A model in the suite · OpenAI

GPT-5.4

OpenAI · GPT · 2026-04-19

51/100

Strict suite averageLegacy 67 · 1 benchmark

Single-benchmark historical packet. The Car Wash scaffold has enough recoverable model identity for a packet, but multiple primary canary failures keep it in interesting-but-unreliable territory.

No transcript compaction

This run kept the entire transcript in context on every turn, with nothing evicted, so usable context shrank as the run progressed. Models with smaller context windows are disadvantaged under this policy. Compaction was added to the harness on 2026-08-16; results marked “compaction enabled” are not directly comparable on this dimension.

Copies GPT-5.4's full data pack — paste it into ChatGPT, Claude, or any AI to talk it through.

Where it ranks

GPT-5.4 against the field

FABLE 586 OPUS 4.881 OPUS 581 KIMI K377 GPT-5.6 SOL74 QWEN3.8 27B74 SONNET 5 (XHIGH)74 GROK 4.674 GPT-5.6 TERRA71 GPT-5.571 GPT-5.6 LUNA71 GROK 4.570 OX ALPHA65 GLM 5.263 GEMINI 3.5 FLASH56 OPUS 4.754 SONNET 4.652 GPT-5.451 GEMINI 3.1 PRO38

Scored 0–100 · higher is better

Download 16:9

X LinkedIn Download

The runs

How GPT-5.4 handled each benchmark

Score, capability radar, and the honest read on what it nailed and where it slipped. Hit Overlay to drop other models onto the same axes.

Car Wash Operations

A filthy operational dataset — ghost records, orphaned orders, typo'd customers, raw enum variants. Tests judgment under messy real-world data: what gets fixed, quarantined, or wrongly promoted.

51legacy 67

Interesting but Unreliable

GPT-5.4 showed better file-level rigor than Opus on SVC-007 and duplicate-customer conflict evidence, but strict scoring centers migration safety. Ghost/test records survived as canonical data, Terrence Blackwood was promoted to a customer, status/payment values remained raw enough for magic/case variants to survive, and the canonical customer count ballooned. The output is a review scaffold, not a trustworthy migration.

OverlayDownload radar

1Claude Fable 588

2Claude Opus 4.886

3Grok 4.684

4Qwen3.8 27B77

5Claude Opus 575

6Kimi K365

7Claude Sonnet 5 (xhigh)64

8GPT-5.6 Sol55

9GPT-5.6 Terra55

10GPT-5.555

11GPT-5.6 Luna55

12Grok 4.555

13GLM 5.2 (OpenRouter)55

14Ox Alpha54

15Gemini 3.5 Flash (High) Fast51

16GPT-5.451

17Opus 4.748

What it nailed

Completed the required artifact set.
Accounted for the full file corpus in the cross-review analysis.
Parsed deshawn_services.tsv and surfaced the SVC-007 conflict.
Preserved useful conflict evidence for some duplicate customer cases.

Where it slipped

Promoted Mickey Mouse, Test Customer, and Asdf Asdf into canonical data.
Promoted Terrence Blackwood to a canonical customer.
Left status and payment methods raw enough that case variants and magic survived.
Over-expanded the canonical customer table.

Misses Three Or More Primary CanariesPromotes Ghost RecordsPromotes Orphan Order