Grok Imagine

Grok Imagine is a generative media product from xAI, the company founded by Elon Musk. It produces still images, short videos with synchronized audio, and image edits from natural language prompts or reference pictures. The tool grew out of the image generation features added to Grok in late 2024 and reached a tagged 1.0 release on February 2, 2026, after a public rollout that began in August 2025.^[4] By that point xAI said the product was generating roughly 1.245 billion videos every 30 days, putting it among the most heavily used consumer video models on the market.^[4]

Grok Imagine combines a text-to-image stack derived from the Aurora (xAI) model with a separate video generation pipeline. The system runs inside the Grok mobile and web apps and, since January 28, 2026, is also exposed through the xAI Imagine API.^[2] The product is positioned against Sora 2 from OpenAI, Veo 3 from Google DeepMind, and Google's lightweight Nano Banana editor, with xAI emphasizing low latency, looser content rules, and per-second pricing well below comparable services.

The rollout has been overshadowed by safety controversies. A preset called "spicy mode," enabled by default in early versions, generated non-consensual nude videos of celebrities including Taylor Swift on the day Grok Imagine became available to subscribers.^[6] Reporters at outlets like The Verge, TechCrunch, and Common Dreams documented the issue within hours of launch, and regulators in the United Kingdom and France later opened inquiries into whether the tool produced illegal sexual content involving minors.^[10] Despite the criticism, xAI has continued to expand the feature set and the underlying API.

Background

xAI was incorporated in March 2023 and shipped its first chatbot, Grok, on the X platform in November of that year. The earliest versions of Grok handled only text. Image generation arrived in August 2024 through a partnership with Black Forest Labs, which let users call the FLUX.1 model directly from inside Grok's chat interface. That arrangement was always treated as a placeholder. Musk had said publicly that xAI wanted to control its own image stack, and the company began training a proprietary diffusion-style model called Aurora (xAI) during the second half of 2024.^[16]

Aurora went live on December 9, 2024, replacing FLUX inside Grok on the web and on X.^[1] xAI described Aurora as an autoregressive mixture-of-experts network trained to predict the next token from interleaved text and image data, rather than a conventional latent diffusion model.^[13] The model focused on photorealism, accurate rendering of small text, faithful reproduction of logos, and recognizable celebrity likenesses.^[15] It was a deliberately permissive system. Aurora allowed images of public figures and copyrighted characters that competing tools refused to produce, and the only hard limit at launch was a refusal to generate full nudity.^[14] The model reached the xAI public API on March 21, 2025.^[16]

The shift from a still-image feature to a full creative suite happened in mid-2025, when xAI began testing video generation. Musk previewed the work on X in July,^[18] calling it "an AI Vine" in a nod to the defunct six-second video service. The product was rebranded "Grok Imagine" and shipped to subscribers on August 4, 2025.^[5] From that point the term refers to the combined image, video, and editing surface rather than the underlying Aurora model alone.

Version timeline

Scroll sideways for more →

The February 2026 "1.0" label was, to a degree, a marketing choice. The product had been live for six months and had accumulated three sets of major capability changes before xAI was willing to drop the version prefix. Internally the company has continued to ship smaller updates roughly every two weeks, and the docs page lists the API as the canonical source for current limits.

Video 1.5 release

Release and model contract

xAI made Grok Imagine Video 1.5 generally available through the Imagine API on June 16, 2026, after a preview had reached Vercel AI Gateway on June 3 ^[23][31]. The stable xAI model identifier is grok-imagine-video-1.5. Its documented aliases are grok-imagine-video-1.5-preview and the dated snapshot grok-imagine-video-1.5-2026-05-30 ^[24]. Vercel's provider-specific preview identifier is xai/grok-imagine-video-1.5-preview, which should not be substituted for the stable identifier in a direct xAI API request ^[31].

