Recraft V3review, pricing and limits

Recraft V3 remains available in 2026 as the more stylized, lower-cost option next to Recraft's newer V4 and V4.1 models, which have taken over as the default for photorealistic and compositional work. It once beat Flux's 1143 Elo score to top the Artificial Analysis Text-to-Image Leaderboard, and still suits teams who specifically want its distinct illustrative look.

Recraft V3 is a text-to-image and vector model that scored a 1172 Elo rating and a 72% win rate to top the Artificial Analysis Text-to-Image Leaderboard in October 2024. Unlike most rivals at the time, it also generates native SVG vector output instead of only raster images.

Provider: Recraft · Family: Recraft

More about Recraft on HokAI

Input modalities: text, image · Output: image

About Recraft V3

Recraft V3 is the third-generation image model from the London AI design startup Recraft, released on October 30, 2024 under the development codename Red Panda before the company confirmed authorship. It is a text-to-image and text-to-vector model, reported at roughly 20 billion parameters, that generates raster images (JPEG/PNG) and native vector graphics (SVG) from a text prompt or a reference image, sitting in Recraft's lineup between the 2024 V2 release and the 2026 V4 and V4.1 generations. V3 took first place on the Artificial Analysis Text-to-Image Leaderboard on release, scoring a 1172 Elo rating with a 72% win rate in head-to-head human comparisons, ahead of Flux1.1 Pro at 1143 Elo and Midjourney v6.1 at 1093 Elo. Independent coverage at the time confirmed it also beat DALL-E 3, Stable Diffusion and Ideogram on the same leaderboard, which is why the model spent weeks circulating anonymously as the mystery entry Red Panda before Recraft claimed it. As an image model rather than a language model, V3 has no token context window in the LLM sense; input is a text prompt plus, optionally, one or more reference images for style matching. Recraft's docs describe an Artistic level parameter that trades conventional composition for more unusual angles and framing at higher settings, and the model supports a curated style library plus an effectively unlimited library of AI-generated custom styles. V3 accepts text and reference-image input and produces image output only: no audio, video, or code. It was the first Recraft model able to render mid-size text accurately inside a generated image and, at launch, the only model on the market able to place that text at a specific position in the frame rather than somewhere approximate. Two output modes exist: Recraft V3 (raster JPEG/PNG) and Recraft V3 Vector, which emits native SVG paths rather than a traced raster file, with far fewer anchor points than a typical auto-traced vector, cutting real cleanup time for a designer. Via the Recraft API, a raster generation is priced per image and takes about 7 seconds to return; the vector variant costs roughly double the price and about twice the latency, close to 16 seconds. Third-party hosts including fal.ai, Replicate and OpenRouter mirror the same per-image pricing as of 2026 rather than setting their own rates. Inside the Recraft Studio subscription, V3 draws from the same monthly credit pool as the other model generations rather than being billed separately. V3 is reachable through Recraft's own API, through the Recraft Studio web and mobile apps, and through third-party inference hosts fal.ai, Replicate and OpenRouter, each with their own API key and billing. It is closed-weight: there is no downloadable checkpoint, no self-hosting path, and no quantized or fine-tuned variant, so every request runs on Recraft's or a licensed host's infrastructure. Recraft has not published a system card or named red-team partners specific to V3. At the company level, Recraft states that user-generated and uploaded assets are not used to train external models and that personal data handling follows GDPR, including honoring deletion requests; the company's SOC 2 Type II audit and AIUC-1 certification were completed after V3's release and apply to the platform as a whole rather than to V3 specifically. V3 remains a reasonable pick for teams that specifically want its more stylized, illustrative look and its lower per-image cost and faster generation time than the newer V4 family. Teams that need V3's benchmark-topping photorealism updated for 2026, or better compositional control, should move to Recraft V4 or V4.1 instead; teams that need the single most photorealistic raster image regardless of vendor should still benchmark against Midjourney and Flux directly, since both remain close competitors on the same leaderboard V3 once led. Recraft does not disclose V3's training data sources or a specific training cutoff date. The company's general data policy applies: assets a user uploads or generates are not used to train models outside the account, and EU users can request deletion under GDPR. No HIPAA or dedicated enterprise data-residency commitments are published for V3 specifically. V3 was followed by Recraft V4 on February 17, 2026, which Recraft documents as a ground-up rebuild improving photorealistic rendering, text-in-image accuracy and compositional control, and by Recraft V4.1 on May 14, 2026, a refinement pass with stronger prompt understanding. Recraft's own model docs now position V4 and V4.1 as the default choice for most generation tasks, with V3 kept live mainly for its distinct stylization and lower cost.

Pricing

Recraft V3 costs $0.04 per raster image (40 API credits, about 7s) and $0.08 per vector image (80 credits, about 16s) via the Recraft API; the same $0.04/$0.08 rates are mirrored by fal.ai, Replicate and OpenRouter. Inside Recraft Studio, V3 draws from the same monthly credit plans as other Recraft models rather than being billed separately.

Key Features

• Native vector output: Recraft V3 Vector emits real SVG paths with substantially fewer anchor points than a traced raster file, so exports are clean enough to hand off to a designer directly.
• Positioned text rendering: V3 was the first Recraft model able to place readable text at a specific spot in a generated image rather than somewhere approximate in the frame.
• Artistic level control: An Artistic level parameter trades conventional, direct compositions at low settings for unusual angles and perspectives at high settings, in the same request.
• Style library and custom styles: Pick from a built-in style catalog or generate a new AI style on demand, then lock a prompt onto that specific visual language without retraining anything.
• Multi-host API access: Available through Recraft's own API as well as fal.ai, Replicate and OpenRouter, each running the same model on its own infrastructure and billing.

Pros

• Topped the Artificial Analysis Text-to-Image Leaderboard on release, beating the leading Flux and Midjourney models in head-to-head comparisons.
• Native SVG vector output means exported logos and icons are already editable paths, not a raster file that needs manual tracing.
• Priced per image rather than per subscription seat, so occasional or API-only users are not forced onto a monthly plan.

Cons

• Superseded by Recraft's newer V4 generation for photorealism and compositional accuracy, so new projects generally default to it instead.
• Recraft has never published a dedicated safety writeup or named outside evaluators for this specific generation, unlike frontier language models from larger labs.
• Nothing about the weights is downloadable or self-hostable, so uptime for every request depends entirely on Recraft's servers or a third-party host's.

Benchmarks

• artificial analysis image arena elo: 1172
• artificial analysis image arena win rate pct: 72

Visit Recraft V3 Official Page