LLM Spend

Workroom Productions uses LLMs. I refuse to use LLM subscription billing (mostly), and I use local models for some local coding, to transcribe audio, and to power DrawThings. I've not included services that are LLM-powered behind the scenes (i.e. DuckDuckGo's search summaries) if they generally pass me by.

I want to share this to hold my company to account in some way – the inference costs are a fraction of the wider impact these tools have. I have a horror of the costs incurred by friends and colleagues as we march forwards attending to our individual monthly bills, not to the wider implications.

Here's what's been spent, in rough order of how much I use each provider. Costs calculated to Sept 2026.

Anthropic $280

My daily doofer; typically Haiku and Sonnet via Msty. I put $10-$15 in the bucket every 2-3 months, but the bank can see projects where I used Claude Code, as I can spend than $15 in a day when using agents. When I run a workshop, I usually chuck in $20 and have half left over.

Kagi and Raindrop ~$10 month (subscription!)

When I say I refuse to use LLM subscription billing, there's some nuance to that. I wouldn't work as easily without my subscriptions to raindrop.io (£25/y, bookmarking) or kagi.com (£6/m, search), and I use their LLM-based tools extensively – using Raindrop's Stella to dig into my bookmarks (and the text saved with them) and Kagi's assistant to run, aggregate and filter searches.

OpenRouter $25

Whenever I need to test out a new model, and in particular for running small local-capable models (i.e the various flavours of Qwen) remotely when I can't devote my machine or my evening to tokens.

Minmax $30

Their minimum spend – I'll need to use this before it expires like my OpenAI credits. I'm currently using it to build out a text-based wiki-ish knowledge base of basic skills (digging a hole / making gloves) as an experiment in agentic stuff.

Amp / SourceGraph Cody $140

I was a Cody subscriber from Feb24 at $9/m for about a year. When they switched direction to Amp they gave me $100 in free use, and I've spent more. Amp seems fabulously effective at explaining code and finding ways through problems, but it's expensive and was very clear about what it was costing. I used their $10 a day allowance for a while, and found that I handed it too much stuff.

So I started to use it purposefully for knotty work – if I'm getting nowhere, I fire up Amp. Excellent though it is, I've not used it much recently.

Jina $0

Somehow I'm still on a free plan – I go through 1M tokens every few months. Msty uses a Jina key to pick up web pages.

Replit $225

Bart and I had workshops that used Replit, but their terms got more and more restrictive, which is understandable. This is 10 months subscription from the moment they stopped being free, and I didn't turn it off soon-enough. This isn't all LLM – we used it initially to offer environments, but the LLM use built up as it became a safely-isolated vibe coding tool. I found I got more and more lost in their IDE and was hating it in the end.

Replicate $14

For a while, the simplest way to engage with an LLM on the commandline. Used for several workshops in 2023.

StableDiffusionOnline $10

Not an LLM, I suppose. Used to make unique (and weird) New Year's cards in 2022.

OpenAI $25

I can't be doing with ChatGPT's memory of what I've done: It's creepy, so I use it very little. Most of this spend is because I put $20 in for some workshops in Summer 25, participants used $5(!), and the rest expired.

There's more, as I remember more – I used continue for a bit, and had an account with a provider or two for their multi-model approach. A fiver a piece, I reckon, but I don't yet recall which ones.