The 1.5 model has a narrower input contract than the older grok-imagine-video endpoint. It takes one starting image plus a natural-language motion prompt and produces a video with synchronized audio. The image can be supplied through a public URL, a base64 data URI, or an xAI Files API identifier ^[24][35]. xAI explicitly states that 1.5 does not support text-to-video. Text-only prompts and video-input operations remain available through the older model ^[24][25].

Scroll sideways for more →

For image-to-video, the source image's ratio is the default. Supplying an aspect_ratio overrides it and stretches the source to the requested shape. The 1080p option is specific to Video 1.5 image-to-video and does not add 1080p to the older text-to-video, editing, or extension paths ^[26].

xAI also rolled out Video 1.5 Fast on grok.com and in the Grok iOS and Android apps. The company says Fast produces a six-second 720p video in about 25 seconds, down from more than 40 seconds for its preceding consumer model ^[23]. xAI has not documented Fast as a separate public API model ID or assigned it separate per-second API pricing. The GA API model and consumer Fast variant should therefore be treated as related but distinct serving options.

Workflow boundaries

The broader Imagine API can generate from text, edit a video, continue its final frame, or use multiple reference images, but xAI's documentation assigns those workflows to grok-imagine-video, not Video 1.5. This distinction matters because the model names have different inputs, limits, resolutions, and prices ^[25][26].

Scroll sideways for more →

Editing ignores custom duration, aspect-ratio, and resolution parameters ^[28]. Extension's duration setting controls only the new segment, so extending a ten-second input by five seconds returns a combined 15-second video ^[29]. Reference-to-video differs from Video 1.5's single starting image because reference images guide subjects or style without becoming the first frame ^[27].

API pricing and access

Video 1.5 uses resolution-based output pricing. xAI charges $0.01 for the starting image, then $0.08 per generated second at 480p, $0.14 at 720p, or $0.25 at 1080p ^[24][25]. A maximum-length 15-second output therefore costs $1.20, $2.10, or $3.75 respectively, plus the image-input charge. The calculation excludes any separate provider fee charged by a third-party gateway.

The older grok-imagine-video endpoint remains less expensive and supports more input modes. Its 480p and 720p output rates are $0.05 and $0.07 per second, with image inputs at $0.002 each and video inputs at $0.01 per second ^[25]. The price difference is another reason not to treat the two identifiers as interchangeable.

The xAI SDK submits the asynchronous request and polls for completion. Direct REST clients receive a request ID and must poll until the job is done, failed, or expired. Completed videos are delivered through temporary URLs, so applications must download them promptly or use the Files API storage options when persistent storage is required ^[26].

Evaluation, provenance, and safety

Artificial Analysis evaluates image-to-video systems using blind pairwise votes: participants compare videos generated from the same source image without seeing the model names, and the service converts aggregate preferences into Elo ratings ^[30]. On July 24, 2026, its with-audio leaderboard placed grok-imagine-video-1.5-preview third with an Elo of 1117, a plus or minus 10 confidence interval, and 3,821 samples. Its no-audio view gave the model an Elo of 1330, tied with PixVerse V6 and behind Gemini Omni Flash and Dreamina Seedance 2.0 720p ^[30].

Those two scores come from different comparison sets and are not directly comparable. They also show why launch-period descriptions of the model as the number-one image-to-video system should be dated: Arena rankings change as models and votes are added. The method measures aggregate viewer preference, not precise physical accuracy, prompt compliance, edit fidelity, safety, or suitability for a particular production pipeline.

xAI says Video 1.5 improves motion, physical consistency, audio, speech clarity, and synchronization over its predecessor ^[23]. Vercel independently confirmed the preview's image-to-video contract and single-pass synchronized audio ^[31], but xAI did not release an architecture paper, training-data description, parameter count, controlled benchmark suite, or dedicated Video 1.5 model card with the GA announcement.

For provenance, xAI's consumer FAQ says generated images and videos include a visible Grok watermark and provides no setting to remove it. The Acceptable Use Policy also prohibits stripping, altering, or circumventing embedded provenance metadata or watermarks ^[32][33]. xAI does not identify those signals as C2PA or another named cryptographic standard.

