GPT-5.4 (xhigh)
OpenAI · proprietary · writing benchmark
The short version
GPT-5.4 (xhigh) sits at #21 of 130 on ToneBench, in the middle of the pack, with a writing Elo of 2202 and an overall score of 87.5 out of 100. We measured it by having it write all 9 of our real YouTube scripts, five times each, then scoring every draft blind against our own finished versions. Here is how it shook out.
Within OpenAI it ranks 6th of 36. GPT-5.6 Sol (ultra) is the family's top writer here, about 113 Elo ahead, so this config trades some quality for whatever it saves you in cost or speed.
Among proprietary models, it comes in 19th of 83. That ranking is against the stronger half of the board: closed models still set the pace on voice fidelity here.
One thing that stands out: it is unusually consistent. Across its 45 drafts the overall score only moves ± 2.0 points run to run, versus a board median of ± 3.8. What you get on the first try is close to what you get every try.
What it does best is Visual Cues: it ranks 10th on that dimension at 90.2, well above the board average. Its softer spot is Length Discipline (81.8, 41st), which is the thing to watch if that metric matters most for your use.
It held steady across all 9 scripts, with its best run on the personal roadmap / opinion (88.9). Consistency like that is underrated, honestly, since it means fewer surprises when you actually put it to work.
On cost, it runs about $0.279/task. Claude Opus 5 (low) scores higher for less money, at $0.115/task, so on pure value this config is not the frontier.
If you're tuning reasoning effort, we also tested GPT-5.4 at other settings. None of them beat this config; the closest is GPT-5.4 (Elo 1729), so the only question is how much quality you want to trade for cost or speed.
Skill profile
The shape is the story: the further a corner reaches, the stronger that dimension. Biggest edge over the board: Visual Cues (+15.7); smallest edge: Hook (+8.2).
Per-metric scores
Nine writing dimensions, each scored 0–100 and blended by the editorial weight shown. Rank is against all 130 current-ranked models. The thick bar is this model; the thin lines above and below are the current field's best model (darker) and average (lighter) on the same scale. The small ± number is its run-to-run variation.
Per-article scores
The same 9 real scripts every current-ranked model writes, scored individually. Different formats stress different skills.
Consistency: run-to-run variability
Every cell on this page is the mean of 5 independent runs per script; the ± numbers are the run-to-run standard deviation. Overall, GPT-5.4 (xhigh) varies by ± 2.0 points between runs versus a board median of ± 3.8, so it is noticeably steadier than the typical model here.
Its steadiest is Anti-Slop (± 1.5 vs ± 2.6 board median), which you can rely on run after run.
| Metric | This model | Board median | Verdict |
|---|---|---|---|
| Tone & Voice ? | 88.6 ± 1.5 | ± 2.2 | typical |
| Writing Craft ? | 87.6 ± 1.7 | ± 2.0 | typical |
| Substance & Value ? | 89.0 ± 1.8 | ± 2.6 | typical |
| Flow & Emotion ? | 84.9 ± 2.3 | ± 2.6 | typical |
| YouTube Structure ? | 86.7 ± 2.3 | ± 3.6 | typical |
| Hook ? | 88.4 ± 2.2 | ± 3.0 | typical |
| Length Discipline ? | 81.8 ± 12.2 | ± 8.6 | typical |
| Anti-Slop ? | 93.1 ± 1.5 | ± 2.6 | steadier than most |
| Visual Cues ? | 90.2 ± 2.2 | ± 4.7 | steadier than most |
Measured, not modeled
Numbers we log on every run and rarely talk about. None of these affect the writing scores; cost and latency are informational.
What each judge scored it
The published overall of 87.5 is the consensus of three family-disjoint judges scoring the same 45 stored drafts blind with the identical rubric. The highest and lowest judge differ by 8.3 points on its overall. A typical amount of judge disagreement for this board. How the panel works: methodology.
How we ran GPT-5.4 (xhigh)
- Accessed viaCodex CLI (OpenAI subscription) · model page ↗
- Exact model / route id
gpt-5.4atreasoning_effort: xhigh - Generations5 independent runs × 9 scripts = 45 drafts, stored immutably after generation
- Run dates2026-07-29 → 2026-08-11
- Judges3-judge panel: Claude Opus 5, high effort (Anthropic) + DeepSeek V4 Flash (OpenRouter) + GPT-5.6 Sol, medium reasoning (Codex CLI): every draft scored independently by all three with the identical rubric, blind to model identity, evidence quotes required before each metric score; published score = per-metric consensus mean, verified with a Bradley-Terry refit (methodology)
- Samplingprovider-default sampling, no fine-tuning, no system-prompt changes; every current-ranked model gets the identical brief, style contract and research packet
- What stays privatethe exact rubric and any non-public reference material; source-publication timing is disclosed for exceptions (why)
Frequently asked questions
How good is GPT-5.4 (xhigh) at writing?
On ToneBench it ranks #21 of 130 with a writing Elo of 2202 and an overall score of 87.5/100. That score comes from writing our 9 real YouTube scripts five times each and scoring every draft blind against our own finished versions across nine writing dimensions.
Is GPT-5.4 (xhigh) the best OpenAI model for writing?
Not quite. Within OpenAI it ranks 6th of 36; GPT-5.6 Sol (ultra) is the family's best writer here.
Is GPT-5.4 (xhigh) good value for the money?
It costs about $0.279/task. Claude Opus 5 (low) scores higher for less, so it is not the value pick.
What are GPT-5.4 (xhigh)'s strengths and weaknesses?
Its strongest dimension is Visual Cues (10th on the board, 90.2). Its weakest is Length Discipline (41st on the board, 81.8). The full nine-metric breakdown is on this page.
Does GPT-5.4 (xhigh) write some formats better than others?
It was fairly even across all 9 scripts, topping out on the personal roadmap / opinion (88.9).
How consistent is GPT-5.4 (xhigh) between runs?
We run every script 5 times. GPT-5.4 (xhigh)'s overall score varies by about ±2.0 points between runs, versus a board median of ±3.8. That makes it one of the steadier models we test. The full per-metric variability table is on this page.
How was GPT-5.4 (xhigh) evaluated?
Via Codex CLI (OpenAI subscription) using the exact model/route id gpt-5.4 at reasoning_effort xhigh, run on 2026-07-29. 5 runs per script, provider-default sampling, no fine-tuning; every draft scored blind with a fixed rubric by a three-family judge panel. Full details in the 'How we ran it' section and the methodology.
Is GPT-5.4 (xhigh) open source?
No, it is a proprietary (closed-weights) model. Among proprietary models it ranks 19th of 83 for writing.