I’ve been playing with a few myself, but certainly not a power user or that experienced as a coach.
Has anyone done any comparisons across all the many AI tools posted in this forum for us to play with. Anyone willing to share your reviews as each specific AI tool threads seems to be more trouble shooting or feedback vs reviews and comparisons.
Happy to pay and seems like most are similarly priced.
Not looking for a DC rainmaker review, what are most folks planning to stick with and why? Thanks
That seems to be a kind of review too…
The question is what you expect to gain from it. And the question is also whether you know what an LLM (“AI”) is.
An LLM is a language model, not a world model. It generates probabilities, not truths. It understands neither your situation nor your goals. It only recognizes statistical patterns in texts and generates what has frequently occurred in similar contexts. A coach, on the other hand, has to support decisions and weigh risks. An LLM only generates plausible sentences.They also often reflect what you want to hear. A real coach will contradict you if you get carried away. It definitely reinforces your tendencies, even if they are destructive. An LLM also has no memory across sessions. Used too many tokens, and you start from “the beginning.” An LLM can also tell you something wrong with absolute conviction without realizing it.
An LLM, on the other hand, can help you sort through your training data and notes and put them into a different textual form. It can summarize training logs or describe trends. If you don’t understand a concept related to training or nutrition, an LLM can explain it to you in simple language. It can also be motivating, thanks to its positive phrasing.
But it’s also often used to create workouts and plans. Can an LLM create workouts? Maybe. An entire training plan? Maybe, maybe not. An LLM does not know how to control progression. How much volume should it increase, how should intensity be distributed, periodization, regeneration, what do sensible workouts look like? Strict limits must be set for LLMs so that they do not generate nonsensical workouts. If you let the LLM decide, I’m not sure if it will always come up with the best workout choice.
Personally, I don’t see any benefit for myself.
If I feel tired because work has been too stressful, or I haven’t had enough sleep, or … then I skip the VO2max session and cycle at a relaxed pace, but I don’t need to ask a chatbot for that.
In the analysis, I look at HR, power, and time. I don’t need a 4-page analysis of a single activity (which always sounds the same anyway). Over longer periods of time, I use fitness charts, which give you an overview at a glance.
So, sorry, probably not what you wanted to read…
But I would still be very interested to know why others see advantages in it.
This just about 100% replicates what I feel/think about it.
We are not yet at a point where ‘AI’ really ‘thinks’, at least not the models available to the mass now.
I am conviced that it will come, but right now it is still in ‘learning’ mode.
@R2Tom and @MedTechCD, I’d push back a little, not because you’re wrong about what LLMs are doing under the hood, but because I think you’re overestimating what humans are doing differently.
At a fundamental level, I’d argue we’re doing essentially the same thing as an LLM. Our “context window” has just been built up, compressed, and refined thousands of times over the course of an entire lifetime of experience.
That said, I think current frontier models can probably outperform the average person on the majority of everyday tasks. Are they a replacement for an elite coach with decades of real-world experience? No, not yet. But that’s a pretty high bar.
What makes this an odd space to evaluate right now is that the underlying models are improving at an exponential rate with meaningful new releases coming out almost every month. On top of that, most “AI-powered” apps aren’t actually using the best or most capable models available, and many aren’t maximizing the context they provide to the model either. So a lot of what people experience as the ceiling of AI is really just a ceiling of a particular product’s implementation.
Ultimately, they’re tools and the value depends on how you use them. If you already have a coach, an LLM can be a fantastic filter for ideas and questions before your sessions, which significantly increases the value you get out of that time. If you don’t have a coach, they’re a pretty solid substitute. You just need to invest a little time in learning how to use them effectively.
Personally, I can see the appeal of a product that takes the guesswork out of all this and simplifies it down to “here are my last x days of data, program my next workout.” But in my opinion that’s actually one of the weaker ways to use these tools. The real value comes from the back and forth — providing rich context about your goals, your life, how you’re feeling — and treating it more like a conversation with a knowledgeable training partner than a vending machine you put data into.
To step away from any marketing pitch and just be brutally honest about the reality of what we are building here…
@R2Tom makes a fair point if we are talking about a raw, vanilla LLM just guessing the next word. But that is fundamentally misunderstanding where the technology is right now.
AI is no longer just “generating plausible sentences.” Through Agentic architectures, the LLM doesn’t invent workouts. It acts as an orchestrator that strictly respects established training methodologies, calls upon precise mathematical tools to analyze metrics with more exactness than a human ever could, and adapts plans based on hard data.
Let’s be realistic: any industry that relies on a computer screen is currently threatened or evolving. The sports coaching industry is no exception. A well-architected AI possesses vastly more raw, aggregate knowledge about sports science and physiology than any single human coach ever will in a lifetime. The challenge isn’t whether the AI “knows” enough; it’s learning how to extract and structure that knowledge correctly. I strongly believe that the professionals who will remain competitive are the coaches who integrate AI into their daily workflows right now. Do we really think top-tier coaches aren’t already using these models to extract specific knowledge and spot trends in their athletes data?