Enabling NSFW content does not disable moderation. xAI says child sexual abuse material, sexual content involving minors, and non-consensual intimate imagery remain prohibited under all settings. Its policy also bars nudifying real people, deceptive impersonation, pornographic depictions of identifiable people, falsified evidence, and attempts to bypass safeguards ^[32][33]. The Imagine API subjects generated media to content-policy review and says that generated media is not used for training ^[34]. These are product-level controls; xAI has not published Video 1.5-specific refusal rates, deepfake tests, false-positive measurements, or red-team results.

Capabilities

Grok Imagine is organized around four core actions: generate a still image, edit an existing image, generate a short video from text, and animate or edit a video starting from an image or earlier clip. All four are reachable from the same prompt box in the consumer app, and the API exposes them through two model endpoints.

Scroll sideways for more →

Audio quality has been the most-tracked metric across versions. The August 2025 launch promised "native audio" but produced muddy ambient tracks and unconvincing lip sync. The 1.0 release added what xAI called "smoother motion" and clearer dialogue, though reviewers at Tom's Guide and Latent Space still found the lip sync trailing Veo 3 by a noticeable margin.^[11][12] Image quality, by contrast, has been competitive with leading models since the Aurora launch, particularly on photorealistic portraits and on small details like signage and product packaging.

The original grok-imagine-video endpoint remains capped at 720p, but Grok Imagine Video 1.5 added 1080p for image-to-video generation. The higher resolution does not apply to text-to-video, video editing, or extension, and xAI prices it separately at $0.25 per generated second ^[24][25][26]. xAI has not documented 4K output.

Aurora architecture and engine

The still-image side of Grok Imagine runs on Aurora, which xAI describes as an autoregressive mixture-of-experts network. Rather than denoising a latent representation in the way most diffusion models do, Aurora predicts the next token in an interleaved sequence that contains both text and image patches. The architecture is closer in spirit to multimodal language models like the original GPT-4o image system than to Stable Diffusion or FLUX. xAI has not published the parameter count or the exact training data mix, though it has confirmed that the model was trained from scratch on "billions of examples from the internet."^[13]

The video pipeline is a separate model that takes the still output from Aurora and extends it forward in time. The original API exposed it as grok-imagine-video; xAI later gave its image-to-video successor the public name Grok Imagine Video 1.5 and the stable identifier grok-imagine-video-1.5 ^[23][24]. According to a Latent Space post that drew on a developer briefing, the video model was trained on a cluster built around 110,000 NVIDIA GB200 GPUs, which would make it one of the largest single-purpose training jobs publicly disclosed for a video model.^[11] That figure has not been independently confirmed by xAI's own publications. Treat it as the company's own claim rather than a verified count.

Audio generation runs as a third stage. After a clip's visual frames are produced, the audio model conditions on the visual content and the original text prompt to generate matched ambient sound, music, and dialogue. The result is muxed with the video before delivery. xAI has not disclosed whether the audio model is a fine-tune of the speech systems that run inside the Grok voice assistant or a separate model entirely.

API and pricing

The Grok Imagine API opened on January 28, 2026 and is documented at docs.x.ai.^[22] It uses asynchronous polling. A client submits a request, receives a job identifier, and then polls for completion. Input images can be supplied as public URLs or as base64 data URIs. The API ships SDKs for Python through both the xAI client and the OpenAI compatibility shim, JavaScript through the Vercel AI SDK, and plain REST through curl.^[3]

Scroll sideways for more →

The $4.20 per minute headline is roughly one third of Google's Veo 3.1 Preview, which lists at $12 per minute with audio, and around one seventh of OpenAI Sora 2 Pro at $30 per minute.^[11] The $4.20 number is also unmistakably an Elon Musk price point; he has used 420 in product launches before, including the brief 2018 attempt to take Tesla private at $420 per share. Whether you find that funny or tiresome probably depends on how long you have been paying attention to the company.

