Some thoughts on how many AI companies would have to pause vs how long a pause could last before their closest competitor catches up. I used ECI (Epoch Capability Index) as a measure of capabilities and the following method to estimate the lag:
1) Find company X’s best model (highest ECI), let’s call it’s ECI score x.
2) Build OpenAI’s frontier curve: sort OpenAI models by release date, keep only the running maximum of ECI.
3) Find when that curve first reached x. Since the frontier jumps in discrete steps, the exact moment almost never exists, so linear interpolation is used.
4) Lag = today minus that date, converted to months. If x is below OpenAI’s oldest tracked model, you only get a lower bound.
OpenAI: 0 months behind (by construction)
Anthropic (best model: Claude Fable 5): 3 months behind (90% confidence interval includes 0 though)
Moonshot AI (best model: Kimi K3): 5 months behind
Mistral (best model: Mistral Medium 3.5): 20 months behind. This figure could be an overestimate because Epoch doesn’t have a score for Mistral Large 3
Microsoft (best model: Phi-4): 26 months behind (wide confidence interval)
Amazon (best model: Amazon Nova Pro): ≥34 months behind (it’s the only Amazon model on the Epoch leaderboard)
01.Ai (best model: Yi-34B): ≥41 months behind
This can be used to calculate “You need to stop X companies to get a Y-month pause”—how many months it would take for the next-best unpaused competitor to catch up. Now, we’re going to make quite the assumption: companies pause in a clear order, starting from ones with the best (highest ECI) models. Another important assumption is that rate of progress of any given company doesn’t depend on how many others have paused.
With these assumptions you get this staircase graph.
At N=2 (OpenAI and Anthropic) you get a 5-month pause until Moonshot AI catches up.
To get a 9-month pause you need N=7.
To get a 12-month pause you need N=9.
To get a 18-month pause you need N=11.
To get a 24-month pause you need N=13.
To get a 36-month pause you need N=14.
So a 5-month pause is quite feasible since only OpenAI and Anthropic would have to pause. Anything more than that would require a lot of companies cooperating.
You need to stop X companies to get a Y-month pause
Some thoughts on how many AI companies would have to pause vs how long a pause could last before their closest competitor catches up. I used ECI (Epoch Capability Index) as a measure of capabilities and the following method to estimate the lag:
1) Find company X’s best model (highest ECI), let’s call it’s ECI score x.
2) Build OpenAI’s frontier curve: sort OpenAI models by release date, keep only the running maximum of ECI.
3) Find when that curve first reached x. Since the frontier jumps in discrete steps, the exact moment almost never exists, so linear interpolation is used.
4) Lag = today minus that date, converted to months.
If x is below OpenAI’s oldest tracked model, you only get a lower bound.
OpenAI: 0 months behind (by construction)
Anthropic (best model: Claude Fable 5): 3 months behind (90% confidence interval includes 0 though)
Moonshot AI (best model: Kimi K3): 5 months behind
Alibaba (best model: Qwen 3.8 Max): 6 months behind
xAI (best model: Grok 4.5): 8 months behind
Meta (best model: Muse Spark 1.1): 8 months behind
Google DeepMind (best model: Gemini 3.1 Pro): 8 months behind. This figure could be an overestimate, it’s plausible Google isn’t that much behind
DeepSeek (best model: DeepSeek V4 Flash 0731): 9 months behind
Z.ai (best model: GLM-5.2): 9 months behind
Thinking Machines (best model: Inkling): 12 months behind (wide confidence interval)
MiniMax (best model: MiniMax-M3): 13 months behind
Nvidia (best model: Nemotron 3 Ultra): 20 months behind
Mistral (best model: Mistral Medium 3.5): 20 months behind. This figure could be an overestimate because Epoch doesn’t have a score for Mistral Large 3
Microsoft (best model: Phi-4): 26 months behind (wide confidence interval)
Amazon (best model: Amazon Nova Pro): ≥34 months behind (it’s the only Amazon model on the Epoch leaderboard)
01.Ai (best model: Yi-34B): ≥41 months behind
This can be used to calculate “You need to stop X companies to get a Y-month pause”—how many months it would take for the next-best unpaused competitor to catch up.
Now, we’re going to make quite the assumption: companies pause in a clear order, starting from ones with the best (highest ECI) models. Another important assumption is that rate of progress of any given company doesn’t depend on how many others have paused.
With these assumptions you get this staircase graph.
At N=2 (OpenAI and Anthropic) you get a 5-month pause until Moonshot AI catches up.
To get a 9-month pause you need N=7.
To get a 12-month pause you need N=9.
To get a 18-month pause you need N=11.
To get a 24-month pause you need N=13.
To get a 36-month pause you need N=14.
So a 5-month pause is quite feasible since only OpenAI and Anthropic would have to pause. Anything more than that would require a lot of companies cooperating.