1
00:00:00,119 --> 00:00:01,439
You know what is wild, Lachlan?

2
00:00:02,039 --> 00:00:05,679
This episode of our show is actually
sponsored by Jellypod AI,

3
00:00:06,199 --> 00:00:10,359
which is, uh, pretty fitting because we
are talking about serious developer tools

4
00:00:10,399 --> 00:00:10,679
today.

5
00:00:11,259 --> 00:00:14,840
I mean, did you see the Claude Code
version two point one point two hundred and

6
00:00:14,880 --> 00:00:16,340
nineteen update that just dropped?

7
00:00:16,859 --> 00:00:17,920
It is massive.

8
00:00:18,399 --> 00:00:21,139
They made claude opus five the default
Opus model.

9
00:00:21,619 --> 00:00:27,319
And we are talking about a native, under
the hood, one million token context window

10
00:00:27,399 --> 00:00:28,680
directly in your terminal.

11
00:00:29,407 --> 00:00:30,807
One million tokens.

12
00:00:31,767 --> 00:00:33,387
Directly in the command line interface.

13
00:00:33,947 --> 00:00:37,027
That is like, I mean, okay, you know I
love my old motorbikes,

14
00:00:37,107 --> 00:00:37,247
right?

15
00:00:37,787 --> 00:00:40,867
Imagine someone dumps a vintage postie
bike on your driveway,

16
00:00:41,427 --> 00:00:43,007
but it is entirely disassembled.

17
00:00:43,427 --> 00:00:47,127
Every single bolt, gasket, and piston ring
is just loose in this giant,

18
00:00:47,247 --> 00:00:48,347
rusted wooden crate.

19
00:00:48,827 --> 00:00:52,207
Before, with smaller context windows, you
could only reach in and grab,

20
00:00:52,287 --> 00:00:56,667
say, the carburetor and hope it fits the
spark plug you pulled out ten minutes ago.

21
00:00:57,227 --> 00:01:00,467
Now, with a million tokens, it is like
having the entire blueprint,

22
00:01:00,807 --> 00:01:04,827
the full history of every mechanic who
ever touched it, and every single part laid

23
00:01:04,888 --> 00:01:06,787
out on a clean workbench all at once.

24
00:01:07,287 --> 00:01:10,748
You can see the whole codebase, start to
finish, without the model forgetting where

25
00:01:10,787 --> 00:01:12,147
the main function even lives.

26
00:01:12,854 --> 00:01:13,615
Exactly!

27
00:01:14,134 --> 00:01:16,254
It is the whole garage in your head.

28
00:01:16,694 --> 00:01:22,015
But, okay, let us talk about the actual
math of this because a million tokens is not

29
00:01:22,074 --> 00:01:25,134
cheap if you are slamming the API
constantly.

30
00:01:25,694 --> 00:01:29,914
They introduced this new, uh, fast mode
pricing model for opus five.

31
00:01:29,954 --> 00:01:35,114
It is what they are calling ten, ten, ten,
fifty per million tokens.

32
00:01:35,654 --> 00:01:40,274
So, to break that down, that is ten
dollars per million tokens for input,

33
00:01:40,814 --> 00:01:46,255
ten dollars for cache writes, ten dollars
for cache reads, and then fifty dollars

34
00:01:46,375 --> 00:01:48,174
per million tokens for output.

35
00:01:48,348 --> 00:01:53,767
Wait, so ten dollars for input, write, and
read, and then fifty for output?

36
00:01:54,927 --> 00:01:58,367
That actually makes prompt caching the
absolute hero of this setup.

37
00:01:58,847 --> 00:01:59,807
Because, think about it.

38
00:02:00,227 --> 00:02:04,607
If you are running an AI agent, and it has
to scan your entire monorepo on every

39
00:02:04,708 --> 00:02:09,128
single turn, you would burn through your
API budget before you even finished writing

40
00:02:09,167 --> 00:02:10,267
a single unit test.

41
00:02:10,767 --> 00:02:13,927
But because the cache read rate is a flat
ten dollars per million,

42
00:02:14,427 --> 00:02:18,727
once that code is in the cache, those
follow up prompts, those little refactoring

43
00:02:18,787 --> 00:02:21,467
loops, they are incredibly cheap.