Regarding to the original question: why are the threads filled with troubleshooting instead of reviews?
Let me give you the developer’s perspective: genuine usage feedback is an incredibly “expensive” commodity. Human nature dictates that people only go to a forum when a bug happens or a sync fails. But when the app works perfectly they don’t come back to the forum to say, “Hey, I love how you did this!” They just grab their bike and go ride.
The same goes for feature suggestions. It’s rare for someone to spend their own time thinking about how a feature could be redesigned to be much better or how it could perfectly fit their specific use case. I would absolutely love for people to do that! In fact, I’ve tried to incorporate almost every small improvement or idea users have asked for so far.
What I desperately want for my tool is exactly what you are asking for: sincere reviews on whether this actually works for training.
I’m not looking for quick paying users right now. I am a software engineer, not a sports coach. I built this because I needed a way to get faster on my bike. It’s helping me improve, but I need real athletes to tell me if the training logic holds up in the real world for them too.
I don’t want to sound like a fake altruist, my long-term vision is absolutely to build a financially sustainable platform. But that only happens if the tool genuinely empowers athletes and coaches to improve. So, if anyone tries it, I don’t just want bug reports. Tell me if it’s actually making you a better cyclist, or tell me how a feature could be improved to fit your routine better.
I think they are filled with troubleshooting for a few reasons:
For a good example of this, take a look at Section11,
specifically the section11.md file. It’s literally dozens of pages with instructions for the LLM on how to respond and how not to respond.
I think the current tools that have active threads on intervals can be grouped in few types.
App style coaches
These have user friendly interfaces, with calendars, graphs showing your progress, etc. They use AI to analyze your training, create plans, etc. They also have a AI chat function allowing you to asked further questions, and make some changes to the plan.
I would put CoachWatts, IntervalCoach and AIEndurance in this category
Pure AI chats
These also can generate plans, workouts, analysis, etc. but do it much more in a chat format, with less focus on the UI. I would put Intervals.pro and LeCoach in this category
I have tried or am still trying the ones I listed above.
When choosing one, ask yourself what are you most comfortable with: an AI prompt, or a visually well designed app
I personally gain a lot of insights from graphs and well organized metrics. So for me option 2 works better
Of those I picked IntervalCoach as my choice for now. AIEndurance interface was mediocre, with the calendar function not working very well. CoachWatts is maybe the best solution at the moment, but for me it felt like overkill with the detailed food advice etc. Also it was priced higher than IntervalCoach.
IntercalCoach had the right mix of functionality and simplicity. It generates plans based on your goals and availability. Provides feedback on your activities, and adjusts workouts based on your readiness. One thing I haven’t seen much of adjusting plan based on progress.
Most coaches have a free trial period, but you may have to spend a number of days to really understand whether it is for you
You forgot mine planwatts 
I might be late to the party. I started developing before joining that forum… even so, I strive to make integration my differentiating factor and ensure a truly user-friendly chat experience. If you haven’t tried it, I encourage you to do so
One more lens worth considering when evaluating these tools: the economics behind them.
If a tool is “free” without requiring you to bring your own LLM API key, they’re either planning to monetize you later, or cutting costs with a cheaper, less capable model. Neither is great for getting quality coaching output.
On the flip side, subscriptions of $10+/month sound reasonable until you look at what frontier model time actually costs. At current Claude Opus pricing, $10 buys you roughly 1M input tokens and ~200K output tokens, which is an enormous amount of coaching interaction. The developer’s time and infrastructure add cost on top, sure, but honestly with the agentic coding tools these days it really is extremely easy to develop and maintain these tools.
That gap is where the real business model lives: companies are financially incentivized to minimize how many tokens they feed the model and use the cheapest model that still seems “good enough.” More context means better coaching, but it also eats into their margin.
Personally, that’s not the trade-off I want to make. I’d rather use the best frontier model available and give it as much of my data as possible, full activity history, HRV trends, sleep, whatever I can feed it. That’s where the real value of AI coaching lies, and it’s something the subscription tools are structurally disincentivized to offer.
AI tools for Intervals.icu are either app-style (graphs, visual insights) or chat-based (flexible, less UI). IntervalCoach balances simplicity and function. AI can summarize data and suggest plans but can’t fully replace a real coach.
