From everything we know so far, Opus 4.5 seems to be the best-...
Sam Bowman@sleepinyourhat
24 views
Dec 08, 2025
~2 min read
Advertisement
1
From everything we know so far, Opus 4.5 seems to be the best-aligned model out there in a bunch of ways.
I follow the training process closely as part of my work on alignment evaluations. Here's my guess about the two things that are most responsible for making 4.5 special. 🧵
I follow the training process closely as part of my work on alignment evaluations. Here's my guess about the two things that are most responsible for making 4.5 special. 🧵
2
First: The spec.
@AmandaAskell and collaborators prepared a great deal of writing on *what it means to be a good Claude.*
I think the resulting model spec is more coherent and nuanced and mature than anything that's been used in model training to date, here or elsewhere.
@AmandaAskell and collaborators prepared a great deal of writing on *what it means to be a good Claude.*
I think the resulting model spec is more coherent and nuanced and mature than anything that's been used in model training to date, here or elsewhere.
3
As the internet seems to have noticed, we also used this written material in a way that seems fairly distinct.
4
With Opus 4.5, we didn't just use this spec to produce *rewards* or *demonstrations* that teach the model the behaviors described in the spec. We also trained the model directly on the text of the spec.
Amanda Askell@AmandaAskell
I just want to confirm that this is based on a real document and we did train Claude on it, including in SL. It's something I've been working on for a while, but it's still being iterated on and we intend to release the full version and more details soon.
5
It's becoming increasingly clear that a model's self-image or self-concept has some real influence on how its behavior generalizes to novel settings.
6
This kind of direct supervised training on text about the model's character seems to be an especially direct way of influencing the model's self-concept.
7
Second: Alignment researchers are involved in every part of training.
We don't have a clear split between alignment research and applied finetuning. Alignment-focused researchers are deeply involved in designing and staffing production training runs.
We don't have a clear split between alignment research and applied finetuning. Alignment-focused researchers are deeply involved in designing and staffing production training runs.
8
We watch our models closely during the parts of training that are most responsible for setting their behavior and character, and regularly make adjustments in response to what we see.
9
This work has at times involved changes to most parts of our training pipeline, and to most aspects of our models' alignment and behavior.
10
A cook who knows what to look for, and is constantly adjusting their technique as they prepare a dish, is going to get better results than someone who rigidly follows a recipe.
11
This is hard, though: It demands a big hybrid team that can respond quickly with engineering savvy and research intuition and creativity and taste.
12
The company has been getting better at this with each model launch, and I think it went especially well with Opus 4.5. I've been really impressed by the speed and quality of some of the alignment and model-behavior research that has gotten done *during* recent training runs.
13
There are many, many people involved in aspects of this hands-on alignment work, but @sprice354_, Jon Kutasov, @MinaeKwon, Monty Evans, and Richard Dargan have played especially central roles.
