In it you will find this graph, depicting the results so far for Anthropic's R&D Automation Index, using a scale developed by Epoch AI:
First of all, I’d like to thank Anthropic for sharing this information. Tracking things like this, and making the results public, is key for keeping up with the rapid development for those of us observing things from outside the major AI companies. So, thank you Anthropic!
The Automation Index “runs from AL0 (no AI involvement) to AL5 (AI operates fully autonomously, with no human in the loop). In AL3, AI “collaborates”: it can do large chunks of work under close human direction. In AL4, AI “leads”: it can complete most of the task end-to-end from a high-level prompt, while the human supervises.”
A brief description of the method: for each week of July 2026, Anthropic sampled 20% of staff from each department in the model R&D loop and had a Claude agent list the tasks they worked on. The resulting ~15,000 tasks were organized into a tree of 542 nodes (378 of them leaves). For each month, an independent Claude judge assigns every node an automation level, and each node carries a weight based on person-time spent on it (a proxy for how important that work is to the R&D). The latest data point is August 2026.
With the data provided, we can investigate what different scenarios for future AI automation might look like. Claude Opus 5 helped me extract data from the graph and make some projections.
For instance, we can fit a logistic curve to the AL4 data (AI “leads”)[1]. Then let’s assume that AL5 lags behind AL4 as much as AL4 has lagged behind AL3 so far (7.3 months on average) and add that to the graph (pinkish area). Let's also project a second scenario where the AL4→AL5 lag matches the AL3→AL4 lag at each level (dashed line). Then we get this:
AL5 projections built on AL4's own trend, with a constant 7.3-month lag (pinkish area) and a level-matched lag (dashed line), giving 80% automation in November 2027 (constant lag) to February 2028 (level-matched lag).
What if the AL4 curve instead looks more like AL3? We can project AL4 forward with the assumption that it is logistic and has the same slope as AL3 (k=6.9/yr). Then add AL5 projections: one with the same 7.3-month lag as before, and two dotted lines spanning a lag range of 6.3 to 8.5 months (roughly how much AL4 has lagged behind AL3 at different AL4 scores so far). Then we get this:
AL4’s slope is set to equal AL3’s slope, pulling AL5's 80% crossing forward to August 2027, assuming a 7.3-month lag between AL4 and AL5 (or July - September 2027 for a lag range of 6.3 to 8.5 months).
According to these projections, it seems plausible that AL5 reaches 80% between mid-2027 to early 2028[2]. Of course, the slope for AL5 could look completely different from AL3 and AL4. Perhaps the lag is much longer than 7.3 months (it probably isn’t much shorter, since AL5 is still at 0%.)
I expect the lag to be longer than 7.3 months, since the step change from AI “leads” to full automation seems quite large to me, and I expect the slope to be steeper for AL5 than AL4 due to recursive self-improvement.
These projections point to two things to watch over the coming months:
Will AL4 rise to ~50% in October 2026? This would indicate that AL4 is roughly following the same slope as AL3.
Will AL5 rise to 1-2% in Sep - Nov 2026? This would indicate that the AL4→AL5 lag is roughly equal to the AL3→AL4 lag (~6-8 months).
I hope Anthropic continues to publish regular updates for their R&D Automation Index so we can get answers to these questions relatively soon.
A few things these projections don’t consider:
The AL3, AL4 and AL5 curves are not independent (e.g. the AL3 curve is really the combined share of work at or above AL3). This means that errors are correlated across the AL3 and AL4 curves, and month to month relative changes are probably more reliable than absolute AL scores.
The projections don’t account for how the AI automation affects its own growth, initiating an intelligence explosion. Once AL5 starts growing, it might climb much faster than AL3 or AL4 did.
AL5 is partly about permitting AIs to automate everything, not just a capability. Does Anthropic want to remove humans from the process? Will they do much more extensive testing and safety training before they dare move from AL4 to AL5, increasing the lag? Are there tasks that Anthropic will never automate entirely, lowering the ceiling that AL5 can reach?
Also, I haven’t double-checked all the extracted data, assumptions, and code produced by Claude. These graphs are for illustrating potential future trajectories, rather than get everything exactly right. The graphs look reasonably accurate to me, but there are probably some minor mistakes. Feel free to check Claude’s work here.
Thank you for reading! If you found value in this post, consider checking out my blog: Forecasting AI Futures.
