AI image generation has moved well past the era of distorted hands and garbled text. The models available in mid-2026 can produce marketing assets, photorealistic scenes, and legally defensible stock images - and the gap between the best and second-best has never been wider. GPT Image 2 reset the competitive map in April, HiDream proved you don't need 32 billion parameters to top benchmarks, and Google made a serious bid for the high-volume API market with its free Gemini app access. This leaderboard ranks what's actually worth using right now.

Rankings are drawn from the Artificial Analysis Text-to-Image Arena, a blind human preference evaluation where users compare two anonymously produced images from the same prompt and pick the better result. All Elo scores below are from that arena as of June 2026. Models that don't participate in the arena (Midjourney, Ideogram, Firefly) are assessed separately in the Specialist Models section using independent evaluations.

TL;DR

GPT Image 2 leads by 70+ Elo points with "Outstanding" text rendering - the first image model to simultaneously win on text accuracy and photorealism
Nano Banana 2 from Google is free in the Gemini app and scores third overall, making it the default choice for volume users
HiDream-O1-Image is the best open-weight option - a 8B model that beats 32B FLUX.2 [dev] on five benchmarks, available free under MIT

How This Leaderboard Works

The primary ranking metric is Artificial Analysis Arena Elo, derived from blind human preference comparisons. Every model in the main table participates in the same arena, so comparisons are apples-to-apples.

Three qualitative ratings supplement the Elo score:

Text rendering - accuracy of in-image typography, signage, and labels. Tested with complex multi-word prompts.
Photorealism - how convincingly a created image passes as a photograph in blind evaluation.
Artistic quality - composition, lighting, style coherence, and visual interest judged via preference voting.

Ratings are: Outstanding, Excellent, Very Good, Good. Models not in the arena are excluded from the main rankings table.

Methodology note: Arena rankings measure human preference in side-by-side comparisons, not absolute quality. A model with a narrow strength (say, photorealism of landscapes) can outscore a more flexible competitor in arena voting depending on what prompts evaluators submit. Treat the Elo as a reliable proxy for general preference, not a ceiling measurement.

Main Rankings: Arena-Tracked Models

Rank	Model	Provider	Arena Elo	Text Rendering	Photorealism	Artistic Quality	API Pricing
1	GPT Image 2	OpenAI	1338	Outstanding	Excellent	Excellent	$0.211/image (high)
2	GPT Image 1.5	OpenAI	1268	Excellent	Excellent	Excellent	$0.133/image
3	Nano Banana 2	Google DeepMind	1260	Very Good	Very Good	Very Good	$0.067/image (free in app)
4	Nano Banana Pro	Google DeepMind	1219	Excellent	Very Good	Excellent	~$0.03/image
5	Recraft V4.1 Utility Pro	Recraft	1204	Outstanding	Good	Good	$0.04/image (Utility)
6	HiDream-O1-Image	HiDream-ai	1191	Very Good	Very Good	Good	Free (MIT open weights)
7	FLUX 2 Max	Black Forest Labs	1168	Very Good	Excellent	Very Good	~$0.05/image
8	FLUX 2 Flex	Black Forest Labs	1157	Good	Very Good	Very Good	~$0.04/image
9	FLUX 2 Pro	Black Forest Labs	1153	Good	Very Good	Good	$0.055/image

Elo scores from Artificial Analysis Text-to-Image Arena, June 2026. GPT Image 1.5 remains available via API but is expected to be deprecated when GPT Image 2 rolls out fully.

Specialist Models: High Quality, No Arena Participation

Some top models don't participate in the Artificial Analysis arena but remain important depending on your use case. These are evaluated using independent benchmarks and hands-on testing.

