Somebody Already Ran This Experiment

Most comparisons of this kind are arguments. This one has a result, produced by someone with no stake in it.

One thing worth stating before any of it: OpenAI does not describe ChatGPT as a training platform, and that is not a technicality. Every other product on this site calls itself a training tool of some kind. ChatGPT is a general-purpose language model being used as one, and most of what follows comes from that gap rather than from any failing of the model.

In 2026 Cycling Weekly ran a six-week test, written up by Zach Nehr, an elite rider and qualified coach. Two cyclists trained for six weeks. One followed a plan generated entirely by ChatGPT, with no coach, no check-ins and no adjustments along the way. The other worked with Ulisses Abbud, a Miami-based former professional, using Vekta as the analytical layer behind every decision. Nehr evaluated both plans blind at the end.

His verdict: ChatGPT 6 out of 10. Vekta and a human coach, 9 out of 10.

Six out of ten is a pass, and it should be read as one. Jeff Herman, the rider on the ChatGPT plan, went from 200 to 227 watts, an improvement of 13.5 percent. He got fitter. The plan worked.

It is also worth stating the caveat that makes the comparison honest rather than convenient. Jeff started at 200 watts. Ben Binet, on the coached plan, started at 389 and finished at 397 — only 2.1 percent. As Ulisses put it, a rider at 200 watts has far more room for rapid improvement than one already training at a high level, where even a small gain in absolute power is much harder to achieve. Vekta measured Ben’s Critical Power rising 13 percent over the same block, and he lost around 1.5 kg, so the raw FTP number understates what happened. But nobody should read those two percentages side by side and conclude anything simple.

The interesting part of the experiment was never the wattage. It was what each rider experienced getting there.

What ChatGPT Can Actually Do Now

Any comparison written a year ago would have said ChatGPT cannot see your training data. That is no longer true, and pretending otherwise would be the easiest way to lose a reader who has already tried it.

Third-party connectors now exist that give ChatGPT read access to more than twenty sources — Garmin, WHOOP, Oura, Withings, Apple Health and others — using OAuth so no password is shared. Through those it can reach heart rate, power, cadence, speed, distance, elevation, lap splits and GPS, and for activities synced from August 2026 onward, Training Effect and time in heart-rate zones taken straight from the device recording. It can see recovery data too.

It also remembers. ChatGPT’s memory carries facts between conversations, updating on its own as you talk to it, and project-level memory keeps one context separate from another. And it can run code: upload a power file and it will spin up a Python environment with pandas and SciPy, fit a curve and hand you both the answer and the code that produced it.

Those connectors are not free — they run at roughly nine to ten dollars a month on top of whatever ChatGPT plan you are on — but they work, and they are good.

So the honest question is not whether ChatGPT can see your data. It is what happens to the data once it arrives.

Where The Data Stops

There is one hard wall worth knowing about before you plan around any of this.

On 1 June 2026, Strava rewrote its API policy. It now prohibits using Strava data, in its own words, “directly or indirectly, in connection with the development, training, evaluation, or operation of any AI Application” — and it names the mechanisms explicitly, including retrieval-augmented generation and ingestion into context windows. It also prohibits anyone from operating any MCP server or agent-mediated interface that exposes Strava data. Strava launched its own connector exclusively with Anthropic.

In practice that means Strava data cannot legally reach ChatGPT, and the route is closed rather than merely missing.

Vekta does not integrate with Strava either, though for a different reason: aggregators process and smooth FIT data before passing it downstream, and Vekta reads raw from Garmin, Wahoo, Hammerhead, COROS, Zwift and Stryd to preserve the power, cadence and torque precision the model depends on. The outcome is the same on that one point. The difference is that Vekta reaches the original file by another road, and ChatGPT does not.

It Can Compute A Model. It Will Not Maintain One.

This is the real distinction, and it is narrower and more durable than the one most comparisons reach for.

Give ChatGPT a power file and ask it for your Critical Power, and it can genuinely work it out. It will write the Python, fit the curve and show you the code. That is not a trick and it is not nothing.

