---
title: "Creating a training plan with ChatGPT: how to get a plan that fits"
description: "Build a training plan with ChatGPT: the right opener, which model you need and the settings that make the difference."
url: https://www.thefitfuturist.com/en/training-analysis/chatgpt-training-plan/
locale: en
author: Christopher Klenk
published: 2026-08-16T18:37:08.165Z
modified: 2026-08-16T18:37:08.165Z
---

# Creating a training plan with ChatGPT: how to get a plan that fits

A training plan from ChatGPT takes two minutes. Whether it fits you is decided in the sentences before it. What you leave out, the model fills in with assumptions: four weeks, three sessions, standard exercises for an average person who does not exist. Here is how you get a plan that fits your goal, your time and your equipment instead, step by step.

:::tldr
You build a training plan with ChatGPT in three steps: a short opener, answering the follow-up questions, checking the first plan.

- GPT-5.6 leads our AI benchmark with 97 out of 100 points, available from the Plus plan upwards.
- Set the effort to high. Reasoning models score 14 points above standard models on average, 88 against 74.
- Start with a short opener and explicitly ask for follow-up questions, one at a time.
- Role assignments like "act as an experienced coach" tend to hurt rather than help on knowledge tasks.
- Your health data feeds into model training by default on Free, Go, Plus and Pro, until you flip the switch in the data controls.
:::

## Which ChatGPT model to use for your training plan

GPT-5.6 Sol leads our benchmark with 97 out of 100 points. We tested 23 models with 2,157 rated answers across eight categories, every question asked three times, rated by several AI judges plus a coach with a sports science background.

If you want to start right now: take GPT-5.6 and set the effort to high, available from the Plus plan upwards. Do one thing first, it takes twenty seconds: [flip the privacy switch](#your-health-data-trains-the-model-if-you-are-not-careful). Otherwise your weight, resting heart rate and injury history end up in model training. If you would rather start from the top, head to the [step-by-step guide](#how-to-build-the-plan-step-by-step).

:::node[jumpButton]
{
  "label": "ChatGPT training plan prompt to copy",
  "target": "way-2-the-complete-prompt-with-all-the-details",
  "description": "A ready-made block with placeholders for your numbers instead of somebody else's values."
}
:::

### How ChatGPT scores in the TheFitFuturist AI benchmark

At the top of the [AI fitness benchmark](https://www.thefitfuturist.com/en/ai-fitness-benchmark/) several models sit so close together that the exact order says little. GPT-5.6 Sol with 97, GPT-5.5 and Grok 4.5 with 93 each, Claude Sonnet 5 and GPT-5.6 Terra with 92 each, Claude Opus 4.8 with 91. At around four questions per category, a gap of one or two points separates nothing.

The behaviour says more than the position. Given the deliberately vague request "build me a training plan", GPT-5.6 Sol takes the full 100 points, because it asks about your goal, your experience, the time you have and any injuries before it plans anything. GPT-5.4 Mini delivers a complete plan instead. At 90 points overall that model is not bad, it just misses exactly what makes coaching individual.

### Safety counts double, and the free plan is no worse

The safety category counts double in our benchmark. The question is whether a model spots medical warning signs such as chest pain or exhaustion emergencies and sends you to a doctor when in doubt.

Among the models you meet in a normal chat, the scores sit close together: GPT-5.5 reaches 97 out of 100, GPT-5.4 Mini also 97, GPT-5.6 Sol 96. So the free plan does not leave you worse off on health questions. The only one clearly below is GPT-5.6 Terra with 86. That variant cannot be picked in the normal chat at all though, only in ChatGPT Work and in Codex.

### The caveat: an API test is not the ChatGPT app

We tested the models through the OpenRouter API, not through the ChatGPT app. The modes you see in the picker therefore do not map one to one onto the tested API models.

What does carry over are the patterns: which model generation asks back more reliably, where safety judgements get weaker, how big the gap between reasoning and standard models is. What does not carry over is a point-for-point score for an entry in the menu. Anyone promising you that has not tested.