44
00:02:21,507 --> 00:02:25,248
It actually makes deep, iterative agent
work financially doable.

45
00:02:25,473 --> 00:02:25,812
Yes!

46
00:02:25,953 --> 00:02:26,433
You get it.

47
00:02:26,792 --> 00:02:30,753
The prompt caching is what makes a million
token context actually usable for daily

48
00:02:30,813 --> 00:02:33,012
coding instead of just a neat party trick.

49
00:02:33,553 --> 00:02:37,253
And if you want to set this up right now,
the CLI commands are super

50
00:02:37,312 --> 00:02:37,972
straightforward.

51
00:02:38,492 --> 00:02:44,572
You literally just type slash model or
slash config model equals claude opus five

52
00:02:44,652 --> 00:02:45,872
right in your terminal.

53
00:02:45,912 --> 00:02:48,072
And boom, you are on the new hotness.

54
00:02:48,652 --> 00:02:53,132
They also updated the slash fast mode, so
now it exclusively routes to Opus five and

55
00:02:53,212 --> 00:02:54,712
Opus four point eight.

56
00:02:54,753 --> 00:02:58,513
And they actually completely removed the
older Opus four point seven from the fast

57
00:02:58,572 --> 00:02:59,052
rotation.

58
00:03:00,382 --> 00:03:02,323
Ah, rest in peace, four point seven.

59
00:03:03,042 --> 00:03:03,862
You served us well.

60
00:03:05,142 --> 00:03:08,662
But, honestly, this is great because if
you have custom agent scripts running,

61
00:03:09,142 --> 00:03:13,503
the claude api skill has been updated to
automatically default to Opus five now

62
00:03:13,562 --> 00:03:13,862
anyway.

63
00:03:14,522 --> 00:03:18,322
So you do not even have to go in and
manually rewrite all your custom tooling.

64
00:03:18,642 --> 00:03:21,843
It just, it just works out of the box.

65
00:03:22,134 --> 00:03:22,334
Right.

66
00:03:22,394 --> 00:03:23,235
It is seamless.

67
00:03:23,394 --> 00:03:27,894
And speaking of things just working, there
are some really nice quality of life

68
00:03:27,974 --> 00:03:31,974
updates in this two point one point two
hundred and nineteen release.

69
00:03:32,054 --> 00:03:37,354
One that I am super happy about is sandbox
dot network dot strict allowlist.

70
00:03:37,394 --> 00:03:41,895
It is a new configuration that silently
drops any unauthorized external network

71
00:03:41,934 --> 00:03:42,534
requests.

72
00:03:43,114 --> 00:03:46,894
So instead of the agent constantly pausing
and throwing those annoying permission

73
00:03:46,934 --> 00:03:51,254
prompts because some package tried to ping
home, it just quietly blocks it and keeps

74
00:03:51,334 --> 00:03:51,594
running.

75
00:03:51,804 --> 00:03:54,644
Oh, that is huge for local development
security.

76
00:03:55,264 --> 00:03:56,244
No more prompt fatigue.

77
00:03:57,105 --> 00:04:01,625
They also added a DirectoryAdded hook,
which is going to be brilliant for anyone

78
00:04:01,744 --> 00:04:06,184
working with massive monorepos where
directories are getting spun up dynamically.

79
00:04:07,064 --> 00:04:10,164
And, oh, my absolute favorite minor fix.

80
00:04:10,805 --> 00:04:14,824
If you use Vim mode in the terminal, and
you are sitting on an empty normal mode

81
00:04:14,885 --> 00:04:18,844
prompt, pressing the left arrow key now
takes you straight back into the active

82
00:04:18,944 --> 00:04:19,464
agent view.

83
00:04:20,284 --> 00:04:25,404
It is such a tiny thing, but as a Vim
user, not getting stuck in that terminal limbo

84
00:04:25,465 --> 00:04:26,465
is a massive win.

85
00:04:27,759 --> 00:04:29,659
It is the little details that keep us
sane.

86
00:04:30,199 --> 00:04:33,339
Well, that is the quick run down on Claude
Code and Opus five.

87
00:04:33,739 --> 00:04:36,460
Happy coding, everyone, and we will catch
you in the next one.

88
00:04:36,583 --> 00:04:37,583
Too easy.

89
00:04:37,643 --> 00:04:38,663
Catch you later, mate.