Towards AITowards AIToneBench

Gemini 3.8 Flash (high thinking)

Google · proprietary · writing benchmark

#91 of 146 Elo 1419 Overall 77.0 Proprietary
Rank
#91 of 146
Writing Elo
1419 ±52
Overall
77.0 / 100
Cost / task
$0.062 per script
Family
Google 10th of 17
Type
Closed 71st of 96
Consistency
± 5.0 swingier than most
Avg tokens
26.2k in+out
Latency
46s per call
Family check:Gemini 3.1 Pro (default) is Google's best writer here, +3.4 overall vs this config for $0.081 more per script.

The short version

Gemini 3.8 Flash (high thinking) sits at #91 of 146 on ToneBench, toward the value end of the board, with a writing Elo of 1419 and an overall score of 77.0 out of 100. We measured it by having it write all 10 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 Google it ranks 10th of 17. Gemini 3.1 Pro (default) is the family's top writer here, about 151 Elo ahead, so this config trades some quality for whatever it saves you in cost or speed.

Among proprietary models, it comes in 71st of 96. That ranking is against the stronger half of the board: closed models still set the pace on voice fidelity here.

Worth knowing before you rely on it: this model is swingy. Its overall score moves ± 5.0 points between runs of the same brief, versus a board median of ± 3.1. A great draft and a mediocre one can come from the identical prompt. The shaded bands on the metric bars below show where that volatility lives.

What it does best is Visual Cues: it ranks 44th on that dimension at 83.5, well above the board average. Its softer spot is Length Discipline (49.3, 122nd), which is the thing to watch if that metric matters most for your use.

It was uneven across the 10 scripts. Its best run was the product release announcement / personal observation (82.8) and its weakest was the engineering process walkthrough / presentation adaptation (73.2), a spread of about 10 points. Worth knowing if your writing skews toward one of those formats.

On cost, it runs about $0.062/task. DeepSeek V4 Flash (native chat alias) scores higher for less money, at $0.002/task, so on pure value this config is not the frontier. Still, it beats 3 models that cost noticeably more.

Skill profile

The shape is the story: the further a corner reaches, the stronger that dimension. Biggest edge over the board: Visual Cues (+8.8); furthest behind: Length Discipline (-19.3).

50Tone?Craft?Substance?Flow?YouTube?Hook?Length?Anti-Slop?Cues?
Gemini 3.8 Flash (high thinking)Board averageBoard best per metricMax possible (100)

Per-metric scores

Nine writing dimensions, each scored 0–100 and blended by the editorial weight shown. Rank is against all 146 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.

Tone & Voice ?19% weight · -0.9 vs avg
best · Claude Fable 5 (max effort + 4.8 fallback) · 89.477.6± 2.4 · 91stboard avg · 78.5
Writing Craft ?13% weight · +1.5 vs avg
best · Claude Opus 5 (max effort) · 89.681.5± 2.1 · 74thboard avg · 80.0
Substance & Value ?15% weight · -0.2 vs avg
best · Claude Fable 5.1 (adaptive default) · 89.279.0± 4.0 · 91stboard avg · 79.2
Flow & Emotion ?14% weight · +1.5 vs avg
best · Claude Fable 5.1 (adaptive default) · 88.677.5± 3.3 · 76thboard avg · 76.0
YouTube Structure ?12% weight · +2.3 vs avg
best · Claude Fable 5 (max effort + 4.8 fallback) · 89.077.8± 4.4 · 78thboard avg · 75.5
Hook ?10% weight · +0.8 vs avg
best · Claude Opus 5 (xhigh) · 90.583.0± 3.0 · 92ndboard avg · 82.2
Length Discipline ?8% weight · -19.3 vs avg
best · GPT-6 Astra (max) · 95.249.3± 16.7 · 122ndboard avg · 68.6
Anti-Slop ?5% weight · -2.0 vs avg
best · GPT-6 Astra (max) · 94.381.4± 2.0 · 101stboard avg · 83.5
Visual Cues ?4% weight · +8.8 vs avg
best · GPT-5.6 Sol (ultra) · 88.583.5± 4.8 · 44thboard avg · 74.7

Per-article scores

The same 10 real scripts every current-ranked model writes, scored individually. Different formats stress different skills.

