AI rendering is the first genuinely new thing to happen to architectural visualization pricing in a decade, and it has produced a genuinely confusing market: tools that cost $30 a month sit next to studio quotes in the thousands, both calling their output “renders.” If you’re a developer, architect, or marketer trying to decide what to use where, most of what you’ll read is written by one side of the fight. This is our attempt at the honest version — what the tools actually are and cost in 2026, what they deliver task by task, and the two things almost nobody selling either option will tell you: where the real dividing line sits, and what the new disclosure rules mean if the images are selling property.
The 2026 AI rendering toolbox, priced
The current tools worth knowing, at their published prices: Veras, now part of Chaos, runs inside Revit, SketchUp, and Rhino and re-imagines whatever’s in your active viewport. Chaos restructured its plans in May 2026: Starter is $202.80 a year and Pro $348 a year, billed annually. LookX offers architecture-trained generation from around $20 a month. ArkoAI renders SketchUp/Rhino/Revit viewports from around $39 a month (widely listed, though not on a publicly indexed official pricing page at the time of writing). Fenestra turns sketches, CAD views, or single images into renders and short AI walkthrough clips from $8–$48 a month on annual billing — its per-generation cost works out to a few cents. Midjourney, at $10 a month, remains the mood-board king; even Chaos’s own comparison concedes it belongs to concept work rather than design development.
Notice what all of these have in common: they work from an image — a viewport, a sketch, a photo. What each of them finally produces is an image, not your authoritative dimensioned model — even when the source is a live Revit, SketchUp, or Rhino viewport. That single architectural fact explains almost everything about where they shine and where they collapse, and it’s worth keeping in mind that the most detailed tool comparisons published this year come from Chaos, which owns both V-Ray and Veras — vendor guidance, useful but not neutral.
The task ledger: what AI delivers, task by task
Skip the philosophy; here’s the working split we’d defend line by line:
- Mood and concept imagery — AI, clearly. Chaos’s 2026 AI in Architecture report — a vendor survey, worth the same squint as its tool comparisons — has 48% of AI-using architects naming concept ideation as the biggest time saving, and that matches the tools’ design: fast, cheap, plausible options while nothing is fixed yet.
- Massing and option studies — AI, with a supervisor. Same report, same caveat: 70% of architects say AI visualizations need constant professional supervision to reflect design intent.
- Material and lighting moods — AI is usable, as long as nobody downstream treats the output as a specification.
- A pre-sale marketing set — the model as the source of truth. AI can still help with enhancement and clean-up, but geometry, cameras, and approved design decisions stay anchored to it. The image is a representation of something a buyer will pay for. Geometry drift — the roof edge that moved, the mullion pattern that changed, the door that appeared — is documented across independent 2026 write-ups, and the tools themselves admit it: every product that ships a “geometry override” slider is confessing what the default does.
- Revision rounds — the model, and it’s not close. “Move that window 300mm and keep everything else identical” is a sentence no image-based tool in the list above documents an answer to. These tools do localized edits now — Veras will re-render a selected part of an image against a new prompt — but that’s a different thing from a dimension-controlled revision in the source model. PIXREADY published a concrete renovation-edit case in 2025 where exactly this failed.
- A brand-consistent campaign set — the model. Veras 4.3 added Design Lock specifically to improve consistency between views. Chaos puts its success rate at about 75% — a real improvement, and still a mitigation rather than a guarantee. Campaign work is a guarantee business.
- Walkthrough video — know what you’re buying. AI walkthrough tools synthesise motion from a source image — Veras’s image-to-video needs a finished render to work from. The output can look convincing, but it’s generated video rather than a camera travelling through persistent 3D geometry: the clip itself has no navigable scene behind its frames. Real architectural animation is priced the way it is because the building actually exists in 3D.
The dividing line isn’t “concepts vs finals”
Everyone in this debate — including the AI vendors — has settled on the same slogan: AI for concepts, studio for finals. It’s not wrong, but it hides the real test, which is simpler and sharper: does the image have to match a design that exists? A mood board has no ground truth; nothing in it can be “wrong.” A planning view, an investor set, a show-flat interior — those are representations of a specific, dimensioned, approved design, and every pixel that contradicts the drawings is a liability someone eventually finds. The moment your image acquires ground truth, AI can no longer be the source of truth — not because the pixels aren’t pretty, but because accuracy and revisability are the actual product. AI’s real product isn’t images; it’s optionality early, when options are cheap.
The 2026 disclosure layer nobody puts in the brochure
If the images sell property, this section matters more than the cost one. California’s AB 723 took effect in January 2026. It applies to a real estate broker or salesperson, or someone acting on their behalf, advertising the sale of real property: where an existing image has been altered to add, remove, or change elements, the disclosure has to sit on or adjacent to that image, with a path to the original unaltered version. Wisconsin’s 2025 Act 69 follows on 1 January 2027, with a narrower trigger: it bites where the alteration creates a false or misleading impression of the property. New York’s Department of State issued a consumer alert in late 2025 warning about AI-generated listing imagery, and MLS boards have begun applying altered-image rules to AI content. None of this is legal advice — rules vary by state and board, and your counsel should confirm what applies. The practical takeaway is calmer, and it isn’t simply “AI versus CGI”: what matters is whether the image honestly represents what’s being marketed, whether an image of an existing property has been materially altered, and which state and board rules apply. Ask whoever produces your visualization for real-estate marketing which one you’re getting, in writing.
The honest cost math
Here’s the part where a studio is supposed to get defensive, so let’s not. Per-image, the gap is enormous and real: a few cents to well under a dollar per AI generation, against published 2026 studio market rates of roughly $500–$1,200 for interiors and $800–$4,000 for exteriors (commercial work runs to $5,000). If raw image cost were the whole equation, this article would be one sentence long.
The equation buyers actually live with is cost per approved image. That includes the professional hours steering the tool (the 70% supervision figure again), the regeneration lottery when a revision can’t be surgical, and the risk carried when an image with ground truth turns out wrong. Industry surveys hint at the same gap from the other side: Deltek’s 2026 study found 70% of A&E firms using AI — and only 38% able to tie it to measurable business impact. Meanwhile the pragmatic middle shows up in the same Chaos survey: 86% of AI users report at least some time saving, and the largest gains sit exactly where the ledger above predicts — concept ideation (48%) and image enhancement (40%), not precision modelling or revision control. The market isn’t choosing a winner. It’s sorting tasks.
Using both without getting burned
The workable 2026 playbook, whether you work with us or anyone else: explore with AI while decisions are cheap — options, moods, massing, direction. The moment a design is fixed and the image has to match it, move to a pipeline where the building exists as a model, revisions are surgical, and every output stays consistent with the last. And whoever you hire for that phase, hold them to the standard the tools can’t fake — we’ve written up what to look for in an architectural CGI company, and every test in it applies double in the AI era. Nothing in our pipeline lets an image ship unless it matches the model. That sentence is the entire difference, and it’s the one thing a subscription can’t buy yet.