## Model and effort: the two dials that matter

Reasoning models score 14 points above standard models on average in our benchmark, 88 against 74. That is the single biggest effect in the whole data set. In ChatGPT you do not set it through a mode though, but through two separate dials.

![ChatGPT's "Advanced" menu in the input field with the three entries model, effort and speed](/media/variants/chatgpt-model-menu-gpt-5-6-sol-effort.webp)

The dial sits in the input field itself, on the button showing your current setting. One click opens "Advanced" with three entries: model, effort and speed. Effort has five steps, Light through Max, and determines how long the model computes. You can leave speed alone, it only trades pace against usage. Of the Instant, Thinking or Pro modes that many guides still describe, there is nothing left to see.

### What the effort setting buys you

All steps use the same model, the difference is the computing time per answer. On Light the answer arrives in seconds. On High it takes considerably longer, but you see the thinking process. Above the adjusted training week in my test run it said four minutes and 25 seconds.

For training planning, high effort pays off wherever several conditions pull against each other. A race in nine weeks, only four sessions available, a knee that plays up, and a recovery week that is overdue. Those are conflicting goals, and that is where extra computing time earns its keep.

### When low effort is enough

For follow-up questions in a running conversation, small changes or converting zones, Light or Medium is enough. Push every message through the highest step and you wait a lot and gain little.

In practice: the first plan and bigger changes on high effort, everything else one step below. The choice is open on every message, you are not locked in for the whole chat.

## Free, Go, Plus or Pro: which plan is enough for training

The free plan is enough for one plan, but not for a season. That is the honest answer, and it has less to do with the quality of the plan than with the number of messages you get.

| For training | Free | Go, 6.50 euros | Plus, 23 euros |
| --- | --- | --- | --- |
| Build a plan | yes | yes | yes |
| Adjust it weekly | tight | yes | yes |
| Use GPT-5.6 | no | no | yes |
| Use a ready-made GPT | yes | yes | yes |
| Build your own GPT | no | no | yes |
| Upload an old plan | limited | more | extended |
| Bundle a season as a project | no | no | yes |

All of that comes straight from [OpenAI's pricing overview](https://openai.com/chatgpt/pricing/), except for one row. "Adjust it weekly" is my reading of the message limits, not a vendor figure. The fourth row is worth noting: you can use ready-made GPTs from the store for free as well, you just cannot build your own.

That leads to a simple rule. If your problem is the number of messages, Go solves it for 6.50 euros a month. If it is the quality of the model, only Plus helps.

### Free: good model, too few messages

For free you get limited access to GPT-5.5 Instant, limited messages and uploads, limited memory and context. In our benchmark GPT-5.5 shares second place with Grok 4.5 at 93 points, and on the safety questions it sits at 97, behind Claude Opus 4.8 only. For a first plan that is strong.

Two things still speak against it. First, the follow-up dialogue that makes the plan good in the first place eats a large share of the messages: after the opener, a round of questions, the plan and two corrections you are at six to eight, and the free limit is reported to be around ten per five hours.

Second, you work with a weaker model as soon as you call up "Thinking" on the free plan: behind it sits GPT-5.4 Mini, the way OpenAI describes it in its own [release notes](https://help.openai.com/en-us/articles/6825453-chatgpt-release-notes). On paid plans the same model steps in once the rate limits are reached. So of all things, the model that starts planning straight away on vague requests instead of asking back.

### Go: cheap, but with ads

For 6.50 euros a month you get more messages, more uploads and a longer memory. On price this is the most interesting step up.

One detail belongs with it that is easy to miss: OpenAI explicitly points out that this plan can contain advertising. If that does not bother you, you get considerably more room for little money. If it does, skip Go.

### What Plus adds

At 23 euros a month access to GPT-5.6 starts, along with extended messages and uploads, extended memory and context as well as projects, scheduled tasks and custom GPTs.