Article 1
opinion / warning explainer
76.5
out of 100
Article 2
news-analysis / skeptical explainer
80.6
out of 100
Article 3
personal roadmap / opinion
74.3
out of 100
Article 4
short explainer
75.9
out of 100
Article 5
news-analysis / opinion explainer
78.3
out of 100
Article 6
founder announcement / personal origin story
81.1
out of 100
Article 7
personal technical walkthrough / agentic workflow case study
73.5
out of 100
Article 8
product release announcement / personal observation
82.8
out of 100
Article 9
career guide / hiring analysis
73.8
out of 100
Article 10
engineering process walkthrough / presentation adaptation
73.2
out of 100

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, Gemini 3.8 Flash (high thinking) varies by ± 5.0 points between runs versus a board median of ± 3.1, so it is swingier than the typical model here, worth knowing if you need repeatable output.

Its most volatile dimension is Substance & Value (± 4.0 vs a board median of ± 2.3): two runs of the same brief can land visibly different substance & value scores.

MetricThis modelBoard medianVerdict
Tone & Voice ?77.6 ± 2.4± 2.0typical
Writing Craft ?81.5 ± 2.1± 1.9typical
Substance & Value ?79.0 ± 4.0± 2.3swingier than most
Flow & Emotion ?77.5 ± 3.3± 2.3typical
YouTube Structure ?77.8 ± 4.4± 3.4typical
Hook ?83.0 ± 3.0± 2.4typical
Length Discipline ?49.3 ± 16.7± 9.4swingier than most
Anti-Slop ?81.4 ± 2.0± 2.2typical
Visual Cues ?83.5 ± 4.8± 4.2typical

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.

Latency per script
46sfaster than the board median of 68s
Prompt tokens in
12.0kstyle guide + brief + research packet
Tokens out
14.1kwell above the board median (thinks a lot)
Cost per script
$0.062 ± 0.021measured from actual billed tokens
List price used
$0.75 / $3.75 per M tokinput / output

What each judge scored it

The published overall of 77.0 is the consensus of three family-disjoint judges scoring the same 50 stored drafts blind with the identical rubric. The highest and lowest judge differ by 4.5 points on its overall. The three judges essentially agree on this model. How the panel works: methodology.

Claude Opus 5
Anthropic
74.3
GPT-5.6 Sol (medium)
OpenAI
78.0
DeepSeek V4 Flash
DeepSeek
78.7

How we ran Gemini 3.8 Flash (high thinking)

Frequently asked questions

How good is Gemini 3.8 Flash (high thinking) at writing?

On ToneBench it ranks #91 of 146 with a writing Elo of 1419 and an overall score of 77.0/100. That score comes from writing our 10 real YouTube scripts five times each and scoring every draft blind against our own finished versions across nine writing dimensions.

Is Gemini 3.8 Flash (high thinking) the best Google model for writing?

Not quite. Within Google it ranks 10th of 17; Gemini 3.1 Pro (default) is the family's best writer here.

Is Gemini 3.8 Flash (high thinking) good value for the money?

It costs about $0.062/task. DeepSeek V4 Flash (native chat alias) scores higher for less, so it is not the value pick.

What are Gemini 3.8 Flash (high thinking)'s strengths and weaknesses?

Its strongest dimension is Visual Cues (44th on the board, 83.5). Its weakest is Length Discipline (122nd on the board, 49.3). The full nine-metric breakdown is on this page.

Does Gemini 3.8 Flash (high thinking) write some formats better than others?

Yes. Its best of our 10 scripts was the product release announcement / personal observation (82.8) and its weakest was the engineering process walkthrough / presentation adaptation (73.2).

How consistent is Gemini 3.8 Flash (high thinking) between runs?

We run every script 5 times. Gemini 3.8 Flash (high thinking)'s overall score varies by about ±5.0 points between runs, versus a board median of ±3.1. That is swingier than typical, so expect more draft-to-draft variation. The full per-metric variability table is on this page.

How was Gemini 3.8 Flash (high thinking) evaluated?

Via Google Gemini API using the exact model/route id gemini-3.8-flash at reasoning_effort high, run on 2026-09-03. 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 Gemini 3.8 Flash (high thinking) open source?

No, it is a proprietary (closed-weights) model. Among proprietary models it ranks 71st of 96 for writing.

← Back to the full ToneBench leaderboard