by Recraft
Recraft V3 review, pricing and verdict
Superseded by Recraft V4 (17 February 2026) and V4.1 (14 May 2026), but still sold, and the only Recraft API line with inpainting, outpainting and exact text placement.
- Best for
- Exact text placement in a generated graphic
- Not for
- Long prompts
Should you use Recraft V3?
Superseded by Recraft V4 on 17 February 2026 and V4.1 on 14 May 2026, V3 is still sold with 1K output and a short prompt cap. It suits API users who need exact text placement or inpainting, and anyone wanting its Risograph style; everyone else should start from V4.1.
Best for
- Exact text placement in a generated graphic Recraft's API docs say V3 is the only line that places text at exact positions through the text_layout parameter.
- Inpainting, outpainting and background replacement over the API The docs say these operations run on V3 only.
- V3-specific styles such as Risograph and Hard flash Recraft's model guide names them as styles V3 is home to.
Not for
- Long prompts V3 caps prompts at 1,000 characters; V4 and newer accept 10,000.
- Print-resolution output V3 outputs at about 1K, while V4.1 Pro outputs 2K raster.
- Self-hosted or air-gapped deployment V3 is served only by Recraft and its hosts; no weights are offered.
Recraft V3 pricing
- Pricing model
- per-image
- Free tier
- Recraft Studio's Free plan has limited credits that reset daily; images made on it are owned by Recraft, shown in the public gallery and not licensed for commercial use. The API has no free tier.
| Tier | Rate |
|---|---|
| Raster API | $0.04 per image (40 API units), about 7 seconds per Recraft's docs |
| Vector API | $0.08 per image (80 API units), about 16 seconds per Recraft's docs |
Recraft's API charges $0.04 per raster image and $0.08 per vector (SVG) image for V3, and the same rates for V3 image-to-image, inpainting, outpainting and background edits. API units are bought in advance, never expire and are non-refundable. Third-party hosts bill separately: fal.ai shows $0.04 per image, or $0.08 with a vector style, and Replicate charges $0.08 per output image on its V3 SVG model.
What Recraft V3 can do
Recraft V3 is Recraft's October 2024 text-to-image and native SVG model, which took first place on the Artificial Analysis image leaderboard at launch with 1172 Elo, per Recraft. Recraft's docs now call it a previous generation, yet it keeps operations no newer Recraft model offers: inpainting, outpainting and exact text placement.
What goes in and what comes out
Reads
- text
- image
Writes
- image
Capabilities
- VisionSupported
- Audio inputNot supported
- Audio outputNot supported
- Code executionNot supported
- Function callingNot supported
- Structured outputNot supported
- Tool useNot supported
- Video inputNot supported
- Web browsingNot supported
Key features
Native SVG output
The recraftv3_vector model returns SVG files, and the raster model returns WebP or PNG at about 1K resolution.
Exact text placement
The text_layout parameter takes words and their positions, and Recraft's docs say V3 is the only line that places text at exact positions.
Artistic level control
controls.artistic_level takes an integer from 0 to 5: clean and static at the bottom, dynamic and eccentric at the top.
Editing operations
Image-to-image, inpainting, outpainting, replace background and generate background run on V3 only in Recraft's API.
Show all 6 features
Styles and style references
Curated styles, custom styles and 1 to 10 style reference images steer the look without retraining.
Multi-host access
The same model is served through Recraft's API, fal.ai, Replicate and OpenRouter, each with its own key and billing.
Known limitations
- Recraft's docs label V3 a previous generation behind V4.1, the default API model, so new projects are steered to the newer line.
- Prompts are capped at 1,000 characters on V3 against 10,000 on V4 and newer, and output stops at about 1K resolution with no Pro tier.
- In the API, V3 raster costs more per image than V4.1 and far more than V4.1 Flash, so the 'V3 is cheaper' argument holds only for Studio credits.
- Generated API image URLs last about 24 hours, and the API allows 100 images per minute and 5 requests per second per user.