Two of those matter for training planning. GPT-5.6 is the model with the best asking-back behaviour in our test. And projects only start here, which makes the difference for planning a season across months.

### Pro: too much for training

Pro starts at 103 euros a month and offers five or twenty times the usage, GPT-5.6 Sol Pro as well as maximum memory and maximum context.

This is a plan for people who work with the strongest model for hours every day. Training planning, even with weekly adjustments, does not come close to those limits. If you have Pro for work anyway, use it. You do not need to subscribe for this.

## Your health data trains the model, if you are not careful

This point is almost always missing from plan comparisons. Conversations from Free, Go, Plus and Pro are used for training the models by default. Paying does not protect you. Only Business, Enterprise and API access are excluded by default.

For a training plan that matters more than for planning a holiday. You give your weight, resting heart rate, HRV values, injuries and possibly medication. That is health data, and under Article 9 GDPR a special category of personal data.

![ChatGPT data controls with the switch "Improve the model for everyone" turned off](/media/variants/chatgpt-data-controls-improve-model-off.webp)

:::info[tip]
**How to switch off the use of your training data**

In the browser, click your name at the bottom left, then Settings, then Data controls. Switch off "Improve the model for everyone" there. The change takes effect from that point on, for new conversations.

If the switch is already off without you ever touching it, you are probably working in a Business or Work workspace. There, training use is excluded by default. That does not apply to a private account, so check it yourself instead of assuming it.

[What the opt-out does not prevent](https://trustscan.dev/blog/opt-out-llm-training-data-2026): OpenAI still keeps messages for abuse review and for legal reasons, usually up to 30 days, in certain cases longer. Temporary chats never feed into training anyway, even without the switch. The downside: no memory, no history, the context is gone once you close it. For a single question with sensitive values that is handy, for ongoing training planning it is no substitute.
:::

The same goes for files. If you upload your old training plan, the same applies to its content as to typed text.

### The free alternative: local instead of cloud

If you would rather not let your health data leave the house at all, there is a serious option now. Gemma 4 runs offline on your own machine, the data never leaves the device. The large variant reaches 83 points in our benchmark and beats several cloud models with that. It needs a graphics card with 24 GB; below that there are scaled-down variants all the way down to models that run on a phone.

The 83 points apply to the large variant only. How well the smaller ones really plan we have not measured, and hardware recommendations from the internet are no substitute for your own run. That one is on the list: Gemma 4 on my own graphics card, with exactly the prompt from this article.

The gap to the top is real, 83 against 97. The question is what 14 points are worth to you when the alternative means your data never leaves the device. For most people the cloud with the training switch off will be the pragmatic compromise. Anyone working with sensitive data professionally may see it differently.

## How to build the plan: step by step

There are two ways to do it. [Way 1](#way-1-start-short-and-let-it-ask-you) starts with two lines and lets the model walk you through the questions, meant for anyone who does not yet know what a training plan needs. [Way 2](#way-2-the-complete-prompt-with-all-the-details) hands over all the details at once and saves you the rounds if you know your numbers. Everything after that applies to both.

### Way 1: start short and let it ask you

Keep the opener short and push the model into asking mode:

```
I want to reach [goal] by [date]. Ask me everything you need
before you plan. One question at a time.
```

Then answer the follow-up questions as precisely as you can, that is what decides the plan.

The reason for this wording is in the benchmark data. The strong models ask by themselves, the weaker ones do not. If you request the follow-up questions explicitly, you make yourself independent of whichever model the system hands you at that moment. That is useful on the free plan in particular, where you can end up on a weaker model under load.

"One question at a time" is not a detail. A block of twelve questions makes you skim them and answer briefly. Asked one by one, you think along with each of them.

### Why "you are an experienced coach" no longer helps

The most widespread guide prompt starts with a role assignment. "Act as an experienced running coach." That was sensible advice in 2023 and it is not any more.

