I have a "fringe" theoretical physics idea[1] that I've been grinding through various frontier models to see how far they can push this kind of thing. It's a great test because it is sufficiently "out there" that they can't memorise the answer, and it has absolutely brutal mathematics.
A simple cross-check is to get two models to advance the paper, and then have the other model validate it.
All of the Gemini 4 models -- including Neon -- consistently make simple, "doofus" errors like neglecting minus signs or doubling a scalar factor due to accidentally multiplying by it twice. Similarly, they often use two different sets of assumptions for a long mathematical argument involving multiple proofs, so the whole argument ends up invalid.
For comparison, these basic errors no longer occur with any of the recent Anthropic or OpenAI models, including every GPT 6 Sol or Astra model that I've tested.
Simply put, if you let Astra do 5-10 iterations in a row (~1 hour of work) and feed it to Gemini 4 Neon, it won't spot any errors. Do the reverse of that and even GPT 6.2 Sol will tear Neon's work to pieces.
PS: I suspect that Gemini models have a higher temperature setting, which makes their English sound more natural, but introduces errors that degrade their hard science capability.
[1] But not crank physics, which I would expect to "trip up" models trained on correct, valid physics and mathematics.
Were you using any / commands when it made doofus errors?
I also found Argon made a ton on doofus physics mistakes. All caught by Astra. But it got better (and much slower) when using /teamwork.