1
00:00:00,000 --> 00:00:03,406
Welcome to the show - James,
I need to start with a setting that feels

2
00:00:03,406 --> 00:00:08,484
like finding a missing gear on an old
trail bike: Anthropic just added

3
00:00:08,484 --> 00:00:12,042
xhigh for Claude Code on Opus 4.

4
00:00:08,484 --> 00:00:12,042
7,

5
00:00:12,042 --> 00:00:17,606
and it's now the DEFAULT.The
word default is doing a lot of

6
00:00:17,606 --> 00:00:18,276
work there.

7
00:00:18,276 --> 00:00:20,596
xhigh sits between high and max,

8
00:00:20,596 --> 00:00:20,838
right?

9
00:00:20,838 --> 00:00:24,652
Which tells you the old jump from high to
max was probably too chunky

10
00:00:24,652 --> 00:00:27,880
if people needed a whole new rung in the
ladder.

11
00:00:24,652 --> 00:00:27,880
Exactly.

12
00:00:27,880 --> 00:00:31,960
High could be a bit too shallow for
proper agentic coding work,

13
00:00:31,960 --> 00:00:34,282
but max could feel like, mate,

14
00:00:34,282 --> 00:00:38,682
turning a race car onto a gravel road -
heaps of token burn,

15
00:00:38,682 --> 00:00:43,062
more latency, and suddenly you're paying
premium prices for a job that

16
00:00:43,062 --> 00:00:47,230
didn't need the full orchestra.The phrase I latch onto

17
00:00:47,230 --> 00:00:48,645
there is agentic work.

18
00:00:48,645 --> 00:00:53,040
Because this isn't "write me a helper
function." This is more like a

19
00:00:53,040 --> 00:00:57,376
multi-step schema migration where step
three only makes sense if step

20
00:00:57,376 --> 00:01:01,200
two didn't break some hidden assumption
from step one.

21
00:00:57,376 --> 00:01:01,200
Yeah,

22
00:01:01,200 --> 00:01:02,360
that's the sweet example.

23
00:01:02,360 --> 00:01:07,624
Or a subtle PR security review - not the
obvious "hey, you forgot auth"

24
00:01:07,624 --> 00:01:11,939
stuff, but the sneaky edge case where one
permission check looks fine

25
00:01:11,939 --> 00:01:16,848
alone, then you trace it across three
files and realize the whole thing

26
00:01:16,848 --> 00:01:18,275
leaks sideways.

27
00:01:18,275 --> 00:01:21,157
That's where high might not dig deep
enough,

28
00:01:21,157 --> 00:01:25,160
but max is a bit of a
sledgehammer.And that's the

29
00:01:21,157 --> 00:01:25,160
tension

30
00:01:25,160 --> 00:01:28,800
right away: is xhigh the missing sweet
spot...

31
00:01:28,800 --> 00:01:33,362
or is it one more dial that makes
developers FEEL responsible while the

32
00:01:33,362 --> 00:01:38,679
bill quietly creeps upward in the
background?Yep.

33
00:01:38,679 --> 00:01:41,793
The old "I've optimised the workflow"
story,

34
00:01:41,793 --> 00:01:44,512
and then you check the tab later and
nearly spit your coffee.

35
00:01:44,512 --> 00:01:47,065
I've done that with cloud defaults before.

36
00:01:47,065 --> 00:01:51,006
You tick one innocent-looking box and
suddenly your app's running like

37
00:01:51,006 --> 00:01:53,800
it's servicing the Olympics.

38
00:01:51,006 --> 00:01:53,800
Mine was logging.

39
00:01:53,800 --> 00:01:56,279
Not AI - logging.

40
00:01:56,279 --> 00:01:59,281
I left a verbose mode on for, I think,

41
00:01:59,281 --> 00:02:03,081
a weekend, and by Monday I had a bill
that looked like the app had developed

42
00:02:03,081 --> 00:02:04,281
a gossip problem.

43
00:02:04,281 --> 00:02:10,800
So when I hear "xhigh is now the default
for Opus 4.

44
00:02:04,281 --> 00:02:10,800
7 in Claude Code,"

45
00:02:10,800 --> 00:02:15,879
I hear convenience and risk in the same
sentence.And to be fair,

46
00:02:15,879 --> 00:02:17,796
there is real convenience here.

47
00:02:17,796 --> 00:02:22,036
Anthropic also changed the workflow
around effort selection.

48
00:02:22,036 --> 00:02:25,555
If you run /effort with NO arguments now,

49
00:02:25,555 --> 00:02:28,517
it opens an arrow-key slider with low,

50
00:02:28,517 --> 00:02:31,796
medium, high, xhigh, and max.

51
00:02:31,796 --> 00:02:35,600
You don't have to remember the exact
string and type it from memory like

52
00:02:35,600 --> 00:02:40,760
you're entering a cheat code from 2004.

53
00:02:35,600 --> 00:02:40,760
The arrow-key slider is small,

54
00:02:40,760 --> 00:02:43,157
but that's a legit UX shift.

55
00:02:43,157 --> 00:02:46,994
Because when something becomes
interactive instead of memorized,

56
00:02:46,994 --> 00:02:48,512
people actually USE it.