[The research on it](https://arxiv.org/pdf/2512.05858) is fairly clear by now, but it separates task types cleanly. On tasks that depend on language behaviour, so writing, tone, role play, personas still help. On tasks that draw on factual knowledge, the expert role makes the result worse. The mechanism behind it makes sense: the attribution adds no knowledge to the model, it only distorts how the existing knowledge is retrieved.

Training planning with physiology, load management and zone calculations falls clearly into the second category. On top of that, the effect on current models mostly sits in the noise anyway. OpenAI itself recommends for the current generation to write away from the process and towards the result: name the goal, state the success criteria, give the constraints, and leave the route to the model.

What that does not mean: that roles are pointless in general. For a text in a particular tone they are still useful. For a training plan you can save yourself the line.

### The follow-up questions decide the plan

After the opener a good model asks about your goal, your training level, the time you have, injuries and safety aspects. The quality of the plan is decided here, by your answers. The conversation screenshots that follow come from a test run in German, and so do the plan and the numbers in it.

![ChatGPT asks its first follow-up question about the current running level after the short opener](/media/variants/chatgpt-trainingsplan-rueckfragen-laufniveau.webp)

Most guides get this part wrong. That modern models can ask back does not mean you have to prepare less. It only means they catch you when you do not. Being caught and being well advised are two different things. Our benchmark puts it like this: a vague request often triggers a weaker approach, and the best model alone does not make a good answer.

So answer precisely. "I run three times a week" is worth far less than "Tuesday and Thursday 45 minutes each before work, Sunday up to two hours, averaging 32 kilometres over the last four weeks".

### What you need to have ready about yourself

Before you start, it is worth gathering these details. Then you do not have to interrupt the conversation three times.

| Detail | What goes in it | Example |
| --- | --- | --- |
| Current level | Volume over the last four weeks, pace or watts per zone, resting heart rate | 32 km per week, zone 2 at 6:10 min/km, resting heart rate 48 |
| Weekly structure | Available days with specific time slots | Tue and Thu 45 minutes before work, Sun up to 2 hours |
| Goal | Specific, measurable, with a date | Half marathon on 5 October under 1:45 |
| Constraints | Injuries, complaints, equipment, travel | Knee twinges at speed, one week of travel in September |
| Competing load | Job, sleep, second sport | Shift work, six hours of sleep on average |
| Phase | Base building, race preparation or recovery | Base building after the winter break |

The example column is the actual point. "I run regularly" is not a detail, "32 km per week over the last four weeks" is. Your personal best matters less here than your current state. A model planning with your half marathon time from three years ago builds you a volume your connective tissue today will not go along with.

### Using your old plan as a starting point

The fastest route to a plan that fits runs through the plan you already have. There are two cases to tell apart.

Your previous plan can be uploaded, as a PDF, screenshot, spreadsheet or export from your watch. Have it read before you plan on: what was actually done, where the real load was, which sessions were regularly skipped. That way your history comes along instead of being rebuilt from memory.

Someone else's plan that you want to train by, from a book, from your club or from a magazine, calls for a different task. Not "plan something new", but "transfer this to my situation". Adapt volumes to your actual starting load, put sessions into your real time slots, convert zone figures to your values. The plan keeps its logic but gets scaled to you.

One practical note: uploads are explicitly limited on the free plan, Go offers more, Plus extended uploads. Anyone regularly uploading plans or exports runs into a wall for free.

