AI-Built Photoshop Clone: PhotoCraft Is 45% Ready and Needs 2,300 to 4,100 More Agent-Hours
Updated November 3, 2026 · first published November 3, 2026
PhotoCraft, the Photoshop-style app from the ArtCraft project, rates itself about 45% ready for real work (range 40 to 50%) and still needs an estimated 2,300 to 4,100 more Opus 5.5 agent-hours to reach full parity. It is real, open source and early alpha. The viral claims that AI decompiled Adobe or fully cloned it are not what the project’s own documents say.
Published 3 November 2026 · figures from the project’s own parity assessment and README, read 3 November 2026. We have not run the software; every percentage and hour count below is the project’s estimate, not ours.
What is ArtCraft?
ArtCraft is a set of open-source apps written in Rust that mimic well-known creative and office software. The storytold GitHub organization lists PhotoCraft, VectorCraft, FilmCraft, LightCraft, PdfCraft, EffectCraft, WordCraft, GridCraft and DeckCraft among others, each described as written from scratch or as a clean-room reimplementation. GitHub shows them under Apache-2.0; an unofficial community guide, artcraftsuite.org, describes dual MIT or Apache-2.0 licensing. Coverage of the launch says the developer, Brandon Thomas, built them with Anthropic's Claude Opus 5.5 (VICE, 9 October 2026). PhotoCraft's repository showed about 76,500 stars on 3 November.
What do the headlines say, and what does the project say?
| Claim in circulation | What the project documents say |
|---|---|
| AI decompiled Adobe and rebuilt it in Rust | The README says PhotoCraft is “implemented from public specs and observed behaviour only. No proprietary code, shaders or assets.” That is the project’s own statement; we cannot verify it independently. |
| Fully working clones of Photoshop and the rest | The README calls PhotoCraft “early alpha” and says it “is not yet a Photoshop replacement for daily professional work.” |
| Every Photoshop feature is covered | Every menu item is wired to a command (631 of 631), but the README says this “measures wiring, not behaviour.” The parity page puts readiness for real work near 45%. |
| Clean room | The README says clean room. We found no description of how observation of the original was kept separate from writing the code, which is what a strict clean-room process requires. Treat the label as the project’s, not a verified legal status. |
How ready is PhotoCraft, by its own numbers?
The parity assessment targets Adobe Photoshop 2026 version 27.11.0. Its headline figures date from 10 October 2026 (PhotoCraft 0.6.0); the page was last updated on 1 November with a minor change. It separates breadth (does the feature exist) from readiness (does it work well enough for real work).
| Measure | Value | Kind |
|---|---|---|
| Feature breadth | ~76% | partly measured, weighted |
| Ready for real work | ~45% (40 to 50%) | estimated |
| Mainstream practitioner | ~43% (38 to 48%) | estimated |
| Essentials user | ~61% (55 to 67%) | estimated |
| Stage | alpha | passes the alpha gate, not yet beta |
The gap between breadth (~76%) and readiness (~45%) is the useful lesson. Counting features is cheap and flattering; making them work on real files, quickly and without crashing is the expensive part. Only 3 of 25 budgeted performance scenarios meet their budget, and the Photoshop file-format oracle passes 133 of 256 Photoshop-authored files, per the page.
What does 2,300 to 4,100 agent-hours mean?
The page calibrates its hours from the repository's history: 1,004 merged pull requests in 10.5 days, at about 2 to 4 agent-hours each. That is roughly 2,008 to 4,016 agent-hours in 10.5 days (252 hours), which implies about 8 to 16 agents working at once, our arithmetic. The remaining estimate to full parity is about as much work again as that whole window. To reach beta (about 75% ready) it estimates 1,100 to 1,800 agent-hours; for an essentials user to reach about 95%, 400 to 700.
What would that cost?
The project does not publish a cost per agent-hour, so no honest dollar figure exists. The arithmetic is cost = agent-hours x cost per agent-hour. If you assume a rate, the range is wide:
| Assumed cost per agent-hour | To beta (1,100 to 1,800 h) | To full parity (2,300 to 4,100 h) |
|---|---|---|
| $5 | $5,500 to $9,000 | $11,500 to $20,500 |
| $15 | $16,500 to $27,000 | $34,500 to $61,500 |
| $40 | $44,000 to $72,000 | $92,000 to $164,000 |
These rates are illustrative assumptions, not measurements. Measure your own: tokens per accepted change times your price per token, as in cost per successful task. Our own October Claude Code logs are in Claude Code in October.
What should an AI-spend owner take from this?
- Budget to readiness, not to a demo. The first working version is the cheap part. The project’s own estimate puts a large share of effort in depth, performance, stability and ecosystem.
- Do not trust presence metrics. 631 of 631 menu commands wired coexists with ~45% readiness. Ask for tests against real inputs.
- Re-estimate on measured throughput. Hours calibrated from pull request counts are only as good as the assumption that each PR is comparable. Track hours to a fixed acceptance bar and re-forecast.
- Check licensing and provenance before replacing licensed software with an AI-built clone. The clean-room label is not a legal opinion.
- Treat AI-built replacements as a long project. See AI coding economics and agent total cost of ownership.
What we could not verify
- The parity percentages and agent-hour estimates are the project’s own and partly estimated by it. Its page notes that Photoshop’s binary menu file was not parsed, so newer menu items may be missing from the denominator.
- We did not install or test PhotoCraft, and we did not review the other ArtCraft apps' parity pages.
- Counts of apps differ between sources (the community guide lists 13; launch coverage mentions four or seven).
Frequently asked questions
Is ArtCraft really a clone of Adobe made by AI?
ArtCraft is a set of open-source Rust apps, reportedly built with Claude Opus 5.5, that mimic Adobe and Microsoft software. The project says it is implemented from public specs and observed behaviour, not decompiled code, and calls PhotoCraft early alpha that is not yet a Photoshop replacement for daily professional work.
How complete is PhotoCraft compared with Photoshop?
By its own parity assessment (figures from 10 October 2026), about 76% of feature breadth exists but only about 45% (40 to 50%) is ready for real work. Every menu command is wired (631 of 631), which the README says measures wiring, not behaviour.
How many agent-hours are left to reach parity?
The project estimates 2,300 to 4,100 Opus 5.5 agent-hours to full parity and 1,100 to 1,800 to reach beta. The estimates are calibrated from 1,004 merged pull requests in 10.5 days at about 2 to 4 agent-hours each.
How much would it cost to build a clone like this?
The project publishes no cost per agent-hour, so no verified figure exists. Cost is agent-hours times your cost per agent-hour; at an assumed $15 per hour, full parity would be roughly $34,500 to $61,500, an illustration, not a measurement.
Related
- Cost per successful task
- AI coding economics in 2026
- Agent total cost of ownership
- Claude Code in October: tokens and list-price value
Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →