[cadamsdotcom]

My suspicion (unconfirmed so take it with a grain of salt) is they either used some/all test data to train, or there was some leakage from the benchmark set into their training set.

That said Sonnet 4.5 isn’t new and there have been loads of innovations recently.

Exciting to see open models nipping at the heels of the big end of town. Let’s see what shakes out over the coming days.

[pertymcpert]

None of these open source models actually can compete with Sonnet when it comes to real life usage. They're all benchmaxxed so in reality they're not "nipping at the heels". Which is a shame.

[viraptor]

M2.1 comes close. I'm using it now instead of Sonnet for real work every day, since the price drop is much bigger than the quality drop. And the quality isn't that far off anyway. They're likely one update away from being genuinely better. Also if you're not in a rush, just letting it run in OpenCode a few extra minutes to solve any remaining issues will cost you only a couple cents, but it will likely get the same end result as Sonnet. That's especially nice on really large tasks like "document everything about feature X in this large codebase, write the docs, now create an independent app that just does X" that can take a very long time.

[rubslopes]

I agree. I use Opus 4.5 daily and I'm often trying new models to see how they compare. I didn't think GLM 4.7 was very good, but MiniMax 2.1 is the closest to Sonnet 4.5 I've used. Still not at the same level, and still very much behind Opus, but it is impressive nonetheless.

FYI I use CC for Anthropic models and OpenCode for everything else.

[unsupp0rted]

M2.1 is extremely bad at writing tests and following instructions from a .md, I've found

[stingraycharles]

It’s a shame but it’s also understandable that they cannot compete with SOTA models like Sonnet and Opus.

They’re focused almost entirely on benchmarks. I think Grok is doing the same thing. I wonder if people could figure out a type of benchmark that cannot be optimized for, like having multiple models compete against each other in something.

[c7b]

You can let them play complete-information games (1 or 2 player) with randomly created rulesets. It's very objective, but the thing is that anything can be optimized for. This benchmark would favor models that are good at logic puzzles / chess-style games, possibly at the expense of other capabilities.

[NitpickLawyer]

swe-rebench is a pretty good indicator. They take "new" tasks every month and test the models on those. For the open models it's a good indicator of task performance since the tasks are collected after the models are released. A bit tricky on evaluating API based models, but it's the best concept yet.

[astrange]

That's lmarena.

[satvikpendem]

You are correct on the leakage, as other comments describe.