I could image the AI “thinking” that the added training stress / stimulus would be needed (even if not logic by classic principles at first sight). And that’s not the anti-deterministic part. Anti-deterministic is the prediction changing by adding the 30min workout and then not changing back when removing the 30min workout. I can only assume some other changes not mentioned here or a timing issue.
Short answer: accept the new AIFTP. This is a totally new system with a new training benchmark mark (AIFTP). Don’t compare it to any of your old number. Try it for a month and see how it goes. It works!
Do you recommend we use the new TR FTP for everything? It’s kinda strange to have different ones for Garmin, Interval. Icu, TR, etc…
Not sure @Nate_Pearson if you could provide a graphic like this showing how the model weights the past and the future over a prediction window. It might clear up many of the questions. I don’t think the last chart is right, my assumption is that the predicted value doesn’t just look at the prediction window. But past workouts become less meaningful over time.
I don’t use Garmin, intervals.icu, TrainingPeaks etc. so do have any experience to help you out. The TR metric won’t line up exactly to others I would guess. If it’s within a few percentage points, probably easiest to keep the same number. If it is more different than that, maybe try to figure out a relationship between to two (scaling factor) and adjust based on that? Good luck!
I don’t have any real world use for differentiating my “real FTP” from my “TR FTP”, so I changed them all.
I do outdoor workouts when I can, but I’m doing TR workouts, so I just have TR send the workout to my Garmin and do the watts TR prescribed.
If I was using FTP for any reason outside the TR ecosystem, like, for example to pace a 40K TT, I would do whatever my preferred FTP test is and use the result for that, not the TR FTP.
Nate writes:
“What is the new AI FTP Detection?
The updated AI FTP Detection you’re getting puts you around a level 3 threshold level. This will be consistent between detections, and we think this is a good insert point into your overall zone progressions to get faster.”
If I interpret this correctly, it seems to say that AI “FTP” will be recalibrated every 30 days back to a level 3 threshold, meaning that the threshold level will not progress month over month. It seems that threshold workouts will be very limited to those around level 3. It also means that AI “FTP” cannot be accurately compared month to month.
In more detail, we are all now seeing the “FTP recalibration” happen at the beginning of our first 30 days. For some it goes up, for some it goes down, for some it is much the same. It will be consistent through a 30 day prediction month. Then on the 30th day it will be recalibrated again back to a level 3 threshold. Your February AI “FTP” will not be consistent with your January AI “FTP”, your March AI ”FTP” will not be consistent with your February AI “FTP” and so on for future months. They are all on subtly different scales, and you should not put any meaning on the change in number from month to month, only the gain/loss within a prediction month.
With this continual recalibration how will we keep track of our month-on-month or year-on-year progress? This seems to obfuscate the long-term gains from consistent training, while focusing on month-by-month incremental progress that gets “wiped out” (numerically) over and over again. For many of us FTP gains are incremental, and these could easily be hidden in the recalibration noise. Also, will AI “FTP” diverge more wildly from actual FTP over time?
I get that AI “FTP” is just a number used to select workouts … but we all want a metric that shows improvement over time. In short, it seems like I have to do my own real FTP testing to track where I really am.
TR clearly has an AI algorithm for accurately estimating true FTP based on a activity history - it’s the AI FTP feature that they’ve deployed all through 2025 and before. I think this should be brought back so that we can track real FTP over time. Of course, this means they would have to fix the terminology to differentiate the two “FTP” metrics.
Finally, this seems likely to create a deluge of “what happened to my FTP number” posts every single month into the future …
I agree with you that it’s going to create a ton of confusion
Having said that, I’ve come to realize that FTP is meaningless to me outside using it to plan workouts. Actual completed power by time values are far more valuable to me as a method of tracking my improvements, and TR, Intervals, etc. give me access to those numbers already.
I dont know what i should think about it. My ai ftp is too high. I had never a w/kg ratio that high. 264 ftp, a 10 Minute intervall with 259 Watt bought my heartrate into vo2max territory, of course i couldnt finish a treshold workout with this heartrate. I reduced my ftp manually
I read this post and about 60 of the replies and am still confused. You’re predicting my FTP, how do I test if it is accurate? Previously, I liked to do the ramp tests, but you say you aren’t predicting ramp test. What duration should I be able to sustain this “FTP” power? You mention “level 3 threshold”, but I don’t really understand what that means with regard to your model and estimates.