Now ask what happens next. Tomorrow you ride again. Nobody uploads anything, so nothing is recalculated. Your zones do not move, because they were text in a conversation rather than values in a system. Next week’s intervals are not reclassified against a changed threshold, because nothing classified them in the first place. The carbohydrate target on Saturday’s session does not shift, because there is no session and no target.

It can produce the number. It cannot be the thing that keeps the number true.

In Vekta, Critical Power, W′, D′ and Pmax are recomputed from training as it arrives, every day, without anybody asking. Cycling and running zones regenerate from them. Interval classification recalibrates against the new threshold. Durability is measured against current capacity rather than last month’s. The fuelling target follows. Nothing had to be triggered, and nothing was waiting on somebody remembering to trigger it.

The same applies to continuity, and here it is worth being precise, because ChatGPT’s memory is better than people assume. It carries facts between conversations, updates itself as you talk, and keeps projects separate. What it stores are things you told it. What it does not store is a continuously recalculated model of five thousand files. Load, Strain, Volume and Intensity in Vekta are quantities that accumulate across months and across sports; the question is this athlete absorbing what I am giving them is answerable because the record was never re-described, only extended.

And there is a hard limit underneath all of it. These connectors are read-only by design — in the operators’ own words, it reads, it does not write, and it does not post activities or change settings in your apps. ChatGPT can look at everything and touch nothing. It cannot put Saturday’s session on your calendar or on your head unit, and no connector is going to let it, because that is not what they are for.

You Are The Thing That Has To Remember

Everything in the section above is true, and it is also more work than it sounds. That deserves saying out loud, because the gap between ChatGPT can do this and this is how I train is where most people quietly give up.

Count what actually has to happen for a language model to tell you your Critical Power. You need to know that a critical power model is the thing worth computing in the first place. You need to get the right file out of the right place in the right format. You need to know which model you want fitted, over which durations, and from which sessions — ask vaguely and you will get a number produced from whatever it decided to use. Then you need enough grounding to judge whether the answer is plausible, because it will be delivered with equal confidence either way.

Then you need to do all of it again next week, because nothing about last week’s answer makes this week’s appear.

Memory helps, and it is genuinely good, but it remembers what you told it rather than doing the work again. A stored fact that your CP is 300 watts is not a model. It is a number that was true in August, sitting in a context window in October, being used to answer a question about November.

Now multiply. A coach with thirty athletes does not have a workflow here, they have a second job: thirty exports, thirty conversations, thirty judgements about whether the output is sane, and thirty sets of results to carry somewhere else by hand, because the connectors are read-only and nothing can be written back. Repeat weekly, forever.

In Vekta none of those steps exist. The model was recomputed before anybody opened a laptop. The zones already moved. The intervals in yesterday’s session were already classified against the new threshold, the summary was already written and delivered, and the carbohydrate target on Saturday already reflects it. Nobody exported anything, nobody phrased a prompt, and nobody had to know that any of it was due.

The question is not whether a language model can do the analysis. It is whether you want to be the one operating it, remembering it, checking it and carrying the answers around.

For one curious athlete, once, it is genuinely worth doing. You will learn something about your own physiology and about how these models work, and that is not a small thing. As a way of running a season, or a roster, it is a lot of hours to spend arriving where a platform starts.

The Plan Is The Easy Part

It is worth being clear about what ChatGPT is genuinely good at here, because it is more than people expect.

Generating a structured training plan is not hard, and ChatGPT does it well. It knows the literature. It can build a periodised block, sequence intensity sensibly, explain why a VO2max session sits where it does, and adjust the whole thing to a stated number of hours. For an athlete with no plan at all, this is a substantial improvement on nothing, and Jeff’s 13.5 percent is the proof.

Writing the plan was never where coaching lives, though. The plan is the opening position. Coaching is what happens to it over six weeks of an actual life — the week someone is ill, the block where the numbers say fine and the athlete says otherwise, the race that moves, the session that has to be cut in half because a meeting overran.

ChatGPT will rewrite a plan whenever you ask. What it cannot do is notice that it should.