Model	Provider	Standout Strength	Text Accuracy	Photorealism	Artistic	Pricing
Ideogram 3.0	Ideogram AI	Typography	Outstanding (~92%)	Very Good	Very Good	$0.03-$0.09/image
Midjourney v8.1	Midjourney	Artistic output	Good	Excellent	Outstanding	$10-$120/month
Adobe Firefly 4	Adobe	Commercial licensing	Good	Very Good	Good	Creative Cloud sub
Stable Diffusion 3.5 Ultra	Stability AI	Self-hosting	Good	Very Good	Very Good	Free (open-weight)

Key Takeaways

GPT Image 2: 241-Point Gap Is Real

GPT Image 2 launched on April 22, 2026, under the product name ChatGPT Images 2.0. On arena.ai's text-to-image leaderboard, it debuted with a 241-point lead over second-place Nano Banana 2 - the largest gap between first and second in that arena's history. Its Artificial Analysis Elo of 1338 tells the same story: no other model in the public arena gets close.

The engineering explanation is the O-series reasoning integration. GPT Image 2 has two modes: Instant (fast generation) and Thinking (deliberate planning before pixels are drawn). The Thinking mode is what drives the text rendering breakthrough - at near-perfect accuracy, it handles multilingual signage, multi-line poster copy, and dense data labels in ways that would have required post-processing in any 2025 model. OpenAI's $0.211 high-quality price point is steep for casual use but competitive for professional workflows where rework costs more.

GPT Image 2's 241-point lead on arena.ai is the largest margin any model has held over its nearest competitor in the history of that leaderboard.

The DALL-E 4 API was retired on May 12, 2026. GPT Image 2 is the direct replacement for API users, though the cost structure changed notably. The low-quality tier at $0.006/image gives developers a cheap path for previews; production runs at $0.053/image (medium) hit a reasonable midpoint.

Nano Banana 2: Google's Free Tier Bet Pays Off

Nano Banana 2 (formally: Gemini 3.1 Flash Image Preview) launched February 26, 2026, and rolled out free to all Gemini app users across 141 countries. At Elo 1260, it sits third overall - above its own sibling, Nano Banana Pro (Elo 1219), which previously held the Gemini family's highest position.

The reversal is notable. The Flash-based model now outscores the Pro model in blind human preference, despite being cheaper and faster. Google's bet is that a Flash-grade model free to millions of app users generates enough training signal to keep improving faster than a more expensive, less-used Pro variant.

For API users, Nano Banana 2 prices at roughly $0.067/image - more than FLUX 2 Max but with Google's ecosystem integration and 4K output as offsets. The bigger story is the free tier: image generation without API costs is a material advantage for prototyping and early-stage products.

Artificial Analysis Text-to-Image Arena rankings as of June 2026, showing GPT Image 2 leading at Elo 1338. Source: artificialanalysis.ai

HiDream-O1-Image: The Efficient Open-Source Contender

The most interesting number in the open-source space is the parameter gap. HiDream-O1-Image is a 8-billion-parameter model that scores Elo 1191 in the arena - above every FLUX 2 variant - and beats 32B FLUX.2 [dev] on GenEval, DPG-Bench, HPSv3, CVTG-2K, and LongText benchmarks. Its Unified Transformer architecture eliminates the VAE bottleneck that most diffusion models inherit from their SDXL-era lineage.

For teams that need to self-host, HiDream-O1-Image is now the default recommendation over FLUX 2 Dev. The MIT license is fully commercial, the VRAM requirements are manageable at 8B, and the benchmark results hold across task types. The FLUX family still wins on photorealism in the arena, but HiDream closes most of that gap at a fraction of the inference cost.

The FLUX 2 comparison guide on this site covers the FLUX variants in detail if you're choosing between FLUX 2 Max, FLUX 2 Dev, and FLUX 2 Schnell for your workload.

Midjourney v8.1: Speed First, Arena Optional

Midjourney v8.1 launched on April 30, 2026 as the fastest version Midjourney has shipped - standard jobs render 4 to 5 times faster than earlier versions, and HD mode (native 2048x2048) is on by default. It doesn't participate in the Artificial Analysis arena, so Elo comparisons aren't possible.

