frontier
Claude Opus 5.5 leads the West; Kimi K3 leads China — with open weights.
- Claude Opus 5.5
- 166.8
- Kimi K3
- 157.4
- Gap
- 9.4
-
2026Tracked data
China leads the open road
On the open-weight frontier the lanes swap: Kimi K3 sits 7.2 points clear of the best Western open-weight model.
- Open-weight lead
- China +7.2
- Labs above 145
- 6 : 5
-
2026Tracked data
Five measures, no clean sweep
The West leads FrontierMath, SWE-bench Verified and ARC-AGI-2. China leads Video-MME. METR task horizon has no result on one side.
- Video-MME
- 79.7%
- METR task horizon
- 3.1 hr
- ARC-AGI-2
- 95.0%
-
—Horizon
AGI
Not on this scale. ECI measures public evaluations; it doesn’t forecast arrival. Walk closer and the horizon moves.
Where it stands
The West leads the measured frontier. China leads the open frontier.
Western ecosystem
166.8
Claude Opus 5.5
+9.4
ECI gap · 2026
Chinese ecosystem
157.4
Kimi K3
-
Composite frontier
West +9.4
Highest ECI in each ecosystem.
-
Open-weight frontier
China +7.2
Led by Kimi K3, released with open weights.
-
Labs above 145 ECI
6 : 5
Western labs to Chinese labs.
-
Selected benchmark leads
3 : 1
One of five measures has no result on one side to compare.
The frontier, year by year
Both lanes climbed 39 points. Neither closed the gap.
The highest dated ECI score in each ecosystem at year end. Since both first had a scored model, the Western lead has stayed between 8.5 and 9.9 points.
Why track it this way
The path to AGI is now an ecosystem competition, not a single-company story. So we follow the strongest Western and Chinese systems side by side, who is closing the gap, and where open weights change the balance.
Two plausible paths
The winner depends on what kind of race this becomes.
Lab depth
This is no longer a two-company race.
Each bar is a lab’s highest deduplicated ECI score. Six Western and five Chinese labs now clear 145.
- AnthropicClaude Opus 5.5 166.8
- OpenAIGPT-6 Astra 166.4
- Google DeepMindGemini 4 Argon 164.6
- Meta AIMuse Spark 1.3 156.8
- xAIGrok 4.6 156.3
- Mistral AIMistral Large 4 146.9
- MicrosoftPhi-4 130.4
- Moonshot AIKimi K3 157.4
- Alibaba / QwenQwen 3.8 Max 156.9
- Zhipu AI / Z.aiGLM-5.3 155.9
- DeepSeekDeepSeek V4.1 Flash 155.6
- MiniMaxMiniMax-M3 147.1
- 01.AIYi-34B 117.4
- BaichuanBaichuan2-13B 103.3
Evidence desk
Ask what a score means before deciding what it proves.
Answers are written from this site’s method notes and the current snapshot. The desk explains the record; it can’t change the score.
Pick a question, or type your own about a benchmark, a score, an access label or a method term.