57
00:02:48,512 --> 00:02:55,320
It's the difference between "there is a
setting" and "this setting exists in my

58
00:02:48,512 --> 00:02:55,320
brain when I need it."Spot

59
00:02:55,320 --> 00:02:55,560
on.

60
00:02:55,560 --> 00:02:57,479
And there are a few entry points.

61
00:02:57,479 --> 00:03:02,272
You can launch straight from the CLI with
claude --effort xhigh.

62
00:03:02,272 --> 00:03:04,193
You can pick it in the model picker.

63
00:03:04,193 --> 00:03:07,955
Or you can set it persistently through
/config,

64
00:03:07,955 --> 00:03:16,567
which now writes to ~/.claude/settings.

65
00:03:07,955 --> 00:03:16,567
json.~/.claude/settings.

66
00:03:07,955 --> 00:03:16,567
json

67
00:03:16,567 --> 00:03:20,104
is the memorable bit for me, because
that's where this stops being a one-off

68
00:03:20,104 --> 00:03:21,878
choice and becomes policy.

69
00:03:21,878 --> 00:03:25,875
A team can say: for serious Opus 4.

70
00:03:21,878 --> 00:03:25,875
7 sessions,

71
00:03:25,875 --> 00:03:27,635
our default is xhigh.

72
00:03:27,635 --> 00:03:32,113
Then nobody has to remember to re-assert
it at the top of every new task.

73
00:03:27,635 --> 00:03:32,113
And

74
00:03:32,113 --> 00:03:33,443
that's a real time saver.

75
00:03:33,443 --> 00:03:35,123
Tiny, but real.

76
00:03:35,123 --> 00:03:40,477
Anyone who's ever reset the same dev
environment setting ten times knows

77
00:03:40,477 --> 00:03:45,049
defaults are where time either gets saved
or quietly bled out.

78
00:03:45,049 --> 00:03:50,110
It's like setting your editor to use
spaces or tabs - sounds boring until

79
00:03:50,110 --> 00:03:55,188
you waste an afternoon fixing the
mess.But here's the catch: xhigh

80
00:03:55,188 --> 00:03:58,043
is Opus 4.

81
00:03:55,188 --> 00:03:58,043
7-only.

82
00:03:58,043 --> 00:03:59,882
Not Sonnet, not Haiku.

83
00:03:59,882 --> 00:04:03,780
And if you switch models,
the effort setting resets to whatever that

84
00:04:03,780 --> 00:04:05,470
model actually supports.

85
00:04:05,470 --> 00:04:08,588
So this is not a universal "Claude got
smarter" button.

86
00:04:08,588 --> 00:04:13,439
It's a specific mode on a specific model.

87
00:04:08,588 --> 00:04:13,439
That's an important little gotcha.

88
00:04:13,439 --> 00:04:16,320
Because if you're bouncing between models,

89
00:04:16,320 --> 00:04:19,487
you might think you're still in xhigh
land when you've actually dropped

90
00:04:19,487 --> 00:04:21,675
back to a different supported effort
level.

91
00:04:21,675 --> 00:04:23,035
Easy mistake.

92
00:04:23,035 --> 00:04:24,797
Very human mistake.

93
00:04:24,797 --> 00:04:27,597
Even a kangaroo could trip over fresh
code there.

94
00:04:27,597 --> 00:04:31,712
And Anthropic is
pretty explicit about the intended

95
00:04:31,712 --> 00:04:35,228
use: intelligence-sensitive agentic tasks.

96
00:04:35,228 --> 00:04:36,589
That's their lane for it.

97
00:04:36,589 --> 00:04:40,982
But "more reasoning" almost always cashes
out as more tokens and more

98
00:04:40,982 --> 00:04:41,632
latency.

99
00:04:41,632 --> 00:04:45,362
There isn't some magic free lunch where
the model thinks harder for the

100
00:04:45,362 --> 00:04:47,640
same price and same speed.

101
00:04:45,362 --> 00:04:47,640
Right.

102
00:04:47,640 --> 00:04:50,845
Better reasoning chains usually mean
longer chains.

103
00:04:50,845 --> 00:04:54,120
More internal work,
more generated analysis,

104
00:04:54,120 --> 00:04:54,840
more waiting.

105
00:04:54,840 --> 00:04:58,602
So if you're using xhigh for every little
file tweak,

106
00:04:58,602 --> 00:05:03,025
you're basically bringing a crane to move
a desk chair.

107
00:04:58,602 --> 00:05:03,025
The budget surprise

108
00:05:03,025 --> 00:05:07,663
gets sharper because xhigh is now the
default for Max subscribers in Auto

109
00:05:07,663 --> 00:05:07,999
mode.

110
00:05:07,999 --> 00:05:11,884
That means some people may see higher
spend without ever making a conscious

111
00:05:11,884 --> 00:05:14,239
"yes,
I want more expensive reasoning" decision.

112
00:05:14,239 --> 00:05:18,577
The default made it for them.
And that's where accountability

113
00:05:18,577 --> 00:05:19,161
comes in.