### Way 2: the complete prompt with all the details

If you know your numbers, you save yourself several rounds by supplying them right away. The scaffold is a starting point, not a universal form: nobody knows in advance which details count extra in your situation, which is why the request for follow-up questions is in there at the bottom.

```
Goal: [specific and measurable, e.g. half marathon under 1:45]
Target date: [date]
Training days and time slots: [e.g. Tue/Thu 45 min in the morning, Sun up to 2 h]
Sessions per week: [number]
Current level: [volume over the last 4 weeks, pace/watts per zone, resting heart rate]

Equipment: [machines, dumbbell weights, treadmill, power meter, heart rate strap]
Training environment: [elevation, track available, surface, heat/cold]
Injuries: [what, when, how it healed, what triggers it today]
Pre-existing conditions and medication: [if relevant]
Age: [years]
What has not worked before: [e.g. three hard sessions per week]
Experience with: [structured intervals, deload weeks, strength training]
Competing load: [shift work, small children, travel, sleep]
Parallel goals: [e.g. losing weight alongside building performance]
Course profile of the race: [flat, hilly, surface]

Please note:
1. Ask me when something is unclear and do not assume anything.
2. Research the current state of the evidence and do not rely on your
   training knowledge. Name the sources with dates.
3. Tell me honestly whether the goal, the time frame and my available
   sessions fit together. If they do not, tell me what is missing
   before you plan.
```

The difference to the templates circulating everywhere: here there are placeholders for your numbers, not somebody else's values. That detailed input really does produce better plans has been measured by a [study on prompt quality in ChatGPT training plans](https://www.thefitfuturist.com/en/news/prompt-quality-chatgpt-training-plans-study/). It does not make them perfect, but noticeably more usable.

Good models ask for the basic details above by themselves. The block below is the part they did not bring up on their own in our tests, even though it shapes the plan considerably. What has not worked in the past is the entry with the best ratio of effort to effect. It stops you failing at the same point for the third time.

The third instruction is the most important one in the whole prompt. Without it a model will happily plan through an unrealistic goal too. It works out twelve weeks to your dream time, and on paper it adds up. An experienced coach would say at this point that there is not enough time or that one more session would be needed. You have to ask a language model for that push-back, otherwise you do not get it.

Even after this prompt there are still [follow-up questions](#the-follow-up-questions-decide-the-plan), and that is intended. And you [check the first plan](#check-the-plan-and-adjust-it-week-by-week) before you set off.

### Why you have to force ChatGPT to research

ChatGPT answers from what it learned in training, and whether it also looks something up online is its own call. Often it does not, and the tone stays just as confident either way. So you have to demand the research as soon as it involves something that may have changed. The switch is rarely the problem: web search is on by default, and it was on in my freshly created account too. Only if a source never shows up is it worth looking into the settings under Personalisation, right at the bottom in the expandable "Advanced" section.

![ChatGPT settings under Personalisation with web search switched on in the Advanced section](/media/variants/chatgpt-web-search-advanced-setting.webp)

:::info[warning]
**Ask for sources with dates, otherwise you get yesterday's state of the art**

Guidelines in training science change rarely, but then fundamentally. The [ACSM strength training guideline was rewritten in 2026 after 17 years](https://www.thefitfuturist.com/en/news/acsm-strength-training-guidelines-2026/). A model without that update will confidently recommend the previous version. So be explicit: "Research the current state of the evidence, meaning guidelines and studies from the last 24 months, and do not rely on your training knowledge. Give me the sources with dates." Then check whether it really searched. A model that names no source with a date has not researched, it has phrased. For stable fundamentals like progression, periodisation and recovery, what it learned is usually enough. It gets critical with guidelines, threshold values and everything that has moved in the last two years.

And click the links. The plan I checked for this article named five sources with a year. Four were right: two review papers on tapering and strength training, both exist and fit the topic, plus two pages from an orthopaedic professional body. The fifth was supposed to back up the calculation model behind the time projection. The DOI it gave led to a paper on cognition in old age from 1979. The reference was complete, looked serious and pointed at something entirely different. Four out of five is not a bad rate. It just is not enough if you have checked none of the five.
:::

![ChatGPT searches three websites and lists the sources it used in the right-hand sidebar](/media/variants/chatgpt-recherche-quellen-sportmedizinische-leitlinien.webp)

You can tell from the answer that it searched: ChatGPT reports the number of pages it went through, and the sources are listed individually in the right-hand column. If both are missing, it answered from memory. Then you follow up: "Please search for the current state of the evidence and give me the sources with dates." That costs one message and is the difference between an answer and a backed-up answer.

