Vision and multimodal understanding
Reading images, charts and documents (MMMU-Pro style). Arena votes are not used.
No current, open, machine-readable vision leaderboard exists. The OpenCompass OpenVLM data stops at September 2025 and Epoch's vision data stops in 2025. The current independent source is Artificial Analysis MMMU-Pro, which needs their free API key (ARTIFICIAL_ANALYSIS_API_KEY). Until it is set, no pick is shown.
Pick updated 17 Nov 2026, 08:34 AEST
Text to image
UniGenBench++ semantic consistency score, judged automatically. Price is fal.ai per image.
Top pick
GPT-4o-1.5
Value pick
No cheaper priced model scores close enough.
GPT-4o-1.5 leads UniGenBench++ on overall semantic consistency (0-100) at 95.77 (higher is better). On fal.ai it is token-billed, see page. The top pick is billed in a different unit (see its price link), so no cheaper alternative can be confirmed and no value pick is named.
Source: UniGenBench++ · source data updated 23 Feb 2026 · fetched 17 Nov 2026, 08:34 AEST
Open benchmark (MIT). Scores prompt-following, judged by a model, not image beauty. Models released after the source date are not ranked yet.
Prices are the base rate on each fal.ai model page; size, quality and resolution can change them.
Pick updated 17 Nov 2026, 08:34 AEST
Image editing
GEditBench v2 overall score (automated judge, Elo scale). Price is fal.ai per edit.
Top pick
Nano Banana Pro
Value pick
Seedream 4.5
Nano Banana Pro leads GEditBench v2 on overall score (automated judge, Elo scale) at 1096 (higher is better). On fal.ai it costs $0.15 per image. Seedream 4.5 scores 1089, close to the top, for $0.04 per image.
Source: GEditBench v2 · source data updated 31 Mar 2026 · fetched 17 Nov 2026, 08:34 AEST
Open image-editing benchmark. Uses the automated Overall column, not the human arena column.
Prices are the base rate on each fal.ai model page; size, quality and resolution can change them.
Pick updated 17 Nov 2026, 08:34 AEST
Image to video
VBVR-Bench overall score: reasoning-heavy image-to-video tasks. Price is fal.ai per second.
Top pick
Seedance 2.0
Value pick
No cheaper priced model scores close enough.
Seedance 2.0 leads VBVR-Bench on overall score (0-1) at 0.5437 (higher is better) among models still available. Sora 2 scores 0.5457 but is excluded: OpenAI shut down the Sora 2 API on 24 Sep 2026. On fal.ai it costs $0.014 per 1000 tokens. The top pick is billed in a different unit (see its price link), so no cheaper alternative can be confirmed and no value pick is named.
| # | Model | Score | fal.ai price |
|---|
| 1 | Sora 2 · excluded: OpenAI shut down the Sora 2 API on 24 Sep 2026 | 0.5457 | Sora 2: $0.10 per second |
| 2 | Seedance 2.0 | 0.5437 | Seedance 2.0: $0.014 per 1000 tokens |
| 3 | Veo 3.1 | 0.48 | Veo 3.1: $0.40 per second |
| 4 | Runway Gen-4 Turbo | 0.4031 | n/a |
| 5 | Wan2.2-I2V-A14B · open weights | 0.3714 | Wan 2.2 A14B: $0.08 per second |
| 6 | Kling 2.6 | 0.3691 | Kling 2.6 Pro: $0.07 per second |
| 7 | LTX-2 · open weights | 0.3129 | LTX-2: $0.06 per second |
| 8 | CogVideoX1.5-5B-I2V · open weights | 0.2727 | n/a |
Source: VBVR-Bench · source data updated 15 Apr 2026 · fetched 17 Nov 2026, 08:34 AEST
Image-to-video reasoning benchmark with rule-based scoring. It measures whether the video does the task, not cinematic quality. The authors' own VBVR baselines are excluded.
Prices are the base rate on each fal.ai model page; size, quality and resolution can change them.
Pick updated 17 Nov 2026, 08:34 AEST
Text to speech
TTS Arena V2 blind-vote Elo (a community TTS arena, not LMArena). Price is fal.ai per 1,000 characters.
Top pick
Aurora
Value pick
No cheaper priced model scores close enough.
Aurora leads TTS Arena V2 on blind-vote Elo at 1687 (higher is better). No cheaper model with a published price scores close enough for a value pick.
| # | Model | Score | fal.ai price |
|---|
| 1 | Aurora · ±45, 312 votes | 1,687 | n/a |
| 2 | Inworld TTS MAX · ±27, 1,193 votes | 1,567 | n/a |
| 3 | Inworld TTS 1.5 MAX · ±37, 390 votes | 1,565 | n/a |
| 4 | Hume Octave · ±27, 949 votes | 1,551 | n/a |
| 5 | Inworld TTS · ±22, 1,712 votes | 1,543 | n/a |
| 6 | OpenAudio S2 · ±41, 322 votes | 1,539 | n/a |
| 7 | Eleven Multilingual v2 · ±27, 968 votes | 1,507 | ElevenLabs Multilingual v2: $0.10 per 1,000 characters |
| 8 | Eleven Turbo v2.5 · ±25, 951 votes | 1,505 | ElevenLabs Turbo v2.5: $0.05 per 1,000 characters |
Source: TTS Arena V2 · source data updated 17 Nov 2026 · fetched 17 Nov 2026, 08:34 AEST
Community blind A/B voting on speech samples (not LMArena). Live leaderboard; preliminary and suspended entries are excluded.
Prices are the base rate on each fal.ai model page; size, quality and resolution can change them.
Pick updated 17 Nov 2026, 08:34 AEST
Speech to text
Hugging Face Open ASR Leaderboard average word error rate. Lower is better.
Top pick
OtterAI/speech-v3.1
Value pick
No cheaper priced model scores close enough.
OtterAI/speech-v3.1 leads Open ASR Leaderboard on average word error rate % (English, short-form) at 3.8844 (lower is better). No cheaper model with a published price scores close enough for a value pick.
| # | Model | WER (lower better) | fal.ai price |
|---|
| 1 | OtterAI/speech-v3.1 | 3.8844 | n/a |
| 2 | reson8/resonant-1-flash | 4.0238 | n/a |
| 3 | zoom/scribe_v1 | 4.15 | n/a |
| 4 | microsoft/azure-speech-06-2026 | 4.2588 | n/a |
| 5 | sophea/asr-k1 (preview) | 4.6144 | n/a |
| 6 | zoom/scribe_v2_pro | 4.7686 | n/a |
| 7 | microsoft/Phi-4-multimodal-instruct · open (mit) | 5.0213 | n/a |
| 8 | ibm-granite/granite-speech-5.0-2b · open (apache-2.0) | 5.0622 | n/a |
Source: Open ASR Leaderboard · source data updated 20 Oct 2026 · fetched 23 Oct 2026, 08:34 AEST
Hugging Face's open leaderboard, run on public test sets. Lower word error rate is better.
Prices are the base rate on each fal.ai model page; size, quality and resolution can change them.
Pick updated 17 Nov 2026, 08:34 AEST
Voice cloning
Cloning a voice from a short sample.
There is no independent public leaderboard for voice cloning. TTS Arena V2 tests stock voices only and says so. No pick is shown rather than guessing.
Pick updated 17 Nov 2026, 08:34 AEST