Who it is for
- API teams that need exact text placement in generated graphics
- Developers using Recraft's inpainting or outpainting endpoints
- Designers who want V3-only styles such as Risograph and Hard flash
Recraft V3 benchmarks
Standout result
Artificial Analysis text-to-image Elo at launch, October 2024 (vendor-reported)
Source: Recraft launch post of 30 October 2024, opened 12 November 2026; The Decoder reported the same figureStrengths and limits
Strengths
- Inpainting, outpainting and background edits are offered on V3 and on no newer Recraft model, per Recraft's API docs.
- Studio's credits table prices V3 at half the credits of V4.1 and a shorter generation time, per Recraft's docs.
- Launched first on the Artificial Analysis text-to-image board, per Recraft's launch post and The Decoder.
- The only Recraft API line that supports inpainting, outpainting and background operations, and the only one that places text at exact positions, per Recraft's API docs read on 12 November 2026.
Show 2 more
- Separate raster and vector models (recraftv3 and recraftv3_vector) with an artistic level control from 0 (static, clean) to 5 (dynamic, eccentric).
- Inside Studio, Recraft's credits table prices V3 raster at 1 credit and about 4 seconds, against 2 credits and about 6 seconds for V4.1.
Limits
- Recraft's docs label it a previous generation, with V4.1 as the default API model.
- Its prompt cap and output size are below V4 and newer, and it carries no 2K Pro tier.
- Its launch-era Elo cannot be read against current Artificial Analysis figures, which sit on a newer board version.
- In the API, V4.1 raster is cheaper per image, so V3 saves money only in Studio credits.
Recraft V3 alternatives
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Recraft V3 This page | Recraft | -- | -- | text, image → image |
| Recraft V4.1Pick V4.1 for new work, longer prompts and 2K output; pick V3 only for exact text placement or inpainting through the API. | Recraft | -- | -- | text, image → image |
 FLUX.2Pick FLUX.2 if you want open-weight variants you can run yourself; pick V3 if you need Recraft's text placement or native SVG. | Black Forest Labs | -- | -- | text, image → image |
 GPT Image 2.6Pick GPT Image 2.6 if a higher Artificial Analysis score decides it; pick V3 for exact text positions and editable vector output. | OpenAI | -- | $8 / $30 | text, image → image |
Recraft V3 specs
- Latency (P50)
- 7000 ms
- Latency (P99)
- 16000 ms
- Mixture of experts
- No
- Parameters
- Not disclosed
- Architecture
- Text-layout-conditioned image generator with a language-model layout planner, per Recraft's ML team; exact architecture not disclosed
Training data
Recraft has not disclosed V3's training data. Its ML team says V3 was trained from scratch, with an in-house OCR model used to build text-layout training data.
API endpoints
| Name | Endpoint | Auth |
|---|---|---|
| Recraft API | https://www.recraft.ai/docs/api-reference/models/recraft-v3 | API token |
| fal.ai | https://fal.ai/models/fal-ai/recraft/v3/text-to-image | API key |
| Replicate (raster) | https://replicate.com/recraft-ai/recraft-v3 | API token |
| Replicate (SVG) | https://replicate.com/recraft-ai/recraft-v3-svg | API token |
| OpenRouter | https://openrouter.ai/recraft/recraft-v3 | API key |
SDKs
- Python
- JavaScript
Prompting patterns
Show 2 prompting patterns
Set the artistic level deliberately
controls.artistic_level of 0 for literal product shots, 4 or 5 for exploratory illustration
Recraft's docs say lower values give standard images and higher values give less conventional compositions and angles.
Draft in raster, finish in vector
Generate with recraftv3, then request recraftv3_vector for the approved logo or icon
The vector model costs twice as much per image in the API, so reserve it for the final asset.
Known quirks
Show 5 known quirks
API default model is V4.1, not V3
Without a model parameter the API uses recraftv4_1, unless the style names a V3 style.
Workaround: Set model to recraftv3 or recraftv3_vector explicitly in every request.
