Hi - please bear with me on plans - there’s a lot of work going on with this and I should be working on the weekend on this!
1 Like
A small pause in feature dev - to approach a wider set of markets we need to change our payment provider to reduce the growing admin burden - this isn’t Facebook scale and so any reporting duties we bring on ourselves really hurt from both an administrative and cost perspective. Please bear with me - the planning + workout creation improvements are coming!
2 Likes
Any updates recently?
Hi @Fln_dtr - coming soon. The project now has significant paying users, and as such it ends being ‘experimental fling it’ and more consolidated and careful development process - our release cadence will now become roughly monthly in order to carefully test and evaluate changes.
We have an upcoming change which is inevitable but frustrating, to change the payment provider, which is why this one is a little more delayed than usual.
Upcoming features in the next release:
- stripe → paddle
- fit file parsing (retrieve raw from intervals.icu - so we can parse stuff like running dynamics from garmin etc.)
- token/model efficiency (another stab)
- planning improvements (big feature)
- daily checks - token efficiency (cost reduction)
Tomorrow should be the day I get this all consolidated and out. Thanks for your patience!
1 Like
UPDATE:
Visual Workout Builder — When the AI creates workouts, you now see charts showing structure, intensity zones, and duration right in the chat. Multi-workout
plans are grouped by week. You can ask for changes ‘in real time’ by continuing to prompt the AI
Workout Validation — Workouts are validated before hitting your calendar. Recovery gaps between hard intervals are flagged, common format mistakes are
auto-corrected (e.g. “400m” → “400mtr” for swimming), and errors show the exact line.
Better Approval Flow — Inline approval panels replace the old popup modal. Improved mobile touch targets and scrolling.
New Billing via Paddle — We’ve switched our payment processor to Paddle, who act as Merchant of Record. What this means for you:
- Prices in your currency — the pricing page now automatically shows amounts in your local currency (EUR, USD, GBP, etc.) based on your location. - No more mental currency conversion.
- Faster checkout — instead of redirecting to a separate payment page, checkout opens as an overlay right on the site.
- VAT/tax handled for you — Paddle takes care of all global tax compliance, so the price you see is the price you pay.
- Existing Stripe subscribers — nothing changes immediately. Your current subscription runs until its renewal date. You’ll receive an email with a personal discount code (20% off) to switch over to Paddle at your own pace.
- Your data and subscription tier stay exactly the same. You won’t be double-charged — Stripe cancels at period end, not immediately.
Wellness-Aware Coaching — The AI now factors in your self-reported wellness (mood, fatigue, stress, sleep, etc.) when making recommendations.
FIT File Analysis — The AI can parse your FIT files for detailed metrics: running dynamics, power data, HRV, GPS/elevation.
Under the Hood — Longer-lasting sessions (no more random logouts), security hardening, mobile layout fixes.
As always feedback welcome. This was a big one!
In today’s update, were there any changes to “token/model efficiency”? It still seems to be burning credits faster recently including for today’s failed Daily Training Check.
Yes! Plenty of token optimisations - but unfortunately one of them has lobotomized it a little (in testing I’m seeing some issues, I’m reluctant to switch back to a top tier model just yet though as the main complaint is cost). It’s such a tough balance - the models that can do this analysis/planning/tool calling well are not cheap. We’ve got multiple tiers and perhaps tomorrow I’ll try to do some reconfiguration to make it ‘better and cheaper’ if that is possible!
OK so this prompted me to spend more time than I should have researching models, how we use them, comparing them, doing some analysis side-by-side and tonight’s deployment has pretty much done a great job.
I need to do some work on long-horizon planning that’s still a weakness, but please bear with me it’s very nearly there, a few tweaks and maybe a few more model swaps and we should be doing a much better job of balancing ‘getting it right’ with ‘not costing the earth’!
Thanks for the patience, and feedback 
UPDATE:
-
Models & credit usage - massive improvement
-
Planning - different stratosphere - faster, can probably do a few 20 weeks plans on the free tier
-
Daily checks - email now sending via us:
- it’s unfair to expect intervals.icu to deliver our mail
- we get better control over formatting etc.
- it’s clear replies would to to support@ rather than the AI picking it up (coming soon feature I guess - though I’m not sure email correspondence would work so well for this)
-
UX/UI - improved mobile, improved font, improved vertical spacing
-
Per-athlete Memory - memory system will retain facts and use them i.e. ‘runs a 3:30 marathon’ (I wish rn… haha) - this augments the training availability system accessible via chat/settings
As always - that is a LOT. Please let me know of issues! Happy to resolve!