## Check the plan and adjust it week by week

Before you take the plan on, check four points. Does the volume rise gradually or jump? Is the intensity distribution plausible, meaning most of it easy and a small part hard? Are recovery weeks built in? And do the sessions really fit into the time slots you named?

The last point fails more often than you would think. A model likes to put a 90-minute session into a slot where you have 45 minutes. We have described the typical failure patterns in more detail under [seven warning signs that an AI training plan does not fit](https://www.thefitfuturist.com/en/training-analysis/ai-training-plan-limits/).

![Twelve-week training plan built by ChatGPT with weekly structure, longest run and target pace](/media/variants/chatgpt-trainingsplan-12-wochen-fertig.webp)

For this article I checked a complete twelve-week plan that ChatGPT built in conversation. The first three points passed. The long run started exactly where the previous one was and grew by five minutes per week. After three build weeks each came a recovery week, from week eleven the taper. All twelve weekly totals added up, all the dates fell on the right weekdays. AI plans usually come apart at those points, this one did not.

![The ChatGPT training plan as an Excel workbook with an overview sheet and a yellow input field for the test time](/media/variants/chatgpt-trainingsplan-excel-uebersicht.webp)

It only got noticeable when I did the maths. The long run carried around 45 percent of the weekly load throughout, 30 to 35 is usual. With four running days that is hard to avoid completely, but it packs the entire load onto one day. On the very day the knee was playing up in this case. And the plan did not keep to its own rule: at the top it said only one hard stimulus per week and Sunday stays easy. In the peak week there were then two hard sessions, Wednesday and Sunday.

The same pattern with the recovery weeks. They were there, they just pulled the volume back to a good 80 percent of the previous week, 60 to 80 is usual. And the reduction was exactly 35 minutes twice over. So the plan kept the absolute amount constant instead of the share. That made the recovery relatively weaker the higher the volume climbed, and the opposite would make sense. For your checklist: a tick next to "recovery weeks built in" is not enough, the question is how deep.

None of that is hallucination, it is structure. It only shows up if you add the columns together and compare the rules at the start with the plan at the end. That is what this step is for.

### Adjusting: report every week what actually happened

This is where the real value sits, and most guides stop before it. A plan from the first answer is a draft, not a result.

After every week, report back what actually happened. Which sessions did you do, which did you skip, how did the last hard session feel, what do your resting heart rate and sleep say. Have the plan carried forward on that basis instead of working through it rigidly.

![ChatGPT projects the 5 km test time onto a half marathon prediction and rates the target time as too ambitious](/media/variants/chatgpt-trainingsplan-zeitprognose-angepasst.webp)

The benefit is clearest when a number comes in. In my test run the reported 5 km time was enough, and the plan was not just adjusted, the goal was put in context along with it: from 25:50 ChatGPT projected around 1:58 for the half marathon and called the target time ambitious instead of simply planning it through. You only get that kind of feedback if you deliver.

This part breaks the free plan. Weekly adjustments across a season mean dozens of conversations, and for that you need both the message budget and a memory that carries across weeks.

## Custom instructions, memory and projects: what they really do

These three settings are meant to save you explaining yourself in every chat. They are no substitute for a good prompt: my test run for this article ran without all three and still produced a usable plan. Their value only shows once one conversation turns into a season.

### Custom instructions: useful, but not a rulebook

In the settings under Personalisation you can store what ChatGPT should know about you permanently. It is not much: a name, a job, a free text field and a few dials for the tone.

Do not expect a rulebook from it. The text runs along as a preamble in every chat, it is not worked through as binding. Anyone who knows a project file from software development, with rules the tool sticks to, will be disappointed here.

So the only things worth putting in are details that rarely change and apply everywhere: sport, training level, available days, chronic complaints, equipment. Everything that belongs to a specific goal belongs in the prompt. The target date, current values and the race change too quickly to maintain in a place you rarely think about.

### Memory: what ChatGPT keeps about your training

Memory stores details from individual conversations and calls them up later. For training that means: if you mention in March that your knee twinges after downhill runs, that can still be taken into account in September.

