1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
# Effort capabilities checked 2026-09-23; criteria are uncalibrated routing hypotheses.
# The `anthropic` model ladder for `omni-dev ai jev route` (#1779, #1820):
# sonnet (Claude Sonnet) / opus (Claude Opus).
#
# A rung names a model family and its description a capability level, not a
# version: Jev reads only the description, and mapping a rung to a current
# model is the consumer's job. See docs/jev.md, "Rungs, model versions and
# effort".
#
# The `fable` rung is commented out, not deleted (#1903): Anthropic's own
# comparison reports Claude Opus 5.5 (released 2026-09-22) performing "at the
# level of Claude Fable 5.1 on most work" and beating it on published coding
# benchmarks, so a distinct, more-capable-than-`opus` class is no longer
# supported by the evidence. Restore it by uncommenting the block below if
# that changes. This is a routing-ladder change only: `claude-fable-5-1`
# remains registered in src/templates/models.yaml for other, unrelated uses.
#
# Tiers are listed from least to most capable; `route` ranks them by position.
# The descriptions are the text validated against jev-1.13.0 — rewording them
# shifts every answer, so re-run the evaluation before changing a word. The
# other providers' ladders (`jev-route-tiers-<provider>.yaml`) still carry
# three rungs and reuse these descriptions rung for rung where they overlap —
# an edit to a rung kept here is an edit to every ladder's matching rung.
# `--tiers FILE` replaces the ladder (at least two tiers, unique names; `none`
# is reserved for the design stage).
tiers:
- name: sonnet
models:
- name: claude-sonnet-5
verified: "2026-09-23"
sources:
effort:
kind: configurable
control: output_config.effort
supported_levels:
levels:
- name: low
description: >-
The stage is short, local and familiar, with explicit requirements and an obvious solution or review checklist.
- name: medium
description: >-
The stage needs several reasoning steps and checking ordinary interactions, with a clear approach and bounded scope.
- name: high
description: >-
The stage needs sustained reasoning about subtle behavior, multiple components, non-obvious failure modes or competing approaches.
- name: xhigh
description: >-
The stage needs extended exploration, long-horizon implementation or review, or repeated testing of interacting hypotheses beyond high effort.
- name: max
description: >-
The stage approaches this model's capability limit and requires its deepest supported analysis; lower effort would leave major unresolved risks.
description: >-
Reliable at executing a clear specification and following patterns that already exist in the
codebase. Tends to miss non-obvious interactions between parts and does not question the
specification.
- name: opus
models:
- name: claude-opus-5-5
verified: "2026-09-23"
sources:
effort:
kind: configurable
control: output_config.effort
supported_levels:
levels:
- name: low
description: >-
The stage is short, local and familiar, with explicit requirements and an obvious solution or review checklist.
- name: medium
description: >-
The stage needs several reasoning steps and checking ordinary interactions, with a clear approach and bounded scope.
- name: high
description: >-
The stage needs sustained reasoning about subtle behavior, multiple components, non-obvious failure modes or competing approaches.
- name: xhigh
description: >-
The stage needs extended exploration, long-horizon implementation or review, or repeated testing of interacting hypotheses beyond high effort.
- name: max
description: >-
The stage approaches this model's capability limit and requires its deepest supported analysis; lower effort would leave major unresolved risks.
description: >-
Reliable at reasoning across many files, subtle library or platform semantics, concurrency and
ordering, once the overall direction is set. Can under-explore a genuinely open design space.
# Retired #1903 — see the header comment above. Kept for reference/rollback.
# - name: fable
# models:
# - name: claude-fable-5-1
# verified: "2026-09-23"
# sources: [https://platform.claude.com/docs/en/build-with-claude/effort]
# effort:
# kind: configurable
# control: output_config.effort
# supported_levels: [low, medium, high, xhigh, max]
# levels:
# - name: low
# description: >-
# The stage is short, local and familiar, with explicit requirements and an obvious solution or review checklist.
# - name: medium
# description: >-
# The stage needs several reasoning steps and checking ordinary interactions, with a clear approach and bounded scope.
# - name: high
# description: >-
# The stage needs sustained reasoning about subtle behavior, multiple components, non-obvious failure modes or competing approaches.
# - name: xhigh
# description: >-
# The stage needs extended exploration, long-horizon implementation or review, or repeated testing of interacting hypotheses beyond high effort.
# - name: max
# description: >-
# The stage approaches this model's capability limit and requires its deepest supported analysis; lower effort would leave major unresolved risks.
# description: >-
# Strongest at open-ended design and research: choosing between architectures, working in
# unfamiliar territory with no precedent in the codebase, anticipating failure modes, and
# security-critical judgement.