Image URLs expire after about a day
Recraft keeps generated API images for about 24 hours and says the duration may change.
Workaround: Download each file right after generation, or request b64_json to get the bytes inline.
Vector output costs twice raster
recraftv3_vector is priced at twice the raster rate, which surprises users who expect one flat per-image price.
Workaround: Iterate in raster and switch to vector only for the final asset.
Prompt cap is 1,000 characters
V3 and V2 accept up to 1,000 prompt characters; V4 and newer accept 10,000.
Workaround: Shorten the prompt or move long briefs to V4.1.
Artistic level changes composition
Higher artistic levels give less conventional compositions and more unusual angles, not only a stylized look.
Workaround: Keep it low for literal briefs and raise it only for exploration.
Recraft V3 facts
- Deployment
- API · fal.ai · Replicate · OpenRouter
- Trains on your data
- Yes
- Licence
- Proprietary (Recraft Terms of Service)
- Safety posture
- Balanced
- Released
- 30 Oct 2024
- Current version
- recraft-v3-20241030
- Family
- Recraft
- Variant
- Multimodal
- Data retention
- Recraft's docs say API inputs and outputs are never used for model training and are not stored beyond generating the result, with generated image files kept about 24 hours for download. Studio content (images, prompts, chat) is used for training by default on Free and paid accounts, opt-out is not available on the Free plan, and Teams plans are opted out by default.
Data governance
- Soc2 type2
- Yes
- GDPR compliant
- Yes
- Trains on inputs
- Yes
Version history
Show all 4 versions
recraft-v4.1 · 14 May 2026
Successor generation; Recraft's docs call it the default and most capable API model.
recraft-v4 · 17 Feb 2026
Successor generation; described in Recraft's docs as a ground-up rebuild.
recraft-v3-vector · 30 Oct 2024
Vector variant (recraftv3_vector) that returns SVG.
recraft-v3-20241030 · 30 Oct 2024
Initial release; first on the Artificial Analysis text-to-image leaderboard with 1172 Elo, per Recraft's launch post.
What changed
Questions about Recraft V3
How much does Recraft V3 cost per image?
Recraft's API charges $0.04 for a raster image and $0.08 for a vector (SVG) image, and the same rates apply to V3 image-to-image, inpainting, outpainting and background edits. Recraft's current API table prices V4.1 at $0.035 per raster image and V4.1 Flash at $0.007. In Recraft Studio a V3 raster generation costs 1 credit and a vector one 2 credits. fal.ai and OpenRouter list $0.04 per image, and Replicate's V3 SVG model $0.08.
Is Recraft V3 free to use?
Recraft Studio has a Free plan with limited credits that reset daily, and Recraft's credits table lists V3 among the models those credits buy. Images made on the Free plan are owned by Recraft, appear in the public community gallery and are not licensed for commercial use. The API has no free tier: an API token needs a positive API-units balance.
Is Recraft V3 open source, or can I run it myself?
No. Recraft's docs, read on 12 November 2026, list no weights download and describe access only through Studio and the API, while fal.ai, Replicate and OpenRouter serve the same model on their own billing. In practice every request runs on Recraft's or a host's servers.
Should I still use Recraft V3 now that V4.1 is out?
Recraft's docs call V3 a previous generation and make V4.1 the API default, and V3 caps prompts at 1,000 characters against 10,000 for V4 and newer, with output at 1K where V4.1 Pro reaches 2K. V3 is the only API line with inpainting, outpainting and background operations, and the only one that places text at exact positions. In Studio, Recraft's credits table puts V3 at 1 credit and about 4 seconds against 2 credits and about 6 seconds for V4.1.
Does Recraft train on images I make with V3?
It depends where you generate. Recraft's docs say API inputs and outputs are never used for model training, while Studio content (images, prompts and chat) is included by default on Free and paid accounts, opt-out is not available on the Free plan, and Teams plans are opted out by default. The privacy policy effective 20 October 2026 draws the same line between the API and Studio.
