Selected external evaluations for image, video, and writing tasks. Pair the relevant model result with the features, access route, and cost of your workflow.
Scores belong to a specific task, model version, evaluation pool, and leaderboard setting. Compare within that context; use the original source for the current full table. A model result does not automatically describe every application offering that model.
Image generation & editing
ACCESSED 2026-10-21
Generation and single-image editing answer different questions. Use the model label and quality setting that match the workflow you intend to buy.
Artificial Analysis
GPT Image 2.5 Sunburst (max)
text-to-image
Metric
Quality Elo (AA-Image-T2I v2.1)
Score
1222
Position
1
Uncertainty / range
display: 1222±10; confidence level: 0.95; lower: 1212; upper: 1232; rank range: 1; preliminary: Not shown in this snapshot
Sample
13014
Source updated
Not shown in this snapshot
Captured
2026-10-21
Settings / pool
pool: AA-Image-T2I v2.1; Overall; evaluator pool: Validated private evaluators; blind pairwise preference; generation: Published defaults; highest supported resolution downscaled to 1024×1024; 1:1; one image; seed 42; exact api snapshot: Not shown in this snapshot; quality label: max; sample unit: Comparisons, shown as Samples
display: 1176±12; confidence level: 0.95; lower: 1164; upper: 1188; rank range: 3–6; preliminary: Not shown in this snapshot
Sample
6210
Source updated
Not shown in this snapshot
Captured
2026-10-21
Settings / pool
pool: AA-Image-T2I v2.1; Overall; evaluator pool: Validated private evaluators; blind pairwise preference; generation: Published defaults; highest supported resolution downscaled to 1024×1024; 1:1; one image; seed 42; exact api snapshot: Not shown in this snapshot; quality label: Not shown in this snapshot; sample unit: Comparisons, shown as Samples
A candidate for budgeted iterations when paired with the catalog’s 1K output price; this result tests text-only generation, not preservation of fourteen references.
Catalog match: Named model match; exact API revision not shown.
display: 1328±9; confidence level: Not shown in this snapshot; rank range: 4–6; preliminary: Yes; notes: Page labels confidence-interval plots as bootstrapping; interval confidence level is not stated in the extracted leaderboard.
Sample
5310
Source updated
2026-10-07
Captured
2026-10-21
Settings / pool
evaluator pool: Community prompts and anonymous pairwise preference votes; generation parameters: Not shown in this snapshot; exact api snapshot: Not shown in this snapshot; pool: Text-to-Image Arena; Overall; sample unit: Votes; source date label: Displayed leaderboard date
display: 1514±6; confidence level: Not shown in this snapshot; rank range: 1–1; preliminary: Yes; notes: Page labels confidence-interval plots as bootstrapping; interval confidence level is not stated in the extracted leaderboard.
Sample
82078
Source updated
2026-10-19
Captured
2026-10-21
Settings / pool
evaluator pool: Community prompts and anonymous pairwise preference votes; generation parameters: Not shown in this snapshot; exact api snapshot: Not shown in this snapshot; pool: Image Edit Arena; Single Image Edit; sample unit: Votes; source date label: Displayed leaderboard date
Bradley–Terry preferences on an Elo scale, anchored at FLUX.2 [dev] = 1000. English, curated prompts are balanced across use cases and capabilities and refreshed monthly. Private evaluators are screened. T2I and editing are independent pools. Vote-independence assumptions can narrow intervals when observations correlate.
Selected row retains Preliminary status and displayed date 2026-10-07. Model generation parameters and interval confidence level were not available in extracted text.
Published and updated 2026-02-09. Describes overlapping task categories and an LLM prompt filter removing about 15% of noisy text-to-image prompts in that update. That historical percentage is not a current sample attrition estimate.
Users submit prompts, compare two anonymous responses, vote, then see identities. This public community preference pool differs from AA’s screened private evaluators.