It Cannot See You

The Cycling Weekly experiment is at its most useful here, because it exposes something no feature table would.

Jeff’s ChatGPT plan gave him three to five hard sessions a week, week after week, with no structured recovery beyond a Monday. It said push harder, every week, because nothing told it not to.

It is fair to note that a connector could now show it his sleep and his HRV. That is a real change and it matters. What has not changed is that nothing would prompt it to look. A chat window is a thing you open when you have a question, and the athlete who is three weeks into a hole is precisely the athlete who has stopped asking.

He got through it, helped by six weeks of unusually clear calendar. By the end he described feeling “a little frayed”, reflected that the plan had “emphasised short-term FTP increases”, and said he was “unsure whether I’d be able to follow the plan long-term”. Nehr recorded that both riders skirted the edges of mental burnout during the block.

An algorithm can crunch power, heart rate and sleep data all day. It cannot ask how you felt getting out of bed this morning, and it cannot tell when the honest answer to that question matters more than the numbers.

Ben’s experience differed from the first session. Ulisses prescribed around who Ben actually was — an 80 kg rider near 400 watts, getting married and changing jobs, needing precision rather than aggression. Ben said afterwards that the exact duration and wattage targets helped him progress, and that Ulisses explaining the idea behind the programme mattered. That second half is the part software of any kind struggles with. A model can generate the plan. It cannot give an athlete the confidence to execute it.

It Will Not Push Back

This is the failure mode that matters most and gets discussed least.

A language model is agreeable by construction. Tell it you feel good and want more, and you will get more. Tell it you want to hit a number in six weeks and it will build the block that targets the number, because that is what you asked for. The Cycling Weekly write-up lists a tendency toward overtraining among the ChatGPT plan’s weaknesses, alongside zero human interaction and limited feedback.

A coach is the opposite. The value of a coach is frequently that they say no — not today, not that hard, not this week. That is not a data problem and no amount of context window solves it.

Nehr’s conclusion was that AI is currently best used as a tool rather than a boss, and that the ultimate formula is the raw data-crunching power of the algorithm steered by the intuition of a human coach.

That is precisely the model Vekta is built on, and it is worth being explicit that this page is not an argument against AI in coaching. Vekta is full of it.

What Vekta’s AI Actually Does

The difference is not whether there is a model involved. It is what the model is pointed at.

Vekta’s machine learning reads every session automatically, detecting efforts by duration, power relative to that athlete’s Critical Power, cadence and torque simultaneously, then classifying each one by intensity type against their own physiology — Neuromuscular above 180 percent of CP, Anaerobic between 130 and 180, VO2max between 105 and 130. It assigns a training stimulus to the session, detects and classifies races, finds the most similar past sessions through a proprietary similarity score, and writes a summary delivered to coach and athlete without being asked.

None of that generates a plan or makes a decision. It removes the file-reading, the pattern-spotting and the recalculation that used to consume a coach’s week, and hands the coach a clear picture to decide from.

The AI does the structural work. The coach does the coaching. In a chat window, the AI is asked to do both, and only one of them is a data problem.

Where ChatGPT Is Genuinely Better

There are real things here and they should not be grudging.

It is the best explainer in the sport. Ask what W′ actually represents, why decoupling matters, what the difference between aerobic and anaerobic threshold really is, and you will get a patient, accurate answer pitched at whatever level you ask for. Vekta has support articles. ChatGPT has a conversation, and it will keep going until you understand.

It is free to start, and instant. No account to configure, no device to connect, no trial to remember to cancel.

It is excellent for the writing around coaching — drafting a message to an athlete, turning a training philosophy into something readable, preparing for a difficult conversation. Coaches use it for this constantly and should.

And for an athlete with no coach, no structure and no budget, a ChatGPT plan is better than the alternative, which is usually riding hard on Tuesdays because it feels like the thing to do. If you want that structure from something purpose-built rather than improvised, TrainerRoad is the product that does it best. Six out of ten, from a qualified coach evaluating blind, is a real result.