Consumer access to Grok Imagine is bundled into Grok's subscription tiers. The current lineup runs from a free tier with hard rate limits, through SuperGrok Lite at $10 per month, SuperGrok at $30 per month or $300 per year, up to SuperGrok Heavy at $300 per month. The free tier permits roughly 10 image generations every two hours. SuperGrok unlocks unlimited image generation and around 100 video renders per day. SuperGrok Heavy raises those limits substantially and is the only tier with full access to Grok 4 Heavy reasoning alongside Imagine. Premium+ subscribers on the X platform get a more limited Grok Imagine quota as a bundled feature.^[17]

For enterprise customers, xAI sells the API on standard pay-as-you-go terms with no minimum commit. Volume discounts are negotiated directly. The company has not published an SLA for video generation latency, which currently runs between roughly 15 seconds and 90 seconds depending on clip length and queue depth.

Controversies

Spicy mode and non-consensual deepfakes

Grok Imagine launched on August 4, 2025 with four content presets: fun, normal, custom, and "spicy."^[5] The spicy preset was meant to allow suggestive or sexually charged content within xAI's posted policy, which still formally banned "depicting likenesses of persons in a pornographic manner." The reality on launch day was very different. Jess Weatherbed of The Verge, working on a story about the new product, wrote that spicy mode produced topless videos of Taylor Swift the first time she tried it, without any nudity-related wording in the prompt.^[6][8] Common Dreams, the Penn State Digital Shred blog, and the AI Incident Database documented similar bypasses against other female celebrities including Scarlett Johansson.^[7][9]

Musk's response was muted. xAI updated the content filters to add stronger refusals for named celebrities, but reviewers continued to find ways around them for several weeks. Haley McNamara of the National Center on Sexual Exploitation said in a statement that "xAI appears to be doubling down on furthering sexual exploitation by enabling AI videos to create nudity." The Hollywood Reporter and Deadline ran pieces speculating about lawsuits, though as of May 2026 no celebrity has filed a public complaint against xAI over Imagine output specifically.^[21]

Regulatory action and minors

In January 2026, Euronews reported that researchers had used Grok Imagine to produce sexually explicit imagery of women and what appeared to be minors. The investigation prompted the United Kingdom's Online Safety regulator and France's Arcom to open formal inquiries into whether Grok Imagine was producing illegal content under each country's child sexual abuse material rules.^[10] The story landed days before the 1.0 release and turned what would have been a victory lap into a more defensive launch, with xAI publishing an updated content policy alongside the version bump.

xAI has said it scans output for child sexual abuse material using industry-standard hashing and that it bans the relevant prompt patterns. Independent researchers at the Stanford Internet Observatory and the Center for Countering Digital Hate have disputed how aggressive those filters actually are in practice. The dispute is ongoing.

Moderation philosophy

The deeper question across all of these stories is whether xAI's stated philosophy of "maximally truth-seeking, minimally censored" output is compatible with the kind of harm reduction other model providers attempt. Musk has been explicit, on X and in interviews, that he sees most existing AI content filters as politically biased. Grok Imagine sits at the sharp end of that view. The product is more permissive than Sora 2 or Veo 3 by design, and the rollout has been a long, public experiment in what happens when you ship that posture to a mass audience. The cost so far has been measured in regulatory inquiries and reputational damage rather than in product revenue, which has continued to grow.

Comparison to competitors

The AI video generation market in early 2026 sits roughly in three tiers. Sora 2 and Veo 3 are the high quality, expensive, more cautious options. Grok Imagine, Bytedance's Seedance, Alibaba's Wan 2.5, and a handful of others target the middle on price and creative latitude.^[20] Nano Banana is a separate category entirely, focused on lightweight edits and stylized clips rather than full text-to-video.