Note that by “AL4 logistic curve”, I mean the curve separating the AL3 from AL4 areas, which actually corresponds to a projection of the combined scores of AL4 and AL5.
I’m focusing on the 80% threshold here because it’s a point where AI can automate most AI R&D, and because higher thresholds may simply not be reached if Anthropic decides to not permit some tasks to be completely automated (at least for some time). Check the graphs to see timelines for other thresholds.
Anthropic recently released some very interesting information about the degree of AI automation for R&D tasks. I recommend reading the entire article: Measurements for understanding the pace of AI development inside frontier labs (Sep 17, 2026).
In it you will find this graph, depicting the results so far for Anthropic's R&D Automation Index, using a scale developed by Epoch AI:
First of all, I’d like to thank Anthropic for sharing this information. Tracking things like this, and making the results public, is key for keeping up with the rapid development for those of us observing things from outside the major AI companies. So, thank you Anthropic!
The Automation Index “runs from AL0 (no AI involvement) to AL5 (AI operates fully autonomously, with no human in the loop). In AL3, AI “collaborates”: it can do large chunks of work under close human direction. In AL4, AI “leads”: it can complete most of the task end-to-end from a high-level prompt, while the human supervises.”
A brief description of the method: for each week of July 2026, Anthropic sampled 20% of staff from each department in the model R&D loop and had a Claude agent list the tasks they worked on. The resulting ~15,000 tasks were organized into a tree of 542 nodes (378 of them leaves). For each month, an independent Claude judge assigns every node an automation level, and each node carries a weight based on person-time spent on it (a proxy for how important that work is to the R&D). The latest data point is August 2026.
With the data provided, we can investigate what different scenarios for future AI automation might look like. Claude Opus 5 helped me extract data from the graph and make some projections.
For instance, we can fit a logistic curve to the AL4 data (AI “leads”)[1]. Then let’s assume that AL5 lags behind AL4 as much as AL4 has lagged behind AL3 so far (7.3 months on average) and add that to the graph (pinkish area). Let's also project a second scenario where the AL4→AL5 lag matches the AL3→AL4 lag at each level (dashed line). Then we get this:
AL5 projections built on AL4's own trend, with a constant 7.3-month lag (pinkish area) and a level-matched lag (dashed line), giving 80% automation in November 2027 (constant lag) to February 2028 (level-matched lag).
What if the AL4 curve instead looks more like AL3? We can project AL4 forward with the assumption that it is logistic and has the same slope as AL3 (k=6.9/yr). Then add AL5 projections: one with the same 7.3-month lag as before, and two dotted lines spanning a lag range of 6.3 to 8.5 months (roughly how much AL4 has lagged behind AL3 at different AL4 scores so far). Then we get this:
AL4’s slope is set to equal AL3’s slope, pulling AL5's 80% crossing forward to August 2027, assuming a 7.3-month lag between AL4 and AL5 (or July - September 2027 for a lag range of 6.3 to 8.5 months).
According to these projections, it seems plausible that AL5 reaches 80% between mid-2027 to early 2028[2]. Of course, the slope for AL5 could look completely different from AL3 and AL4. Perhaps the lag is much longer than 7.3 months (it probably isn’t much shorter, since AL5 is still at 0%.)
I expect the lag to be longer than 7.3 months, since the step change from AI “leads” to full automation seems quite large to me, and I expect the slope to be steeper for AL5 than AL4 due to recursive self-improvement.
These projections point to two things to watch over the coming months:
I hope Anthropic continues to publish regular updates for their R&D Automation Index so we can get answers to these questions relatively soon.
A few things these projections don’t consider:
Also, I haven’t double-checked all the extracted data, assumptions, and code produced by Claude. These graphs are for illustrating potential future trajectories, rather than get everything exactly right. The graphs look reasonably accurate to me, but there are probably some minor mistakes. Feel free to check Claude’s work here.
Thank you for reading! If you found value in this post, consider checking out my blog: Forecasting AI Futures.
Note that by “AL4 logistic curve”, I mean the curve separating the AL3 from AL4 areas, which actually corresponds to a projection of the combined scores of AL4 and AL5.
I’m focusing on the 80% threshold here because it’s a point where AI can automate most AI R&D, and because higher thresholds may simply not be reached if Anthropic decides to not permit some tasks to be completely automated (at least for some time). Check the graphs to see timelines for other thresholds.