OCG Bench

Task-domain model decisions, expressed as cognitive transitions.

Evidence-boundTask-domain routingModel transition graph

TL;DR

C'Sonnet 4.6 sufficient for isolated bugs under ~300 lines. Escalate to Opus 4.8 when bug involves shared state or requires multi-step causal chain.

Defaultclaude/sonnet-4.6

avoidopenai/gpt-5.5

Decision Card

bug is isolated, context fits one prompt, cost matters

claude/sonnet-4.6

bug involves subtle state, multiple suspects, correctness is critical

claude/opus-4.8

you need precise root-cause location, not symptom description

openai/gpt-5.5

terminal-heavy debugging workflow

openai/gpt-5.5

claude/opus-4.8

Claude Opus 4.8

High-stakes reasoning and uncertainty-sensitive work

SWE-Bench Pro leader
lower hallucination exposure
strong OSWorld verified score

Use when silent failure costs more than latency or price.

claude/sonnet-4.6

Claude Sonnet 4.6

Default high-volume operator

developer-preferred coding workflows
enterprise document reading
cost-sensitive triage

Escalate when the task needs cross-source coherence or implicit constraint tracking.

openai/gpt-5.5

GPT-5.5

Terminal and input-heavy workflow specialist

Terminal-Bench 2.1 leader
CLI-integrated pipelines
input-volume economics

Avoid as the sole source of truth for unverified extraction or citation-heavy tasks.

DEMO · EVENT

Fable 5 is unavailable. What now?

A concrete decision graph showing how model availability changes task routing.

Open demo

Task Domains

All 10 domains from the content master file.

D1 selected

Code Engineering / Single-File Debug

@code_engineering/single_file_debug

5transitions

4decisions

3evidence

Cognitive Transition Graph

c changes through r@D into C'; relation labels come from the selected domain.

cMy Python function returns wrong output on second call

acceptedidentify@single_file_debugclaude/opus-4.8

acceptedvalidate@single_file_debugclaude/opus-4.8

C'Sonnet 4.6 sufficient for isolated bugs under ~300 lines. Escalate to Opus 4.8 when bug involves shared state or requires multi-step causal chain.

Relation schema:c > r@D > C'identifyvalidate

Cognitive Transitions

by r@D and model

r@D	Model	Result	C'	Evidence
`identify@single_file_debug`	claude/opus-4.8	accepted	classifies bug type accurately, including subtle state issues	E-GLOBAL-001, E-D1-001
`identify@single_file_debug`	claude/sonnet-4.6	accepted	matches Opus for common patterns; preferred by devs	E-D1-002
`identify@single_file_debug`	openai/gpt-5.5	bounded	accurate classification but sometimes stops at symptom level	E-D1-003
`validate@single_file_debug`	claude/opus-4.8	accepted	more likely to flag own code flaws vs prior Opus; stronger dishonest-summary resistance	E-GLOBAL-004
`validate@single_file_debug`	claude/sonnet-4.6	bounded	reliable for isolated bugs; can miss subtle flaws silently vs Opus 4.8	E-GLOBAL-004

Triple Block

c / r / C'

My Python function returns wrong output on second call

locate@code_engineering/single_file_debug

Sonnet 4.6 sufficient for isolated bugs under ~300 lines. Escalate to Opus 4.8 when bug involves shared state or requires multi-step causal chain.

cost-benefit threshold at ~300 lines / single-call scope

Boundary Conditions

cost-benefit threshold at ~300 lines / single-call scope

Accepted

claude/opus-4.8 classifies bug type accurately, including subtle state issues

claude/sonnet-4.6 matches Opus for common patterns; preferred by devs

claude/opus-4.8 more likely to flag own code flaws vs prior Opus; stronger dishonest-summary resistance

Bounded

openai/gpt-5.5 accurate classification but sometimes stops at symptom level

claude/sonnet-4.6 reliable for isolated bugs; can miss subtle flaws silently vs Opus 4.8

Rejected

No Rejected transitions in this domain.

Evidence Ledger

Global benchmarks, claims, and update actions from the master file.

35% medium-confidence claims

Benchmarks

E-GLOBAL-001: Opus 4.8 SWE-Bench Pro 69.2% vs GPT-5.5 58.6%
E-GLOBAL-002: GPT-5.5 leads Terminal-Bench 2.1: 78.2% vs Opus 4.8 74.6%
E-GLOBAL-007: Opus 4.8 OSWorld-Verified 83.4% vs GPT-5.5 78.7%

Claims

E-GLOBAL-003: Opus 4.8 hallucination rate 35.9% vs GPT-5.5 86% on AA-Omniscience
E-GLOBAL-004: Opus 4.8 is 4x less likely than Opus 4.7 to miss code flaws; 17x less likely than Sonnet 4.6 to produce dishonest summaries
E-GLOBAL-005: Sonnet 4.6 matches Opus 4.6 on OfficeQA enterprise document reading
E-GLOBAL-006: Opus 4.8 Finance Agent v2: highest recorded score
E-GLOBAL-008: Opus 4.8 Legal Agent Benchmark: highest recorded, first to break 10% all-pass
E-GLOBAL-009: Sonnet 4.6 vs GPT-5.5: Sonnet leads on coding; GPT-5.5 leads aggregate (89 vs 82)

Actions

Update on model version change
Update on new benchmark
Update on capability event

Routing Summary

Opus 4.8

Multi-file code requiring cross-state coherence

Legal analysis beyond first-pass triage

Citation-critical document extraction

Sonnet 4.6

Single-file debugging under ~300 lines

High-volume document summarization

JSON extraction from labeled inputs

GPT-5.5

Terminal-heavy agentic coding

CLI-integrated pipelines

Input-heavy batch tasks where cost is primary constraint

Human Gates

Security code review final approval

Binding legal decisions

Financial outputs for reporting or audit