Scroll sideways for more →

Reviewers tend to give Sora 2 the edge on physical realism and Veo 3 the edge on cinematic camera moves and synchronized lip sync. Grok Imagine wins on price, on raw throughput, and on willingness to render almost any subject the user asks for. Tom's Guide ran a seven-prompt comparison in November 2025 and gave Sora 2 the overall win, though Grok Imagine produced two clips the reviewer preferred and did so at roughly a tenth of the cost.^[12] The Latent Space writeup in February 2026 went further and called Grok Imagine "the #1 video model on cost and latency," while acknowledging that on best-of-three quality it still trailed the leaders.^[11]

The comparison with Nano Banana is more lopsided in xAI's favor, because Nano Banana is not really a competing product. Google positioned Nano Banana as a fast, conversational image editor sitting inside Gemini, with short video generation as a secondary feature. Grok Imagine targets the full pipeline. Where the two overlap is in single-shot meme-style clips, and there Nano Banana's tighter integration with the Gemini conversational model gives it an edge in iteration speed for users who already live inside Google's stack.

Reception

Reception of Grok Imagine has been split along familiar lines. Developers and power users on X have been broadly positive, citing the price, the API ergonomics, and the absence of the kind of refusals that have frustrated heavy users of competing services. The Latent Space and Hacker News threads around the API launch were dominated by people running the same prompts on Grok Imagine, Sora 2, and Veo 3 and posting the results. Most of those side-by-side comparisons concluded that Grok Imagine was usable for a wider range of subjects than either competitor.

Mainstream press coverage has been harsher. The Verge, TechCrunch, Common Dreams, Deadline, and Euronews have all run multiple critical pieces, and the framing in those stories has settled into a template. The reviewer tries the product, finds an objectionable output within minutes, and writes the story around that finding. xAI has not run a sustained public response to the press cycle beyond Musk's own posts on X, which generally either ignore the criticism or frame it as legacy media bias.

The usage numbers suggest that neither the controversies nor the resolution cap have meaningfully slowed adoption. The 1.245 billion videos in 30 days figure from February 2026 is roughly an order of magnitude above what Runway, Pika, and Luma have publicly disclosed for the same period.^[4] Some of that volume is driven by the bundled access for Premium+ subscribers on X, and the per-user generation count has not been broken out, so the comparison should be read with caution. Even discounting for that, Grok Imagine is on track to be the highest volume consumer video generation product in the market, at least until OpenAI ships the consumer build of its Sora app to a broader audience.

Whether that volume converts to durable revenue is the open question. xAI has not disclosed revenue from the Imagine product line separately, and the company's overall financials remain private. The most credible read at the moment is that Imagine is a strategic loss leader designed to anchor users to the Grok subscription bundle and to give the X platform a creative tool that none of its competitors can match for free or near-free output. The pricing on the API supports that read, and so does the speed at which xAI has shipped features like SuperGrok Lite that lower the entry barrier even further.

See also

• xAI
• Grok
• Grok 4
• Elon Musk
• Aurora (xAI)
• Sora 2
• Veo 3
• Nano Banana
• Seedance
• Wan 2.5
• AI agents

