Haider.

opus 5.5 is behaving like a very different model today it had been flawless since launch, but now it's dropping tasks, skipping commits, and claiming to have done work it hasn't. even its deep reasoning seems to have turned into basic pattern matching i wouldn't call it a nerf yet, but is anyone else experiencing this?

astra has a bigger brain than opus and fable but astra was heavily mistreated during pre- or post-training with too much focus on benchmarks and too little on alignment to use a human analogy: you can be smarter but still make stupid mistakes

Sam Altman says: "i get sad when people say AI will automate everything and we'll be like house cats, with the AI deciding how to reward us. that's not what we think" I'm hopeful for a huge wave of scientific discovery and new companies giving everyone the chance to create something new

assuming Astra 6.2 is real, i can see two possible scenarios: 1) this is a bigger and more capable model than Astra which is possibly the rumored "Bel" openai previously chose not to release 2) it's a model from the astra/sol/terra/luna lineup but oai is using a temporary internal name to avoid leaks i don't think they'd release another mid-tier model either

how on earth does anthropic get to a $2T valuation? SpaceX was valued at nearly $1.8T for its IPO, and they don't just have an AI lab and data centers -- they have reusable rockets + a LEO constellation no one has come close to matching even with that $1.8T felt insanely high to me

anthropic is the smartest AI lab when it comes to acceleration right now all those months spent red-teaming and trying to understand model behavior seem to be paying off. they've gotten better at shipping frontier models safely while still moving faster than openai and that's interesting especially if both labs have similar automated research capacity

gemini 4 Argon isn't quite at Astra or Fable's level but it's really good at scanning large codebases or databases and generating extremely large outputs in a single response Carbon is marketed as Astra and Fable-level but by the time it's available to the public, it will already be outdated by Fable 5.5 and Astra 6.x very poor strategy by Google

the biggest difference between google and openai/anthropic is what they have internally i think openai and anthropic both have significantly stronger models than what they've released so far. i don't get that same impression from google gemini 4 might actually be the best model google's own researchers have access to right now

there's a clear pattern i've noticed with codex and claude code: codex often runs a lot of tests. if you look at the context, it usually finishes the main task in a few minutes, then spends the rest of the time running tests and even tests for those tests this is why claude code/models are generally good for vibe coding and one-shotting claude is faster but codex is more resilient

gpt-6.1-sol is a fantastic model the biggest thing for me is that the base intelligence is finally good again. i really hate when the lower tiers of a model feel stupid, and i don't care how smart it can get at xhigh if the base model isn't good base intelligence dictates everything

i don't get the hate or mockery toward gemini 4 when it isn't even out yet their TPU advantage is massive and doesn't get enough attention. while everyone's focused on the chat interface wars, google's building the infrastructure to run AI at scale without relying heavily on Nvidia the developer side still feels overlooked

AI is getting cheaper and cheaper. AI plans are getting more and more expensive. perhaps a frontier model can explain this paradox to me