This is as much on the specific question I face as it hopefully clarifies for me (and possibly others) the fine details of external AI use and setting requirements in general.
Looking at my Openrouter account, I found that DT is – or, seems to be – calling Amazons “Nova Lite 1.0”. Now, invoiced amounts are miniscule in each case (“0,000117”), but I have just found within the last day alone there were 50+ such calls/usages.
I generally wonder what is going on here, as I never selected “Nova AI” anywhere in DT-settings to begin with.
So, I am wondering whether this unintended process, where some data seems to flow from my machine, is initiated by something within DT – or whether I would have to inquire Openrouter about it?
Thing is, I definitely know things are principally somehow connected to DT, as I just set up a dedicated API-key that is only used by DT, and filtered the Openrouter-activity monitor on this API-key. This is where all the “Nova AI” token processings are listed (like “1.826 > 31” token usage in an indivudual entry)
I wonder, is there anything in the settings I have misunderstood that opens up this process (see screenshot for settings?
(Note: I have not used chat at all during the time “Nova AI” usage is registered on my Openrouter account…)
Thanks for any knowlegeable person chiming in on this!
In that case, it is better to use the bespoke provider and model instead of using OpenRouter. The same is true if you have an Anthropic, Gemini, etc. API key.
So, you are saying OpenRouter is billing me on usage of an AI provider I never picked anywhere?
That would be quite big news for everyone in the DT-community using the OR-API-keys, I suppose…
DEVONthink doesn’t use or prefer Nova Lite internally in any way. Which AI features do you use regularly? E.g. the chat inspector/popover, scripts, templates, smart rule and batch processing actions/placeholders? Is the model selected in Settings > AI > Summarization or Settings > AI > Transcription?
So, you are saying OpenRouter is billing me on usage of an AI provider I never picked anywhere?
That would be quite big news for everyone in the DT-community using the OR-API-keys, I suppose…
No, that’s not what I said at all. I commented on you accessing ChatGPT 5 via OpenRouter and said you should use ChatGPT in the AI > Chat > Provider popup instead.
I see!
The strange thing is, I didn´t use AI in any conscious/intended way in that last timeframe where this appeared as linked to my DT-API key
This is why I also included the AI-setting thinking, maybe I opened any route for some “background” processes that DT does (on auto-ocr´ing things, or so…)
So, I am still buffled and puzzled…
And, no, neither is it selected in DT settings (where “standard” is checked; … but then I didn´t use any summarization the last day… but still see 50+ activations of Nova over 3-4 timespans…
… and I am OK with anybody telling me, it´s an OR-side problem.
I am just going this route, as things only (seem to) happen in association with my (exclusive) DT-API-key from OR…
Currently things might “clear up” … even if in a strange way. Openrouter dashboard now suddenly says “Activity logs are temporarily unavailable”.
Dark AI times, I guess.…
OK – kind of an intermittent report, status update:
• the activity-log at Openrouter is back “on” – and still shows those unexplained “Nova Lite” (Amazon) API-calls, billing them to me. And informing me, some data flowed here…
• it still is only showing up with the DT-assigned API-key; the pattern is always/mostly around 1500 tokens in, 20 out – around 50 on a day. So it´s small, but still a problem/issue in principle it seems to me…
• I have now created a ticket with Openrouter to get their take on things. Will see, will report.
• As to the teams questions on what I use with DT: almost never templates; sometimes scripts – but they all date from pre-AI times and non-AI contexts. – I have used chat once around the time where the calls are reported – though I can´t firmly correllate this (also there are distinct 4 timeblocks where the calls to Nova Lite were made, and I definitely used chat only in one go…)
• … so, I am still at a loss. What remains is a little irritation that data (seem to) flow unintended to models (providers) never chosen. But, of course, I can´t pin it down on either DT or Openrouter, or any other third malign Halloween actor, just yet…
• … practically, I await Openrouters reaction, and will then see what to do with it. For now, I have blocked Amazon out in the Openrouter settings, and for now it seems to do the stopping. But, with all this in my back, I will check this for the time to come…
That´s my – inconclusive – update. But it´s where it is.
And definitely no smart rules or batch processing that use the actions Chat - Query or Chat - Continue if…? These actions also support choosing a different model, just like Settings > AI > Summarization and Settings > AI > Transcription > Text recognition via Chat or scripting.
Is anything related to Nova Lite logged in ~/Library/Application Support/DEVONthink/Chat.log?
Thanks, @cgrunenberg – this is an interesting document, indeed!
I think, I have a clue now – though all this is proam-guesswork.
But reading: 2025-10-28 16:31:06,845 ERROR: OpenRouter (openai/gpt-4.1-mini): Key limit exceeded (total limit). Manage it using OpenRouter 2025-10-28 16:33:21,562 INFO: 75 of 346 available OpenRouter models used. 2025-10-28 16:33:51,661 INFO: OpenRouter (openai/gpt-4.1-mini): 177014 input, 186 output tokens used. 2025-10-28 17:30:55,521 INFO: OpenRouter (amazon/nova-lite-v1): 1643 input, 18 output tokens used.
… I get the impression, when my OpenRouter account hit the assigned usage-/cash-limit it seems to have “automatically” swapped over to Amazons Nova Lite (as I haven´t actively blocked it before).
Does that make sense?
So, it really is an OpenRouter issue, of sorts. At least totally related to OR…
… there is a backstory, though, that I might put somewhere else: the funds-limit was hit, as suddenly in the last month my DT-AI-usage was going beyond 50$, without me being aware of really extensive active usage… This is when I changed some of the settings to see what influences this and leads to such unintended amounts within DT (so, I disabled “allow modification of content” as well as “Online OpenAl GPT-4o-Transcription” for audio/video and OCR via Chat (which I didn´t intentionally use anyways))…
So, to not disgress here: thanks @cgrunenberg (and @BLUEFROG ) for following things so far!
Maybe you can verify my reading, to “close the case” of this particular issue/thread?
And maybe this thread will give some insights for future use to other peole here utilizing Openrouter.
PS - proposal: I think the existence of this log is a very valuable and helpful info. I´d propose to include this somewhere explicitily in the documentation. I am sure it will help others sorting their AI-relationships in DT, eventually in case of need.
Only theoretically as DEVONthink logs the provider/model chosen by the user, either in the settings or on demand e.g. in smart rule & batch processing actions, scripts or in the chat inspector/popover. DEVONthink does not prefer or fall back to this model on its own (as the Nova models are not competitive, IMHO).
OK. My reading would remain, as this makes the most sense to me.
I think there was an autoswitch on OpenRouters side once the limit was hit. It seems the only “signal” there ever was, or at least the only one I can see/identify.
At least, I am assuming now it was not on the side of DT, as I positively never selected the model in DT (or anywhere else, for that matter), and the coincidence with the rate/fund limit is positively there.
So, no need to assume anything opaque happened on the DT-side, for now & in my take.
Which of course is a relief.
So, thanks for pointing to the “chat log”-file, which I wasn´t aware of. As said, I think it would be helpful to document that for all those trying to triage or simply follow the DT-AI-interactions. Helps against AI-paranoia. A bit (sic!) at least.
