If you're selling model capabilities for money, and you change the mobile capabilities, not changing the model designator (name and/or version number) is fraud, full stop.

Unlike here where the "phantom bank accounts" fraud was successfully pinned on tellers and branch managers who were all "independently" just trying to meet management goals...

And yes, if you're doing engineering that violates the law, you should be fearful of doing. That's kind of the point of punishing criminals.

https://www.justice.gov/archives/opa/pr/volkswagen-engineer-...

Opus 5.5 was significantly crushing every task I gave it for quite some time.

Suddenly, it starts making the type of mistakes that even Sol 5.6 would not make. It became noticeably worse than Astra.

I've grown so completely tired of the labs changing the quality of the model out from under our feet that I CANNOT wait until there's viable alternatives at the frontier level.

I will gladly forgo the best available power for consistency - and I see that coming soon (hopefully with Qwen 4 or shortly after).

DeepSeek v4.1 Flash is not even remotely in the same league as an un-nerfed Astra or Opus 5.5 - despite what microbenchmarks may claim.

It's a great model, especially for the cost - but there's a huge difference.

Oh good Lord, they stir the pots constantly. Anthropic is hella naughty about this. Someone else on HN said it well, Anthropic has too much of a "researcher" mindset. You cant experiment on human populations tool use the same way you run AI/ML experiments.

I want a stable tool that works consistently, not a hammer that changes mid-swing (credit Mario Zechner for this analogy).

Google and Facebook have proven otherwise. Unless you are very, very lucky you are always part of multiple running experiments on their products, and nobody really cares.

2. Google doesn't pretend that its algorithm has certain capabilities while serving the user a nerfed version

I worked recently for a German megacorp, for whatever reason they gave us unlimited access to a bunch of models, up to sonnet 4.5/Gpt o1 etc, so not the best or latest models. And no agentic tools.

So I built a custom Claude Code clone because they exposed the api from their custom 'LLM Chat' clone in the client and accepted my user credentials for access.

Essentially a frankenstein AI harness using various models for their strengths, including using the OpenAI best model for memory context consolidation and compaction + Sonnet as orchestrator, etc. I was able to spawn 20+ agents to do my work and basically automate my entire job at that corp. I think my internal bill was something like 5k euros per month and I was being careful to not abuse it...

It was better than the best you could get from claude at the time (about a year ago now I think), in terms of task performance completion.

So imagine what someone with essentially infinite spend can do with endless fable agents. AKA US government and Anthropic internal.

Not a great plan

I don't know who they think they're fooling.

https://news.ycombinator.com/item?id=50429947

If I order a 16oz beer and get a 12oz pour, I'm going to call it out.

And then Sam Altman is gonna sue your ass for owning a measuring cup.

Is this 2026's "my macbook feels 'snappy' when I update it", though? (That is to say, not really much in evidence)

House of El covered this in her latest video incidentally.

I'll gladly sign up to be a "Experimental Settings" beta tester (with knobs of how extreme I want to go) for a slight discount (or just bill me the same, just gift me a reset every week? idk whatever).

Doing A/B testing in prod on a frontier model that could really mess with the users in really bad ways is a bad idea.

I suspect it's rather a simple matter of cost.