I agree with other commenter that predicting power duration curve like Stryd would be easier to reason about. This also makes it clear when the model is estimating something (very short sprint, very long endurance) that you have no data on because I haven’t completed hard workouts at those durations.
Also, can you do that thing where you review all my training history and comment on it? I have been doing TR for over a year and felt like I pretty quickly hit plateau when I should have been seeing newb gains. I have been focusing on my nutrition more and TR is predicting I improved, but I would like to do some test to verify…
I guess I don’t know why people up in arms over the predicted number just don’t do a test. Plenty of races in Zwift you can go all out for 1 hour. You can even just TT up Alpe du Zwift. Best measure of performance is performance.
Yeah. A lot of people felt like the old AI FTP was accurate, so I think a lot of the complaints are based on not wanting to have to do a test or having a system across multiple devices/apps that this creates a headache with. I get that when you felt like tests were a thing of the past it sucks to be told “TR is taking away something you had and loved and now you can fix it by doing a test”. No one likes to have functionality removed. The Garmin and Strava reddits certainly prove that out!
I never found the old AI FTP very accurate, so I always tested, so it has pretty much zero impact on me.
They haven’t said it much, but in the FAQ podcast Jonathon mentioned that they won’t change your AI FTP setting in specialty, so then presumably the AI is focused on increasing TTE / Athlete Levels
This is an interesting chart. ~80% of TR users are threshold level ~3.5 or lower.
I looked in the workout library for a threshold level ~3.5 workout. The most representative one I found was Gayley:
- Threshold level 3.6
- 4x8 minute intervals at FTP
- 4 minute recovery between intervals
- Workouts - TrainerRoad
If I’m interpreting correctly, implicitly, the TrainerRoad AI logic is assuming that a workout like Gayley is a proxy for FTP.
The “usual” definition for FTP is the power you can hold for 40-70 minutes. Gaylord in total has 32 mins at FTP, with breaks, so below the lower end of this range. This is a generous assessment of FTP.
Given that the TrainerRoad AI logic is landing on 70% of users showing a decrease in FTP (where FTP = something like Gayley), does this mean that a lot of users had inflated FTPs to begin with? Surely that can’t be the case 
As another data point, Xert won’t ever raise your FTP unless you have a “breakthrough”, in fact by default it will decay your FTP unless you turn that off.
People should go back and watch the GCN Coggan/FTP videos. I find it kind of humorous that they picked the FTP number because of the popularity of 40k TTs and that was the data they had (can confirm they were wildly popular in the 90s/early 2000s…you could almost find a race every week during spring/summer).
Power curves are great and have a lot of uses.
Of course, there is so much more to fitness and training than FTP. But, right or wrong, people love having a single performance score that they can compare and track. And the strong marketing message has been “FTP improvement” and “AI FTP detection” for years. I don’t see the world switching over to “FTP is dead”.
Interesting deeper discussion from Nate and Jonathan, 9 months ago.
+1000. I am skeptical about this model to serve cyclist interested in longer formats well.
To figure out your pacing, you can go to the website under Career > Power Records. Scroll down past the first graph to Power Records and look at the table shown there. It gives your 20 and 60 minute powers.
You can use the 60 minute number or take 95% of the 20 minute number (whichever is higher) as FTP to pace against. It’s not the best approach, but it’s basically giving you the same number most approaches give you.
This approach is only accurate if you have a recent enough all-out effort at 20 minute or 60 minute duration. Meaning that you were well rested, well prepared, warmed up properly, went all-out with great pacing, had no interruptions, and were completely spent at the 20 minute or 60 minute mark. Basically, this is essentially the same as saying take a 20 minute or 60 minute FTP test. 
It’s not certainly not perfect. But even if you do a 8 or 20 minute test perfectly, the percentage that you multiply by isn’t a number that fits everyone anyways. It’s also why I said take the higher of the two values, since during interval workouts you probably didn’t do a 60 minute max effort, but you probably have done a 20 minute one, especially in specialty.
And unless you’re doing a draft-illegal TT on a flat course, you’re not going to just sit at X watts anyways. You’re going to go all out at certain points to get up the hill, stay with the group, get positioned well during the mass start, chase a breakaway, etc.
If you really feel strongly about having the most accurate FTP then you should figure out what method works best for you and manually do it rather than having trusted the TR AI estimate. If you trusted the TR AI FTP before, then you should trust it even more now, since TR says it’s even better.