About Recraft V3
Recraft V3 is the image model Recraft released on 30 October 2024, after it circulated anonymously under the codename red_panda; Dataconomy's May 2025 report says that entry topped the Artificial Analysis benchmark. Recraft's docs now list V3 as a previous generation. Recraft V4, called a ground-up rebuild in its docs, followed on 17 February 2026, and Recraft V4.1 on 14 May 2026 became the default and most capable API model, with V4.1 Flash added in September 2026. V3 has not been retired: Recraft's pricing page says all models across the V2, V3, V4 and V4.1 generations are available through the API (read 12 November 2026).
Read the full background
Recraft's launch post says V3 took first place on Artificial Analysis's text-to-image leaderboard with an Elo of 1172 after four days of blind voting, and The Decoder reported the same figure in October 2024, ahead of Flux and Ideogram. Recraft's docs add that it held first place for five months. These are launch-era figures from the vendor and one outside report. V3 is absent from the Artificial Analysis text-to-image board page (labelled v2.1) read on 12 November 2026, which lists the V4.1 variants. OpenRouter's public V3 page, opened the same day without an account, cites an Artificial Analysis Elo of 869 for V3 and a median end-to-end latency of 7.59 seconds, so the launch-era figure and today's are not on a confirmed common scale.
V3 takes a text prompt of up to 1,000 characters and can use reference images. It returns raster images at about 1K resolution (WebP or PNG) from the recraftv3 model and SVG from recraftv3_vector, with curated and custom styles, style references of 1 to 10 images, negative prompts, a text_layout parameter for word positions and an artistic level control from 0 (static, clean) to 5 (dynamic, eccentric). Recraft's ML team wrote that V3 pairs a language-model layout planner with an image generator conditioned on that text layout, that it was trained from scratch, and that an in-house OCR model built the layout data; the training data itself is not disclosed.
The case for V3 today is narrow. Recraft's API docs say inpainting, outpainting and the background operations run on V3 only, and that no other line places text at exact positions. Recraft's model guide also names V3 as home to styles such as Risograph and Hard flash, and says it costs less and generates faster than V4 inside Studio. The model is reachable through the Recraft API, Studio on web and mobile (see the Recraft tool page), and the hosts fal.ai, Replicate and OpenRouter. Recraft's docs (read 12 November 2026) list no weights download, so the model is closed-weight.
Who should skip it: anyone starting a new project, since Recraft itself points to V4.1 as the default, and anyone who needs prompts beyond V3's cap, 2K output or the newest photorealism. Among rivals, FLUX.2 from Black Forest Labs has open-weight variants, and GPT Image 2.6 from OpenAI scored 1261 Elo on the Artificial Analysis board on 12 November 2026. For the wider field see the image models directory and the best AI image generators guide. Pick V3 when the job needs exact text placement or inpainting through the API; otherwise start from V4.1.
Sources and guides
Official sources
- Recraft V3 model page Vendor
- Licence text Licence
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Recraft's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Hokai guides that mention Recraft V3
Lists that include Recraft V3
The provider
More AI Models on Hokai
Seedream 4.5ByteDance's Seedream 4.5, announced by BytePlus on 3 December 2025, takes up to 10 reference images; Artificial Analysis scores it 1030 Elo for text-to-image.- Seedream 5.0 LiteByteDance Seed's February 2026 Lite model reasons over prompts and can search the web; Artificial Analysis scores it 1020 Elo for text-to-image.
- Seedream 5.0 ProByteDance Seed's July 2026 image model adds point and lasso editing and layer separation; Artificial Analysis scores it 1102 Elo for text-to-image.
- Seedream 5.5ByteDance Seed's November 2026 image model: native RGBA output and multi-round editing, ranked by Artificial Analysis at 1135 Elo.
MAI-Image-2.6MAI-Image-2.6 is Microsoft AI's September 2026 image model for text-to-image and editing, with a 32,000-token context window.
Nano Banana 2.1Google's image generation and editing model in the Gemini API, with 1K to 4K output and multi-image reference mixing, generally available since October 2026.