We have our OpenHands Index. We have

Juan joining us today and I will share

the link to OpenHands Index, but I'm

going to stop presenting so that Juan

can share his screen and go over the

latest and the greatest models.

>> Yes, thank you so much, Amy. So,

um here's the index and you will see it.

>> Yes.

>> Okay, thank you. And um

so we have a a graph um

uh vertical axis is a score, horizontal

axis is a cost. So, it the best model

would be here, which doesn't exist, and

the worst model would be here, which we

wouldn't pay attention to.

Um I want to start by talking about

Fable um by saying that I'm not going to

talk about Fable. Uh Fable

did a great performance at a very high

cost and then disappear. So,

we cannot use it, so we will only talk

about models we can actually use.

Um all models we can actually use uh on

the Pareto line, that means the line

that's has uh the best uh

cost for a given performance. Um we can

see that Opus 48 and Opus 47 are very

close.

So, one could ask which one to use.

Like, they are Opus 48 seems a little

bit more expensive uh but performs a

little bit better. I'm going to break it

down per benchmark so you can decide

which one is better for you for your use

case.

And then on the Pareto line, we have a

new contender, which is Minimaxen 3, and

it's performing very good. Um I will

also break it down across all

benchmarks.

So, first,

issue resolution. This is your standard

hey, OpenHands fix this issue. We can

see that Opus 48 is on the Pareto line

and Opus 47 is performing worse at a

higher uh cost. Um so, if you're just

calling OpenHands late to fix issues,

like that's most of what they do, uh

going for Opus 48 uh would be your

default, not Opus 47.

And Minimaxen 3 uh performs very well.

If if you can pay attention, you notice

that Minimaxen 3 is like in the same

line as Opus 47. It actually performs

like a hair thin worse. They're both at

76. Um

Opus 46 is 76.8 and M M Minimax 3 76.4.

So, less than uh

than half a point. Um but it's very good

news that uh open source models are

performing uh are catching up with uh a

model that at the moment was amazing

like Opus 46.

Moving on towards uh testing, we uh

we then start testing separate from uh

each resolution. And here we can see

that uh Opus performs uh

a tiny bit better at Opus 4 uh although

48 is a tiny bit better at Opus 47.

Still enough to be the the the straight

to the choice. Um and Minimax performs

almost as well as Opus 48. So, if you

are doing test, if you're like just

doing test all day, um

you might want to use Minimax instead of

Minimax 3 instead of Opus 48 because

it's very very close and it's 1/4 of the

cost.

Um moving on towards front end, this is

all your UI fixes, um polishing

interface, anything graphical, and

looking at images and describing what's

inside the image. So,

we have our Pareto line. You can use

Opus. Minimax is also in the if you

can't going to see it because it's uh at

at

it's almost the same as Deep B4 Pro, but

they are both on the Pareto line. Um

my advice on front end is pick the one

you actually need. So, like

if you're testing images, do you need a

full description or do you need like uh

a high-level one? If you need a full

description, I would continue going for

Opus. If you're okay with something uh a

little more high-level, not that much

detail, um

Minimax 24 perform very well at uh much

more accessible cost.

Um then we are going to move to green

field. We call green field like by

coding, like when you are starting

something from a scratch. New tool, new

application, uh starting from zero. Um

Minimax is also on the Pareto line. Um

and we have

Opus

here. So, uh Opus 48 8

Oh,

I seem to Okay.

>> [snorts]

>> Uh

Opus 4 8 performs

uh a little bit better than um

than Opus 4 7 at a little bit more cost.

So, uh

we can see a history here. Opus 4 6 um

was more expensive than Opus 4 7, but

they they perform the same. Now, Opus 4

8 performs better, but at the same cost

as Opus 4 6. And curiously, Fable did

not perform that well. We have some

theories, but we're not going to talk

about them because we cannot choose

Fable anyway, so what's the point? Um

So, if you're about to go with 4 8, it

looks great. GPT-5 4 also keeps looking

great. GPT-5.5

It did have some regression.

Um

So, yeah. Either Opus 4 8 or GPT-5.5 or

Minimax and

Deep Seek Deep 4 Pro, they're all

looking great.

And last but not least, information

gathering. This is the one you do when

you have to create a a few a few

databases, feel know if something

happened in the real world, get an

official link, get evidence. Um GPT-5 is

performing the best here. Uh that's my

go-to when I really really need to find

something out. Um Minimax M3 is not in

the borderline first first of the

benchmarks in which it's not in the

borderline. Uh it's very close, but

Queen 3 uh 6 plus uh beats it.

So, you can use either of those.

Um we have improved our skills on that

end to actually get the link. So, all of

them should perform better as well um

than they used to before. Um

So, yeah. So, uh my advice, use Opus 4 8

if you really need performance for

almost everything, maybe not for

information gathering, and Minimax M3 uh

looks great overall. Um

So, yeah. Uh that's that's the news on

the index for this week.