Video generation
ACCESSED 2026-10-21
Silent and audio-enabled video use separate evaluation pools. AA defaults target 1080p, 24 fps, ten seconds and 16:9 where supported; model-specific settings and prompt coverage remain important.
profile: aa_video_defaults; pool: AA-Video-T2V-Silent v2.1; audio enabled: No; category: Overall; per model settings: Not separately exposed in the captured row.
profile: aa_video_defaults; pool: AA-Video-T2V-Silent v2.1; audio enabled: No; category: Overall; resolution: 720p; prompt coverage: AA reports 61% of its text-to-video prompt set, limited to prompts at most 1,000 characters.
Partial prompt coverage limits generalization. Its image-led UI workflow may still fit an approved still; this text-only result does not evaluate that mode.
Catalog match: Exact model and 720p variant match the Runway record.
profile: aa_video_defaults; pool: AA-Video-T2V v2.1; audio enabled: Yes; category: Overall; per model settings: Not separately exposed in the captured row.
Relevant evidence for audiovisual text generation, not a dedicated lip-sync score or proof of conversational-editing quality. Do not compare its Elo with the silent rows.
Catalog match: Exact model gemini-omni-1.1-flash verified in Google documentation and mentioned in the audio article, but absent from catalog; do not link it to the writing-model Gemini record.
Validated private evaluators make blinded pairwise preferences. Bradley–Terry estimates are reported as Elo; scores refresh hourly. Six task/audio pools are separate: never average or directly compare them. Reported intervals assume independent votes and may understate correlated uncertainty. Silent T2V uses 500 curated prompts; audio T2V uses 1,000. The benchmark is a preference measure, not a universal production-quality guarantee.
Official identifier gemini-omni-1.1-flash; text generation produces audio, while the Interactions API enables conversational editing.
Creative writing
ACCESSED 2026-10-21
Creative-writing preferences help choose a prose candidate. Keep effort settings and uncertainty in view, then pair that signal with the tools needed for your source-to-draft workflow.
Arena
Claude Opus 5.5 (High)
creative-writing
Metric
Arena Score
Score
1520
Position
1
Uncertainty / range
display: ±14; plus minus: 14; confidence level: Not shown in this snapshot; rank spread: 1; 5; preliminary: No
Sample
2241
Source updated
2026-10-20
Captured
2026-10-21
Settings / pool
category: Text / Creative Writing; model effort label: High; style control: Not shown in this snapshot; factuality adjustment: Not shown in this snapshot; adjustment state: unverified; adjustment note: Extraction exposes adjustment choices, not selected state.; temperature: Not shown in this snapshot; system prompt: Not shown in this snapshot
display: ±19; plus minus: 19; confidence level: Not shown in this snapshot; rank spread: 1; 7; preliminary: Yes
Sample
1055
Source updated
2026-10-20
Captured
2026-10-21
Settings / pool
category: Text / Creative Writing; model effort label: High; style control: Not shown in this snapshot; factuality adjustment: Not shown in this snapshot; adjustment state: unverified; adjustment note: Extraction exposes adjustment choices, not selected state.; temperature: Not shown in this snapshot; system prompt: Not shown in this snapshot
display: ±16; plus minus: 16; confidence level: Not shown in this snapshot; rank spread: 14; 72; preliminary: No
Sample
1688
Source updated
2026-10-20
Captured
2026-10-21
Settings / pool
category: Text / Creative Writing; model effort label: Max; style control: Not shown in this snapshot; factuality adjustment: Not shown in this snapshot; adjustment state: unverified; adjustment note: Extraction exposes adjustment choices, not selected state.; temperature: Not shown in this snapshot; system prompt: Not shown in this snapshot
Three selected rows, not a complete ranking. Displayed uncertainty level was not specified in extracted text; no 95% label inferred. Adjustment selection is unknown.
Published 2024-08-29; updated 2025-06-13. Models length and markdown features as covariates; observational, not proof that style and quality are causally separated.
Verifies gpt-6.1-sol, max effort support and Responses tools. Default effort is medium, not the Arena Max row.
Choose by task: identify a promising model, confirm the exact version and available controls, then compare the complete workflow cost. Dated source snapshot; new leaderboard releases may change the order.