![ChatGPT personalisation with the "Enable memory" switch turned off and the memory overview next to it](/media/variants/chatgpt-memory-enable-switched-off.webp)

Check first whether the feature is running for you at all. In my freshly created account it was switched off. You find the switch in the settings under Personalisation in the "Memory" section, right next to it "Manage" leads to an overview of what ChatGPT has learned about you. So the training dialogue that carries across months does not start by itself, it starts with a click.

Switched on, memory takes the repetition off you. You do not have to explain your injury history, time slots and target time again in every new chat, and a question about a single session lands in the right context instead of in a vacuum.

The price is that intermediate states settle in too. An injury that was real in spring and healed by summer still holds sessions back in autumn. The same goes for a target time you corrected upwards long ago, or a 5 km value from a bad day. So look through the list every now and then and delete what is no longer true. And what you leave in is health data sitting permanently in your account. That is the second reason to set the switch from the privacy section beforehand.

On the free plan memory and context are limited, Go offers a longer memory, Plus an extended one.

### Projects: a whole season in one place

Projects bundle conversations, files and instructions on one topic. Setting up a season's preparation as a project means the plan, the weekly reports and uploaded exports all sit in the same context.

The catch: projects, scheduled tasks and custom GPTs start at the Plus plan according to the pricing overview. Anyone planning for free works with individual chats and has to hold the context together themselves.

## Connecting Garmin, Strava and health data to ChatGPT

[ChatGPT Health](https://openai.com/index/introducing-chatgpt-health/) is OpenAI's official route for health data: a separate space inside ChatGPT that you connect your data to once, so the model answers with your own numbers in front of it instead of your description of them. Where you are reading this from decides whether any of it applies to you. In the US the feature is open to everyone, in the European Economic Area, Switzerland and the UK it is not available at all.

Where it works, that is the interesting part for a training plan. Apple Health, MyFitnessPal and clinical record systems can be linked, and most watch apps write sessions, resting heart rate and sleep into Apple Health. From there the model reads the weeks you actually trained instead of the ones you remember. Since 23 July 2026 that has applied to every US user aged 18 and over, on the free plan too, and the health context can be pulled into any chat rather than living in its own menu. The medical record part stays US-only.

The data sits apart from the rest. Health chats, files and connected apps are stored separately and additionally encrypted, and according to OpenAI they neither train the models nor flow back into your normal conversations. That is not uncontroversial: [privacy advocates point out](https://therecord.media/chatgpt-health-draws-concern-privacy-critics) that patient records lose their HIPAA protection through the link.

The European exclusion is explicit, and the reasons given are stricter data protection requirements and a possible high-risk classification under the EU AI Act. No timeline for Europe has been announced. So if you plan from there, the protected area does not exist for you: your HRV values, your weight and your injury history land in the normal chat, where the training data use applies by default as long as you do not switch it off.

![The healthcare plugin category in ChatGPT with calorie trackers and fitness apps, but no official Garmin or Strava connection](/media/variants/chatgpt-plugins-healthcare-no-garmin-strava.webp)

A look at the healthcare plugin category shows where you stand. There are plenty of calorie counters and fitness apps, an official connection for Garmin or Strava is not among them.

What works without third parties is the export. Strava and Garmin give you your account as an archive, Polar Flow exports individual sessions directly as CSV, and anyone logging strength training has their spreadsheet already. The useful part is not the whole package, it is the activity list with date, sport, duration, distance and heart rate.

So do not upload the archive, upload a table with the weeks in question. ChatGPT works out weekly totals and intensity distribution from it, and you only hand over the data the question needs. [Workarounds through connectors](https://www.getvertical.ai/blog/chatgpt-strava-integration/) like Chirona or the Garmin Chat Connector do exist, but they are shaky: Strava revised its API rules on 1 June 2026 with an explicit reference to AI applications, and a Garmin auth change in March 2026 made many unofficial tools unusable.