I have no idea about LLM or API and was, and still am, amazed by what Intervals can do with my workouts. Even after years, I still don’t understand a lot of it, but it’s fun to see how my numbers and graphs change after a workout. I’ve tried most AI apps and suddenly have the ability to modify my training through specific instructions and questions. What should I change about my diet (eat less ;)), should I train tomorrow, sweet spot or threshold intervals, etc.Do I always stick to it? Of course not! Does it always work? Of course not! I ride my road bike for fun and love analyzing my data afterward (which sometimes takes longer than the ride itself) and comparing it with my friends. Currently, PlanWatts helps me with my analysis and performs calculations for me that would otherwise take me time; the program speeds things up. Comparing data between different units and drawing conclusions from that makes the program faster.Is it always accurate? Of course not! Plan Watts is still free, but most of the other tools aren’t anymore. Suddenly I’m one token short (whatever that is) and I’m supposed to wait 24 hours for an answer to my question, or the program suddenly switches to other issues (can’t find values, it’s due to intervals, data not found, and lack of memory, etc. - not my cup of tea). It’s important to know that these programs don’t replace a coach, and that not everything is correct. But if, in addition to the fun of the analysis, you also get a good tip or two - why not? Since I have absolutely no idea what such a tool costs.I can’t say that this is too expensive and that isn’t. For example, I find Intervals’ payment system very good because it’s FULLY functional even in the free version, which isn’t the case with some of these AI tools. What they all have in common is that the developers take very good care of their “baby.” So, I’m done.(Sorry, I’m translating with Google.)
The good enough comment is always an interesting one. it always comes done to the 80-20 rule.
Paying a lower amount to get 80% of the possible solution is in fact good enough for many people. But some athletes may want to spring for that extra 20%, usually at a significant cost increase
The interesting part about all the apps currently in development is that they run the gamut from cheap (or free) but limited to premium-priced and rich in functionality. As to what will be acceptable to the market? 
I believe ambitious amateur athletes who prepare for one or more events a year need a coach or a lot of past experience. I don’t think AI is sufficient for that. If I say I train, every coach will probably tell me, “You just ride a bike.” AI isn’t that cheeky. I learn a bit about training methods, nutrition, and measurement techniques, and I can still ask silly questions at 3 a.m. In short, 80% is perfectly adequate for me. If I were 30 years younger, I would either work with a coach or pay 100% for an AI tailored to my specific needs.
The 80/20 principle is exactly my point.
All these tools charge a fixed subscription and pay for token usage themselves, so they are all heavily incentivized to hit that “good enough” threshold, and this is true regardless of what they charge. The economics push every one of them toward the same outcome.
What frustrates me is that if you actually want that extra 20%, you’re more likely to get it from a free ChatGPT or Claude subscription than from any of these dedicated tools. You just need to learn how to prompt them.
All these apps lack proper product management. With AI making it so easy to ship changes fast, it feels like new features get added before the last one was even tested. I came to this topic with an open mind about a month ago, genuinely trying not to reinvent the wheel — and I was pretty disappointed.
So why would I pay for any of them when:
a) They’re unstable and the UI is unintuitive.
b) The chat interactions are worse than if I just fed the same data to ChatGPT or Claude manually.
The way I see it, an AI tool specifically for training should follow the same philosophy as Claude Code or OpenAI Codex for programming — a framework that lets the AI read and focus on the relevant information, then execute actions on your source (which in our case is the training plan). Layer on some options to configure how it interacts with your data, keep the UI simple, and that’s it.
I don’t need a separate panel for nutrition — I can ask the AI in the chat window before a ride. I don’t need a million charts and AI hints scattered everywhere. Charts will always be better on intervals.icu, and AI hints placed in random spots are mostly useless.
What I need is an expert assistant that has access to my data and doesn’t hallucinate because its context is overloaded. That’s the hard engineering problem — the one tools like Claude Code have actually solved well.
I wanted to use something off the shelf, but nothing cut it — so I built my own. It cost me more than all these subscriptions combined, but I managed to fine tune it to the point that the interactions with AI are more meaningful than just with Claude or ChatGPT directly.
Interesting fact: not always the biggest model is the one that answers best, I guess it depends on the context window size, which is sometimes bigger in faster, non-thinking models.
I’ve been thinking about the fair pricing model for AI coaching tools for a while, and I’ve come to the conclusion that only two models make sense (and they can coexist in the same app):
BYOK (Bring Your Own Key): You manage your own AI usage, but there’s still infrastructure cost — database, servers, sync — so you’d pay the developer a small monthly fee, comparable to or less than a Strava subscription.
Transparent token pricing: You pay for actual usage, with infrastructure and operations cost baked in so the developer doesn’t run at a loss.
Both models solve the same problem you’re describing: they remove the incentive to quietly downgrade the model or starve it of context to protect margins.
Interesting thread and inputs from you and few others. I’m convinced that with the appropriate data as context, AI can outperform majority of coaches. I’ve been using ChatGPT for months and I found it very nice for “self reflection” practice. Now I’m trying to step up and I think I’m ready for go with a subscription to feed some data and have better outputs. What models and AI platforms are the best for endurance athletes?
If you haven’t tried Claude I would say give it a shot personally out of the big three (ChatGPT, Claude, and Gemini) Claude is my favorite currently as I find its writing style strikes a better balance between confident but not too confident, formal but not corporate.
Honestly, I have a hard time seeing any of these AI coaching apps surviving long-term. One or more of the following will make them obsolete within a few years:
Until then, I struggle to pay for what amounts to a slightly more polished copy-paste workflow, one that actually gives me less transparency and control than just prompting directly.