114
00:05:19,161 --> 00:05:22,122
If the tool gets better by default,

115
00:05:22,122 --> 00:05:23,000
lovely.

116
00:05:23,000 --> 00:05:26,520
If the spend rises by default too,

117
00:05:26,520 --> 00:05:28,121
you've gotta look.

118
00:05:28,121 --> 00:05:33,639
Anthropic's safeguard here is /usage -
that replaces the old /cost and

119
00:05:33,639 --> 00:05:34,998
/stats commands.

120
00:05:34,998 --> 00:05:40,581
So /usage is now the place to check
whether your shiny better defaults

121
00:05:40,581 --> 00:05:47,300
are actually changing consumption
patterns./usage replacing both /cost

122
00:05:47,300 --> 00:05:51,195
and /stats is good design, honestly.

123
00:05:51,195 --> 00:05:54,316
One command, one dashboard in your head.

124
00:05:54,316 --> 00:05:59,196
Because if you split "how much am I
using?" from "what is this costing me?"

125
00:05:59,196 --> 00:06:03,545
people tend to check neither until it's
too late.So

126
00:06:03,545 --> 00:06:05,715
when is xhigh actually worth it?

127
00:06:05,715 --> 00:06:09,556
I'd say the strong cases are where
reasoning compounds.

128
00:06:09,556 --> 00:06:11,434
Migrating legacy code.

129
00:06:11,434 --> 00:06:13,552
Designing a new API surface.

130
00:06:13,552 --> 00:06:16,513
Reviewing PRs with subtle security
implications.

131
00:06:16,513 --> 00:06:20,233
Large-codebase reviews where one wrong
assumption cascades into three

132
00:06:20,233 --> 00:06:27,046
more and suddenly you're patching ghosts.

133
00:06:20,233 --> 00:06:27,046
The phrase "one wrong assumption cascades
into three more"

134
00:06:27,046 --> 00:06:28,353
is exactly it.

135
00:06:28,353 --> 00:06:32,353
That's the kind of task where deeper
effort can pay for itself,

136
00:06:32,353 --> 00:06:36,675
because catching the first bad assumption
prevents a whole chain of bogus

137
00:06:36,675 --> 00:06:37,220
edits.

138
00:06:37,220 --> 00:06:39,697
Especially in legacy refactors,

139
00:06:39,697 --> 00:06:46,402
where every module has some weird
haunted-house backstory.Ah

140
00:06:46,402 --> 00:06:53,211
yes, the beloved legacy comment from
2017: "don't touch this." Terrifying.

141
00:06:53,211 --> 00:06:55,929
But for short interactive tasks,

142
00:06:55,929 --> 00:06:58,971
one-off file edits, or quick questions?

143
00:06:58,971 --> 00:07:01,211
xhigh is usually overkill.

144
00:07:01,211 --> 00:07:04,249
The latency tax alone can outweigh the
quality bump.

145
00:07:04,249 --> 00:07:06,652
Sometimes you don't need a
philosopher-mechanic.

146
00:07:06,652 --> 00:07:09,042
You just need a spanner.

147
00:07:06,652 --> 00:07:09,042
Let me try to say it back.

148
00:07:09,042 --> 00:07:12,562
If the task is branching,

149
00:07:12,562 --> 00:07:17,199
stateful, and each decision changes the
next decision,

150
00:07:17,199 --> 00:07:18,643
xhigh makes sense.

151
00:07:18,643 --> 00:07:21,921
If the task is narrow, obvious,

152
00:07:21,921 --> 00:07:25,758
and easy to verify fast,
go cheaper and faster.

153
00:07:25,758 --> 00:07:31,279
Is that basically the rule?

154
00:07:25,758 --> 00:07:31,279
Almost - the missing bit is consequence.

155
00:07:31,279 --> 00:07:34,802
A task can be small but high-stakes.

156
00:07:34,802 --> 00:07:38,073
Like a short PR with subtle security
implications.

157
00:07:38,073 --> 00:07:42,506
That's still a candidate for xhigh
because the cost of being wrong is

158
00:07:42,506 --> 00:07:43,870
bigger than the wait.

159
00:07:43,870 --> 00:07:45,430
So it's not just size.

160
00:07:45,430 --> 00:07:49,790
It's dependency chain plus blast
radius.That's the real

161
00:07:49,790 --> 00:07:50,805
skill then.

162
00:07:50,805 --> 00:07:55,138
Not knowing the feature exists - knowing
when deeper reasoning is actually

163
00:07:55,138 --> 00:07:58,645
needed, versus when you're just soothing
your own anxiety with a more

164
00:07:58,645 --> 00:08:03,351
expensive setting.

165
00:07:58,645 --> 00:08:03,351
And that's the question I can't shake:
once a team gets

166
00:08:03,351 --> 00:08:06,965
used to a just-in-case higher-effort
default,

167
00:08:06,965 --> 00:08:09,846
do they become more careful coders...

168
00:08:09,846 --> 00:08:13,637
or just more expensive ones?
Check /usage and you'll find out

169
00:08:13,637 --> 00:08:14,401
pretty fast.

170
00:08:14,401 --> 00:08:16,241
See you next time.