The numbers, what the leaderboards actually say

Public benchmarks should be taken with a grain of salt, we will get to why in a moment, but they are a useful starting point for comparing models, and the patterns in them are real. The two most-watched leaderboards are the Artificial Analysis text-to-image and text-to-videoText-to-videoGenerating video from text alone, with no starting image. benchmarks, which use human preference voting (people are shown two outputs from different models for the same promptPromptThe text description you provide to a model to specify what you want it to generate. and asked which they prefer), aggregate the votes into ELOEloA rating derived from head-to-head human preference votes, borrowed from chess, used to rank models on leaderboards.-style ratings, and publish the results.

Here are approximate snapshots as of early 2026, drawn from those benchmarks and from the public claims of the labs themselves. Numbers will be out of date by the time you read this, but the rough ordering is what matters.

IMAGE MODEL LEADERBOARD

Model

Lab

Architecture

Open weightsOpen weightsModels whose trained parameters are publicly downloadable, allowing anyone to run them locally or fine-tune them. Contrasts with closed/API-only models.

Strengths

Nano Banana Pro

Google

Gemini-native

No

Text rendering, world knowledge, 14-image composition

FLUX 1.1 [pro]

Black Forest Labs

12B MultimodalMultimodalHandling more than one kind of data, for example text and images together, in a single model. Diffusion TransformerTransformerThe neural network architecture built on attention that underpins modern language and, increasingly, image and video models. It scales well and handles long-range relationships better than convolutional networks. (MM-Diffusion Transformer (DiTDiT (Diffusion Transformer)A diffusion model architecture that uses a Transformer instead of a U-Net. Introduced by Peebles and Xie in 2022. Scales much better than U-Nets and is the basis for almost all current frontier models.))

No (pro)

General photoreal, prompt adherencePrompt adherenceHow faithfully a model follows the specifics of your prompt, meaning the objects, counts, colors, spatial relationships, and text you asked for. High adherence means you get what you described, not just something in the right neighborhood.

Ideogram 3.0

Ideogram

Latent diffusionLatent diffusionDiffusion that operates in a compressed 'latent' space produced by a VAE, rather than directly in pixel space. Dramatically reduces compute cost. Used by essentially every modern image and video model.

No

Typography, design, posters

Recraft V3

Recraft

Latent diffusion + vector

No

Long-form text, vector output

GPT-Image (4o)

OpenAI

LLM-native

No

Instruction following, edits

Seedream 5.0

ByteDance

MM-DiTMM-DiT (Multimodal Diffusion Transformer)A variant of DiT that processes text and image (or audio) tokens in separate parallel streams joined by attention. Used by Stable Diffusion 3, FLUX, and most newer models. Improves prompt following significantly.

No

Cinematic, e-commerce design

FLUX [dev]

Black Forest Labs

12B MM-DiT

Yes (non-comm)

Open ecosystem workhorse

Qwen-Image 2.0

Alibaba

7B MM-DiT

Yes

Chinese text, open license

SD 3.5 Large

Stability AI

8B MM-DiT

Yes

Permissive license

Midjourney V7

Midjourney

DiT (closed)

No

Aesthetic distinctiveness

VIDEO MODEL LEADERBOARD

Model

Lab

Audio

Max length

Cost / sec

Veo 3.1

Google

Yes (joint)

8s @ 1080p

$0.50 - $1.50

Kling 3.0

Kuaishou

Yes (joint)

15s cinematic

$0.30 - $0.80

Sora 2 Pro

OpenAI

Yes (joint)

25s @ 1792x1024

$0.50 - $2.00

Seedance 2.0

ByteDance

Yes

12s @ 1080p

$0.20 - $0.60

Hailuo 2.3

MiniMax

Partial

10s @ 1080p

$0.05 - $0.15

Runway Gen-4.5

Runway

Separate

10s @ 1080p

$0.25 - $0.60

Wan 2.5

Alibaba

Yes

10s @ 1080p

Free (open)

Luma Ray 3

Luma

Separate

10s @ 4K

$0.30 - $0.80

Pika 2.2

Pika

Separate

10s @ 1080p

$0.10 - $0.30

Hunyuan Video

Tencent

Separate

Variable

Free (open)

Notice the bifurcation. The top of the leaderboard is dominated by closed-source models from labs with massive resources (Google, Kuaishou, OpenAI, ByteDance). The middle is a mix of mid-priced commercial offerings (Hailuo, Runway, Luma, Pika). The bottom is open-source models (Wan, Hunyuan) that are free to use but require you to bring your own infrastructure. There is no single 'best' model, there is a quality-cost-openness trade-off, and your right answer depends on which axis matters most to you.

TRAINING COST COMPARISON

Model

Year

Type

Estimated cost

Compute scale

Stable Diffusion 1.5

2022

Image

~$600K

150K A100-hours

SDXL

2023

Image

$2-3M

Several hundred K A100-hours

Stable Diffusion 3

2024

Image

$5-10M

Multi-week multi-cluster

FLUX (BFL)

2024

Image

$10-30M (est)

Undisclosed

Open-Sora 2.0

2025

Video

$200K

Optimized open-source

Hunyuan Video

2024

Video

$5-15M (est)

13B params

Sora 1

2024

Video

$100M+ (est)

Months on thousands of GPUs

Veo 3

2025

Video

$100M+ (est)

Google internal

Kling 3.0

2026

Video

Undisclosed

Kuaishou internal

MODEL FUNDING & VALUATION

Lab

Total raised

Last valuation

Notable backers

OpenAI

$60B+

$500B+

Microsoft, Thrive, Khosla

Anthropic

$25B+

$170B+

Google, Amazon

Black Forest Labs

$450M+

$3.25B (Dec 2025)

a16z, Salesforce, AMP

Runway

$300M+

~$3B (est)

Google, Nvidia, Salesforce

Pika Labs

$135M

~$700M

Lightspeed, Spark, a16z

Stability AI

$100M+

Recovering

Coatue, Lightspeed

Luma Labs

$70M+

~$300M

a16z, Amplify

Ideogram

$80M+

Undisclosed

a16z, Index

Midjourney

$0 (bootstrapped)

Profitable

None - founder funded