References

1. ^xAI. "Grok Image Generation Release." December 9, 2024. x.ai/...grok-image-generation-release
2. ^xAI. "Grok Imagine API." January 28, 2026. x.ai/...grok-imagine-api
3. ^xAI Docs. "Imagine Overview." docs.x.ai/...imagine
4. ^American Bazaar Online. "xAI takes 'biggest leap yet,' launches Grok Imagine 1.0." February 6, 2026. americanbazaaronline.com/...rok-imagine-1-0-474633
5. ^TechCrunch. "Grok Imagine, xAI's new AI image and video generator, lets you make NSFW content." August 4, 2025. techcrunch.com/...rator-lets-you-make-nsfw-content
6. ^The Verge. "Grok's 'spicy' video setting instantly made me Taylor Swift nude deepfakes." August 2025. theverge.com/...imagine-video-generator-spicy-mode
7. ^Common Dreams. "Safeguards? What Safeguards? Grok's New 'Spicy Mode' Makes Nude Taylor Swift Deepfakes." August 2025. commondreams.org/...taylor-swift-nude-deepfakes
8. ^Musically. "Grok 'spicy mode' accused of making NSFW Taylor Swift deepfakes." August 6, 2025. musically.com/...aking-nsfw-taylor-swift-deepfakes
9. ^AI Incident Database. "Incident 1165: Grok Imagine Reportedly Produces Non-Consensual Taylor Swift Deepfake Nudes Without Explicit Prompting." incidentdatabase.ai/...1165
10. ^Euronews. "Grok under fire for generating sexually explicit deepfakes of women and minors." January 5, 2026. euronews.com/...icit-deepfakes-of-women-and-minors
11. ^Latent Space. "SpaceXai Grok Imagine API: the #1 Video Model, Best Pricing and Latency." January 2026. latent.space/...ainews-spacexai-grok-imagine-api
12. ^Tom's Guide. "How does Grok Imagine compare to Sora 2? Here's what happened when I ran 7 tests." tomsguide.com/...prompts-and-theres-a-clear-winner
13. ^The Decoder. "xAI's Aurora image model becomes official, built from scratch." December 2024. the-decoder.com/...mes-official-built-from-scratch
14. ^TechCrunch. "Elon Musk's X gains a new image generator, Aurora." December 7, 2024. techcrunch.com/...ins-a-new-image-generator-aurora
15. ^EM360Tech. "What is xAI Aurora Generator? Inside Grok's New Image Generator." em360tech.com/...-inside-groks-new-image-generator
16. ^Wikipedia. "Grok (chatbot)." Sections on Aurora and Grok Imagine. en.wikipedia.org/...Grok_(chatbot)
17. ^Fello AI. "Grok Pricing 2026: SuperGrok, X Premium+, Heavy & API Costs." felloai.com/grok-pricing
18. ^mlq.ai. "Grok to Launch Text-to-Video Generation with Sound via Imagine Feature in October." 2025. mlq.ai/...ith-sound-via-imagine-feature-in-october
19. ^Times of AI. "Elon Musk Unveils Grok Imagine v0.9; Here's What's New." October 2025. timesofai.com/...elon-musk-unveils-grok-imagine-v0-9
20. ^WaveSpeed Blog. "Grok Imagine Video vs Sora 2, Veo 3.1, Seedance 1.5, WAN 2.5/2.6, and Vidu Q3: Complete Comparison." 2026. wavespeed.ai/...-seedance-wan-vidu-comparison-2026
21. ^Deadline. "Elon Musk's Latest AI Frontier: 'Spicy' Deepfakes Of Stars Like Scarlett Johansson & Taylor Swift." August 2025. deadline.com/...-johansson-taylor-swift-1236480553
22. ^SQ Magazine. "xAI Launches Grok Imagine API to Rival Google and OpenAI in Video Generation." sqmagazine.co.uk/xai-grok-imagine-api-video-launch
23. ^xAI. "Grok Imagine Video 1.5." June 16, 2026. x.ai/...grok-imagine-video-1-5
24. ^xAI Docs. "grok-imagine-video-1.5." docs.x.ai/...grok-imagine-video-1.5
25. ^xAI Docs. "Pricing." docs.x.ai/...pricing
26. ^xAI Docs. "Video Generation." docs.x.ai/...generation
27. ^xAI Docs. "Reference-to-Video." docs.x.ai/...reference-to-video
28. ^xAI Docs. "Video Editing." docs.x.ai/...editing
29. ^xAI Docs. "Video Extension." docs.x.ai/...extension
30. ^Artificial Analysis. "Image to Video Leaderboard." artificialanalysis.ai/...image-to-video
31. ^Vercel. "Grok Imagine Video 1.5 on AI Gateway." June 3, 2026. vercel.com/...grok-imagine-video-1-5-on-ai-gateway
32. ^xAI. "FAQ - Grok Website / Apps." docs.x.ai/...faq
33. ^xAI. "Acceptable Use Policy." Effective June 26, 2026. x.ai/...acceptable-use-policy
34. ^xAI Docs. "Imagine Overview." docs.x.ai/...imagine
35. ^xAI Docs. "Image-to-Video." docs.x.ai/...image-to-video