Based on hands-on testing and the Midjourney vs FLUX comparison, v8.1 still produces the most visually distinctive artistic output in the field. Its "Outstanding" artistic quality rating reflects a model tuned for coherence and intentional aesthetic, not just prompt adherence. Text rendering improved much from v7 but doesn't approach GPT Image 2 or Ideogram 3.0 at the specialist level.

The subscription model ($10-$120/month) makes Midjourney cost-uncompetitive for high-volume API use cases but workable for individual creators and small teams who value the interface and community.

Midjourney v8.1 sample output showing HD quality image Midjourney v8.1 default HD mode produces at native 2048x2048, then optionally upscales to 4K. Source: midjourney.com

Ideogram 3.0: Text Rendering Still Unmatched

Ideogram 3.0 sits in its own category for typography-heavy work. Its ~92% text accuracy - compared to ~30-40% for Midjourney v7 in the same conditions - is a specific architecture decision: Ideogram trained the model explicitly for glyph recognition and kerning rather than treating in-image text as a side effect of general pretraining.

For marketing assets, social graphics, signage mockups, and any workflow where legible text is required in the created image, Ideogram 3.0 is still the specialist. At $0.03/image (Turbo tier), it's also among the cheapest quality options available. The arena ranking gap versus GPT Image 2 reflects that general human preference tests skew toward photorealism and composition - not the narrow text rendering prompts where Ideogram controls.

Practical Guidance

For all-around quality with budget flexibility: GPT Image 2 at medium quality ($0.053/image) gives you the best arena-validated output. Use low quality ($0.006/image) for drafts.

For free or near-free volume: Nano Banana 2 via the Gemini app costs nothing for individual use. The API at $0.067/image is competitive once you factor in 4K support and Google ecosystem integrations.

For self-hosting and open-source: HiDream-O1-Image is the current top pick - Elo 1191, MIT license, 8B parameters manageable on a 24GB consumer GPU. Run FLUX 2 Dev if you specifically need the photorealism edge that FLUX's architecture provides.

For professional creative work: Midjourney v8.1's artistic output is still the most visually distinctive in the field. Not API-friendly, but the right call for image-forward creative workflows where the subscription cost is absorbed.

For typography-heavy commercial assets: Ideogram 3.0 at $0.03-$0.09/image. No other model approaches its text accuracy for marketing and design work.

For enterprise and legally cleared content: Adobe Firefly 4 with Creative Cloud. The commercial licensing clarity and Content Credentials support are worth the premium for organizations with IP exposure.

For the best value in the FLUX family: See our FLUX 2 comparison guide - Max for quality, Dev for open-weight self-hosting, Schnell for speed and cost efficiency.

Full API pricing details across all models are tracked in our image generation pricing comparison.

FAQ

Which AI image model is best overall in 2026?

GPT Image 2 leads the Artificial Analysis Arena at Elo 1338, with a 70+ point gap over the next competitor. It has the best text rendering of any publicly ranked model and wins head-to-head preference tests across photorealism and artistic quality categories.

What is the best free AI image generator right now?

Nano Banana 2 (Gemini 3.1 Flash Image) is free in the Gemini app for all users. With an Arena Elo of 1260, it ranks third overall among all tracked models - above Nano Banana Pro, which previously held Google's top spot.

What is the best open-source text-to-image model?

HiDream-O1-Image, released May 2026, under MIT license. Its 8B architecture achieves Arena Elo 1191 and beats 32B FLUX.2 [dev] on five standard benchmarks. FLUX 2 Max (Elo 1168) remains the best option if photorealism is the primary requirement.

Is DALL-E still available?

No. OpenAI retired the DALL-E 4 API on May 12, 2026. GPT Image 2 is the replacement for API users. The pricing structure is different - the old per-image fixed rate has been replaced by a quality-tiered model ($0.006 low, $0.053 medium, $0.211 high).

How often does this leaderboard update?

Arena Elo scores shift continuously as new votes come in, but meaningful ranking changes happen when a new model launches or a major update ships. This leaderboard is reviewed at the start of each month and updated in full.

Sources: