Code and installation: github.com/unclecode/jevshif… Next I want to do the same for thinking effort, so Jev can pick both the model and how much thinking the task needs.

Another good use I found for @typesafeai’s Jev: automatically switching Claude Code models based on what I’m working on. I built JevShift for this. It looks at the task and recent context, then picks a model. Here’s a quick demo 👇

This is how I matched: modelprint fingerprints 23 known models every day, so when a mystery signature shows up i just compare. modelprint.ai/history btw i also measured the models the viral post says are inside pareto. none of them match what it served me. rumors need measurement too :)

i fingerprinted this "union alpha" which turned out it is (unbiased/pareto) for 15 times in 100 minutes: 02:05-02:34 UTC: a fingerprint matching hermes-4-405b in my library, 4 of 4 probes. 02:37: every probe value changes at once. new model. 02:40-03:46: a signature matching nothing i have measured. unknown model. and in 10 of the 15 runs, the same text sent twice, seconds apart, was billed with two different token counts. people spent the stealth week arguing whether it was openai, llama, qwen or glm. everyone was right. they were fingerprinting different models behind one name. Not only for mystery models. Your daily OpenAI or Gemini API deserves the same check. modelprint.ai

I actually today measured it for multiple time and each time was getting different fingerprints! I wasn’t aware that it’s revealed and it’s a router! I will post later today about the result

I really don’t want end up with 100x subscriptions, I even forget some of em, get charge every month while my usages may not touch the cap!

Terminal: My Precious (Read in Gollum voice haha)

I really wonder how will be the saas model? Here @treg_ai let us access to 2800+ saas and pay as u go!

Fabel 5.1 is where finally I feel claude code talk concise and relevant! It gets the point much faster, do not generate long essay result for a simple question and so far so good!

I think paying for all LLM should become pay as you go! Pay only when u use and not monthly subscription! Unless u r a heavy users. Is that something u look? Check this treg oss library and hosted version

An AI agent to make sense, needs many tools, or better to make many api calls, and imagine if you want to subscribe to each of em and never never to their cap! How come you make that agent into a profitable business? The solution is a proxy that lets you pay just for the API calls you make! Here @treg_ai is doing that and more: tool routing, team account, remvoe the need to share api keys between team members and many more.

Ok the mystery the ox-alpha is GLM-5.3-Flash, And modelprint.ai was right! ox alpha was the first real test I did. day one, the fingerprint said glm family, 4 of 4 tokenizer probes. day five, z.ai confirmed it. today i reran it on the official glm-5.3-flash endpoint: same values, 47, 56, 60, 90. the label changed, the model did not. anyway the guessing game was fun, but it was not the point for me. but the real question: can you fingerprint a model from its behavior alone, without access to weights? even for open models, reading the weights is slow. behavior is fast. if that works, you get something bigger: a version control system for models that does not depend on the lab, the provider, or anyone's word. the API itself is the source of truth. so now it is a tool, and it is open. join, contribute, add probes. and when the next mystery model appears, we run it again. modelprint.ai

update: the official name is out, glm-5.3-flash. i reran the fingerprint on the new endpoint and compared with day one of the stealth run: tokenizer probes 47, 56, 60, 90 then. 47, 56, 60, 90 now. same template delta. the label changed, the model did not. modelprint.ai/history

congrats on the release :) great job you guys have done here, kudos, btw for anyone curious, this was measurable from outside. i fingerprinted ox alpha on day one and the glm family matched all 4 tokenizer probes. the label came 5 days after the data: xitter.cf/unclecode/status/20925…

I wish “feedback” had option to not share my session context!

the interesting part: this was detectable without any inside source. i fingerprinted the api 5 days ago and it matched the glm family on all 4 probes: xitter.cf/unclecode/status/20925…

fun detail: the routes were telling on it the whole time. fingerprinted it 5 days ago, glm 4 of 4 probes: xitter.cf/unclecode/status/20925…

not even apparently, its measurable. i fingerprinted the endpoint 5 days ago and glm matched all 4 tokenizer probes: xitter.cf/unclecode/status/20925…

and it was verifiable before the confirmation. my fingerprint gave glm 4 of 4 probes, gemini 1 of 9, claude and grok 0. measurement first, label later: xitter.cf/unclecode/status/20925…