Mark

[Mark @mark15631]

Same! It's probably like they said officially, it's more oriented toward auto-completing tasks.

I'm guessing they're catering to the majority of users, since most people just use it with a one-liner: "get this done."

Those of us with a more structured workflow are probably the ones being sacrificed.

[Mark @mark15631]

@bcherny Thanks for sharing the internal perspective.

I just tested Claude Opus 4.7 in the web client and noticed something concerning:

1. In long-context scenarios, it’s extremely difficult to trigger think / visible reasoning steps

2. When I shorten the context length significantly, it triggers thinking much more easily

This makes it hard not to suspect that the long-context regression is a genuine model-level issue, not just an MRCR benchmark artifact.

It could be an inherent capability problem or the model has been deliberately nerfed for long-context use cases so it doesn’t enable think / reasoning.

[Boris Cherny @bcherny]

👋 We kept MRCR in the system card for scientific honesty, but we've actually been phasing it out slowly.

Two reasons: (1) it's built around stacking distractors to trick the model, which isn't how people actually use long context, and (2) we care more about applied long-context capability than needle-retrieval. Graphwalks is a better signal for applied reasoning over long context, and internally we've seen this model do really well on long-context code.

MRCR wasn't included in the Mythos Preview system card for these reasons, but Graphwalks was - that will be the case for future models too.

[Mark @mark15631]

@thsottiaux iPhone remote！

[Mark @mark15631]

@ChrisMasterton @adocomplete Because when you subscribe max and you will find that the so called 20 times , is not that much. Especially for opus 4.1. Easily to be used.

[Mark @mark15631]

@thomasdijkstra @pvncher thanks, it’s really helpful!