2 Likes
Fixing one problem I discovered with the tool tiering system where sometimes the AI can’t ‘find an activity’ once it’s gone into deep analysis toolset. Providing the AI with too many tools can confuse it, but in this case I was over-zealous with the cutting down of the toolset in some cases! Will be fixed shortly!
UPDATE:
- email links fixed on daily checks
- analysis improvements - tooling available, and caught a bug on best efforts whilst doing my own vanity prompts about a run I did this afternoon
Daily checks should be good tomorrow, but will try to get some time to clear out an issue where the action would be delete existing + create new - rather than just update. Not a blocker, but a little suboptimal UX.
After what was one of the most terrible weekends of my life doing battle with AI models (evaluating cost/quality, and different properties) - I think the system is stabilised.
It’s worth noting if I get interest I could do a ‘super premium’ version where we use opus4.6 and just all the problems go away - I’ll gauge that one later! Hopefully now we’re pretty stable though - please report any issues to me!
UDPATE:
Today comes out a release that’s been building for a while in parallel with the prior fixes. It makes leaps and bounds on top of the original planning fixes.
- new site
- new more visually easy colour scheme
- daily check robustness
- intervals.icu webhook usage for more seamless experience (daily check wellness instant, and cache invalidations)
- updated the tooling to include some newer analysis endpoints on the intervals.icu API
- loads of smaller bits
2 Likes
It looks like the Ai trainer doesn’t work as good anymore… I’m having difficulties to let it create a training plan for the coming week.. Couple of weeks before it worked like a charme (it did know when I want to cycle and when strength training or stretching ). Also my standard weekend MTB ride (3,5hr) was in the planning. Not anymore… What has changed?
Thanks for the feedback it’s been quite heavy this weekend as you can imagine :p. There are a few reasons for the changes:
- letting the model do it’s own thing wasn’t very scientific
- the system was erroring and generating awful workouts most of the time that wouldn’t parse in intervals.icu and a lot of other failures I could see in our logs
- COST - the biggie. Things are dialled back to cater for most of the users’ cost appetite (read back on the thread here, multiple worries about token consumption cost) - in answer to this I may consider ‘big brain’ mode where you can burn tokens for fun - that is obviously something we could do! (and probably quite easy/good idea).
I’m working hard (and have been all weekend tbh) to get this completely nailed. Ensure you have proper memories and training availability set in the platform - that will help it know when you train and what your preferences are.
The system will limit ramp rate etc. and I’m aware with one small bug on updating workouts in the pending queue -however in our benchmarks and tests this new system performs 100x better for reliability and ‘sensible workout’ selection.
Leave it with me, I’m working on it
1 Like
Hi,
I cannot find the pending Queue in intervals that are supposed to be sent from Intervals.pro. I might overlook somewhere and probably an easy fix but please let me know!
Cheers
This is an external project and not related to intervals.icu directly.
Please ask the developer of that platform.
The ‘chat actions review’ should pop up as soon as the workout/plan/action is ready - please ensure to clear browser cache/cookies etc. and hard refresh this should be seamless. I’ll investigate the issue separately too.
I’ve also noticed a clear degradation in the AI behavior over the past week. I’m spending significantly more time “arguing” with it than before, with more misunderstandings and incorrect actions.
Is this a temporary regression, and is there any improvement in sight?
Also, does using BYOK (e.g. via OpenRouter) make any difference here—especially if this is related to cost or model selection?
Thanks for the feedback! BYOK doesn’t make a difference other than billing in this case.
We have a ‘user support’ fight going on - cost vs quality. I think I will aim to provide a set of choices around ‘model costs’ - this is a fair bit of work. Where I will allow substituting the models for higher cost and better output. It seems there’s not a good balance for our particular workload out there at the minute. Of course opus 4.6 would do a great job, but cost about £1.50/user/day with minimal usage.
Workout generation is still a pain intervals.icu has no validator and there’s very little going on in this space other than some garmin stuff which I’m trying to integrate at the moment. Please leave that with me, we should have a simplification and solution soon.
I believe analysis however is much improved currently due to some innovations on how we’re using the model and the tools it can see.
Thanks for your support and feedback!