Simone Civetta

That's from @swiftconn, returning November 2-3!

I’m not an unconditional Claude Code fan, far from it. But Dynamic Workflows is one of the best features this space has seen since Skills (and maybe even before).
I just can't understand why Codex didn't implement it since day 2. Massive feature gap.

We're all very excited about this edition and so proud to have this event taking place once more in the Ville Lumière!

"Opus 4.7 is a disgruntled employee." "No, it’s the first model that truly gets what I want."

Reviews and feedback on the latest release by Anthropic are divergent, to say the least. But they're interesting, as they highlight how hard is to evaluate - and as a result - improve these models in a way that feels strictly better for everyone..

Two charts help us grab a more definite picture:

- Arena-dot-ai showing Opus 4.7 scoring better in some areas and much worse in others.
- A steep drop in MRCR v2 recall in Opus 4.7’s own System Card - MRCR v2 being the same eval Anthropic used only weeks ago to support Opus 4.6’s 1M-context quality.

For full disclosure Anthropic engineers have recently stated that MRCR v2 is now obsolete - but that's almost secondary: if an eval can move from headline evidence to outdated in a few weeks, the measurement layer is still fragile.

My read is that Anthropic models, especially if compared to OpenAI's, have long suffered from a somewhat worse instruction following but a better capability in understand the user's true goals. And I'm sure nobody at Anthropic is lying when they mention that the team invested much in improving the former.

That said, instruction following is a double-edged sword, as it makes the model more easily steerable, but also more prone to following the _wrong_ instructions, like the ones included in a suboptimal system prompt, or in a messy CLAUDE.md. As a result, a prompt or a skill that worked perfectly on 4.6 can contain nuances that the model will follow more closely, thus impacting existing habits and workflows. Truth be told, breaking existing prompts is a common issue with model releases, no matter the lab, and a risk even Anthropic’s team had warned about.

Another pitfall is the new adaptive thinking behavior; at medium effort, the model often skips reasoning, negatively impacting reliability. Note hat the effort parameter acts as a ceiling, not a floor, for the thinking budget. To maintain high-quality reasoning, Boris Cherny’s proposes using xhigh effort mode (or max for critical work) to prevent the model from under-thinking. But, inevitably, this will increase your plan use.

I also suspect we’re seeing a disconnect caused either by an internal dogfooding bias that doesn't reflect how most of us use the model, or by a shift toward power efficiency that leans too heavily on the thinking budget for steering. Probably, a mix of both, resulting in a model that doesn't translate as well to real-world workflows as they might have hoped.

Either way, that's one more reminder we should treat model drift as a constant, and design our prompts, harnesses, and workflows accordingly.

I'm absolutely not an AI doomer but this is a very interesting take.