The Cost Of Free

ChatGPT’s free tier is genuinely free, and for asking questions it is all most people need. The arithmetic changes once you want it working from your actual data.

The paid tiers run at $8 a month for Go, $20 for Plus, and $100 or $200 for Pro. On top of that, reaching your Garmin data means a third-party connector at roughly nine to ten dollars a month. So an athlete who wants ChatGPT reading their training rather than guessing at it is somewhere around $30 a month, and still has no model, no season history and nothing pushing sessions to their head unit.

A Vekta athlete seat is €19.99 a month or €179.99 a year, with a 14-day trial, and a coach account is free at any roster size.

None of which is the point. If ChatGPT is doing what you need, its price is irrelevant and so is ours.

The comparison only becomes interesting once you want the thing it is not built to do, and at that point you are paying two subscriptions to approximate one.

It Does Not Know What It Does Not Know

This is the quietest risk and the one most likely to catch an athlete out.

A language model answers with the same confidence whether it is right, wrong, or working from something you told it four hundred messages ago and it has since compressed away. In most contexts that is an inconvenience. In training it means an athlete acting for six weeks on a number that was never checked.

Endurance science is not settled, either. There are live disagreements about threshold definitions, about how load should be quantified, about durability and what it means. A model trained on all of it will state a position fluently without signalling that a different one exists, and an athlete without the background to know that has no way to tell.

Vekta is not immune to being wrong, and no platform is. The difference is that it is auditable. Critical Power comes from specific sessions you can open. An interval is classified against a threshold you can see. Durability is measured at stated fatigue levels. Every number traces back to a file, and a coach who disagrees can go and look at why.

A number you can interrogate is worth more than a number delivered confidently, and this is true of us as much as of anyone else.

For Athletes

If you are training without a coach and want structure, use ChatGPT. We mean that. It costs nothing to start, it will build you something sensible, and it will explain the reasoning as many times as you need.

What you should know going in is what it is not doing. It is not modelling your physiology, so the zones it gives you are the ones you gave it. It is not tracking your load across the season, so it cannot tell you that you are three weeks into a hole. It does not know you are tired unless you tell it, and it will not tell you to stop.

Vekta makes sense for an athlete who wants the model rather than the plan: Critical Power and W′ computed from your own training and updated daily without a test, durability measured in every session, a fuelling target on every workout, and a written summary after every ride. An athlete seat is €19.99 a month or €179.99 a year, with a 14-day trial.

If you want both, that is the honest answer. Keep asking ChatGPT to explain things. Let something else do the modelling.

For Coaches

No coach is losing their athletes to ChatGPT, and anyone selling you that fear is selling something.

What is worth paying attention to is the athlete who has never had a coach and now has a plan that is 60 percent as good as one, for free. That is a bigger market change than it looks, and the answer to it is not to argue that AI is useless. Cycling Weekly gave it six out of ten and they were right to.

The answer is the other 40 percent, and the experiment describes exactly what it consists of: reading fatigue, knowing the athlete’s life, adjusting when it changes, explaining the reasoning, and being willing to say no. None of that is a data problem, which is why software cannot take it from you.

What software can take from you is the hundred files a week. Vekta reads all of them, classifies the intervals, flags the races, writes the summaries and surfaces the athletes who need you, so the hours go to the 40 percent instead. A coach account is free at any roster size.

Who Should Use ChatGPT

Anyone learning. It is the most patient teacher in endurance sport and it costs nothing.

Anyone with no plan and no budget, who needs a structure to follow this month rather than a philosophy to adopt this year.

Any coach drafting communication, thinking out loud, or wanting a second read on an idea before taking it to an athlete.

Where it stops being the right tool is the point at which the answer depends on a model of you specifically, maintained over time, and on somebody noticing what you did not say. That is not a limitation anyone is going to engineer away soon, because it is not really a limitation of the software.

Still Weighing It Up?

If you are not comparing us to ChatGPT so much as working out what to use instead of it, we have written that up separately, with pricing for every option: ChatGPT training plan alternatives.