When will an open source AI model match Fable and Astra?

Something I hadn’t quite appreciated until recently is how quickly open models are improving. The best ones are now performing at a level that was state of the art only a few months ago.

I’ve started to notice this in everyday use. I often switch between models on side projects and occasionally blind-test them. Increasingly, I’m seeing open models match or outperform closed alternatives.

But my impressions only go so far. To get a better sense of how open models are faring, and to forecast when they might reach the performance of today’s best closed models (specifically Claude Fable 5.1 from Anthropic and GPT Astra from OpenAI), I dug into Epoch AI’s public data.

Epoch publishes a Capabilities Index (ECI) that combines results from more than 50 benchmarks into a single score. It’s a blunt instrument that obscures some of the differences between models, but it’s a fast and useful way to compare broad model capabilities and how they’re changing over time.

On this index, Kimi K3 is the highest-scoring open model as of 6 September 2026. It scores 158, just a few points below GPT 5.6 Sol and Claude Opus 5 on 162, and Claude Fable 5.1 on 163.

Bar chart of Epoch Capabilities Index scores on 6 September 2026. GPT-6 Astra 169, Claude Fable 5.1 163, Claude Opus 5 162, GPT-5.6 Sol 162, Kimi K3 158, DeepSeek V4 Pro 0813 155. The last two are open weights.

In May this year, Epoch also published an analysis of the gap between open and closed models. Today, I extended this analysis with the latest data and recreated their chart, with help from Astra in ChatGPT Work. I’ve also added some simple forecasts to go along with it.

This capability gap (in other words, how long it takes for open models to catch up) works out at roughly 4 to 6 months. How do you get that range? Take Kimi K3’s score of 158 from September this year and then find the last time a score of 158 led the field. That was early March 2026, just before GPT-5.4 Pro raised the bar with a score of 159.

So going by these numbers alone, open models are about six months behind. But these scores come with uncertainty and Epoch has a comparison rule that accounts for this. Using that rule, the capability lag is roughly 4.5 months. And that’s where the 4.5 to 6 month range comes from.

Step chart of the highest available Epoch Capabilities Index score over time, January 2023 to September 2026, with separate lines for open and closed weight models. The open line trails the closed line by roughly four and a half to six months.

With that in mind, when will an open model reach the capabilities of today’s state-of-the-art closed models, namely Fable or Astra?

If we take the pace of open model progress over the past 3, 6, and 12 months, and then project each pace forward from Kimi K3’s latest score, we get three possible timelines.

Let’s start with Astra. It has an ECI score of 169. That’s around 11 points ahead of Kimi K3. Over the past six months, the trend in open model progress has been around 1.9 points a month. If they keep that up, they’ll get to Astra levels of performance in roughly 6 months. That takes us to March 2027. At the same pace, we get Fable levels in late November 2026.

Bar chart of how long open models might take to reach Astra at 169 ECI and Fable 5.1 at 163 ECI, starting from Kimi K3 at 158. Three scenarios use the pace of the last 90, 180 and 365 days.

Things could move faster. If we use the pace of the last three months, we could see open models on a par with the performance of Astra today by January 2027. In contrast, if we use the average pace of the last year, we would expect Astra-level performance in July 2027.

Of course, these are rough projections. There’s a lot that could change and closed models aren’t stopping either. Still, it’s fun to forecast and make a few simple predictions based on historic data.

I’ll check in on progress in December this year and again in April and August in 2027 to see how these forecasts hold up.