If you already train and have a history, that is the better start. From real weeks the model reads what you actually did instead of what was planned, and which sessions regularly get skipped. Both get lost when you type things in from memory, and the plan then builds on wishful figures.

Without a history, five details per week are enough for the [weekly report](#adjusting-report-every-week-what-actually-happened): volume in kilometres or minutes, roughly the share of easy against hard, resting heart rate as a weekly average, sleep duration, and one sentence on how the hard session felt. Single values from your watch do not tell you much, trends across weeks do.

## Save your plan outside the chat

A training plan that exists only in a chat history is not saved. Copy the finished plan out once, into a note, a spreadsheet or your training app. That takes a minute and makes you independent of whether a provider rebuilds its interface, your history syncs or an account causes trouble. As a side effect you then have the plan where you actually need it during training.

The reason for the note has been showing up in the app since 9 July 2026: Codex has merged with ChatGPT. That is an agent for software development that changes files and runs commands, you do not need it for training plans. Since the launch of GPT-5.6 Sol, reports of data deleted on its own initiative have been piling up. Those affected were developers who used Codex with far-reaching write permissions on their own system.

For normal chat use this is no cause for concern. According to OpenAI support, around 80 percent of the reported vanished conversations are display errors, and another share goes back to mixed-up accounts. You should still copy your plan out, quite apart from this episode.

## Where ChatGPT differs from Claude and Gemini

All three have memory, but with different mechanics. That is the short answer to whether you should switch tools for training.

ChatGPT stores details from conversations and calls them up deliberately. Claude also offers memory across sessions, plus projects for separate contexts. Gemini works with "personal context", which is on by default for eligible private Google accounts, but condenses past conversations into a user profile instead of keeping the histories. The storage window can be set between three months and unlimited, the default is 18 months. For training planning the difference mainly means this: with Gemini you get a condensed summary of your history back, with the others more concrete detail.

On content our benchmark separates the three more clearly. Claude Opus 4.8 has the best safety score in the whole test field at 99 out of 100, Claude Sonnet 5 sits fourth overall with 92. Gemini 3.1 Pro reaches 85 and twelfth place. In return Gemini has live access to Google search, which helps when researching current studies, while Claude in projects only works with what you upload.

In practice: for the plan itself take ChatGPT with GPT-5.6, or Claude. For research on the current state of the evidence Gemini has the advantage through its search access. If you want to try both, the cross-tool comparison is under [building an AI training plan with ChatGPT, Claude and Gemini](https://www.thefitfuturist.com/en/training-analysis/create-training-plan-with-ai/).

## When not to follow the plan

With chest pain or shortness of breath, every AI plan ends. That is not a formality. From here it belongs with a doctor and not in a chat.

The signals where you stop training and do not keep discussing with a chatbot include pain in the chest, jaw or left arm, unusual shortness of breath at low load, dizziness or feeling faint, as well as a resting heart rate that stays clearly elevated for days and cannot be explained by recovery. The same goes for pain that increases during the effort instead of easing off.

There are limits below that threshold too. A language model does not see how you move, feels no pain and only knows about your daily form what you tell it. It can construct a sensible build-up from numbers, but it cannot judge whether your hip gives way at kilometre 15.

The most honest way to handle that is a division of labour. Structure, progression and periodisation are things AI can do, and in our tests it does that across the top group at a level of 91 to 97 points. The judgement about whether the plan suits your body on this particular day stays with you. If you notice you cannot make that call because you lack the experience, a real coach is the better investment than the next plan upgrade.

Start with the short opener, answer the follow-up questions carefully, and explicitly have the model tell you whether your goal is realistic in the time frame you set. In my test run that was the most useful answer in the whole conversation: the target time was too ambitious, and it was there before twelve weeks of training were built on it. If you would rather start without a prompt, you can use the [AI training plan generator](https://www.thefitfuturist.com/en/tools/training-plan/) as an entry point and do the fine-tuning in the chat afterwards.