Claude Just Got a Big Update (Opus 4.1)

Watch on YouTube

Now Playing

Transcript

77 segments

0:00

I guess it's new model week. Anthropic

0:02

just dropped a new version of Claude

0:04

Opus. This is Claude Opus 4.1. And yes,

0:08

it is obviously better than 4.0. Here's

0:11

the blog. Today, we're releasing Claude

0:13

Opus 4.1, an upgrade to Claude Opus 4 on

0:16

agentic task, real world coding, and

0:19

reasoning. We plan to release

0:20

substantially larger improvements to our

0:22

models in the coming weeks. Fantastic. I

0:26

love when a new model drops like Claude

0:28

4 because you know they're just going to

0:31

continue to iterate on it and make it

0:32

better and squeeze every ounce of

0:35

intelligence out of these foundation

0:37

models. Let's get right into the

0:38

benchmarks. So this is SWEBench

0:40

verified. Here's Sonnet 3.7 at 62.3.

0:44

Opus 4 jumped to 72.5 and Opus 4.1

0:48

squeezed out another 2% to get to 74.5%

0:53

on Sweetbench verified. It might not

0:55

seem like a lot, but every percentage

0:57

point matters. And according to the

0:59

post, it also improves Claude's in-depth

1:01

research and data analysis skills,

1:03

especially around detail tracking and

1:05

Agentic Search. I love to hear this

1:08

because the better it works within

1:09

Agentic Frameworks, the more

1:11

capabilities it's going to have.

1:13

Scaffolding is all you need. Let's look

1:15

at some more benchmarks. Here it is with

1:16

Terminal Bench, its ability to use the

1:18

terminal at 43.3 as compared to 39.2. So

1:23

another singledigit percentage bump here

1:25

is GPQA diamond which is graduate level

1:27

reasoning. It is a minor bump 80.9 as

1:30

compared to 79.6.

1:33

We have towbench. This is a gentic tool

1:35

use a very important benchmark for a

1:38

gentic frameworks. Here it is for retail

1:41

towbench 82.4 as compared to 81.4 a 1%

1:46

bump. But for airline it actually went

1:48

down 56% as compared to 59.6 six for

1:52

multilingual Q&A 89.5 as compared to

1:56

88.8

1:57

and visual reasoning another single

2:00

point bump and Amy 2025 a 2 and 1/2

2:03

point bump up to 78%. Now if we're

2:05

comparing all of these to OpenAI's 03

2:09

and Gemini 2.5 Pro we can see on Sweet

2:13

Bench it is beating both of them.

2:15

Terminal Bench same thing. Now, for GPQA

2:18

Diamond, it is being beat by both of

2:20

these other models, and it's more or

2:22

less the same for a Gentic Tool use. But

2:25

look at this. For high school math

2:27

competition, it is getting pretty

2:29

dominated, to be honest. 03 at 88.9,

2:32

Gemini 2.5 Pro at 88% and Opus 4.1 at

2:36

78%. None of these benchmarks really

2:39

matter. What really matters is when you

2:41

get in and you start using it, how does

2:43

it work? How does it perform? Right now,

2:46

Claude is still best known for being the

2:49

best coding model on the market. We'll

2:52

see if that changes soon, but for now,

2:53

it is kind of widely understood that

2:55

Claude is currently the best at agentic

2:58

coding with agent-driven development.

3:00

So, that's it. It's a very short blog

3:02

post. That's all the information I have

3:04

so far. I'm definitely going to be

3:05

testing it. And if you do, let me know

3:07

what you think. If you enjoyed this

3:09

video, please consider giving a like and

3:11

subscribe. and I'll see you in the next

Interactive Summary

Ask follow-up questions or revisit key timestamps.

This video covers the release of Claude Opus 4.1, highlighting its improvements in agentic tasks, coding, and reasoning over its predecessor. The host reviews various benchmarks, including SWEBench and Terminal Bench, noting both incremental performance gains and areas where the model trails behind competitors like OpenAI's o3 and Gemini 2.5 Pro.