Claude Opus 5.5 Built Seven Adobe Clones in Rust: What the ArtCraft Apps Actually Do

PhotoCraft had 6,691 stars on GitHub by the morning of October 7, 2026, within a week of its first commit. It is one of seven repositories published under the account storytold between September 30 and October 1, each a clean-room reimplementation of an Adobe application in pure Rust: PhotoCraft for Photoshop, VectorCraft for Illustrator, FilmCraft for Premiere Pro, LightCraft for Lightroom, PrintCraft for Acrobat Pro, EffectCraft for After Effects, and DesignCraft for InDesign. The developer behind the ArtCraft project credits Claude Opus 5.5 for the pace, and the repositories' own documentation backs that up. The question worth asking is not whether a model can help produce seven Rust codebases in a week. It is how close any of them is to the application it replaces — and the repositories answer that more honestly than the posts about them do.

Here is the number the celebration posts leave out. PhotoCraft has all 625 of Photoshop's menu items wired to live commands, a figure the project measures itself and reports as 100%. In the same document, the people who wrote it say real Photoshop parity is still well below 50%.

The seven repositories, one week in

All seven live in the open, under MIT or Apache-2.0, and all seven are written in Rust — not a wrapper around an existing C library, but a fresh engine per app. Star counts read from the GitHub API on October 7, 2026:

• PhotoCraft — image editing. 6,691 stars, 836 forks, created September 30. Opens, edits and saves layered PSD and PSB files, with layers, masks, adjustment layers, layer styles, live type, vectors and a brush engine.

• FilmCraft — video editing. 1,624 stars, created September 30. A multi-track timeline, program monitor and video scopes, with a colour and sound toolchain.

• PrintCraft — PDF reading, organising and protection. 1,117 stars, created September 30. Tracks 803 Acrobat Pro features and, by its own roadmap, still needs its own PDF renderer.

• LightCraft — photo library and raw development. 1,244 stars, created September 30. DNG, CR2, Sony ARW and RW2 decode natively; Nikon's compressed NEF and Canon CR3 are blocked by the clean-room constraint the project set itself.

• VectorCraft — vector illustration. 1,039 stars, created September 30. Pen tool, boolean operations on shapes, gradient meshes, envelope transforms and image tracing.

• EffectCraft — motion graphics and compositing. 826 stars, created October 1.

• DesignCraft — page layout and publishing. 535 stars, created October 1.

Every one of them ships installers: macOS, Windows and Linux builds, plus a WebAssembly build that runs the whole engine in a browser tab. PhotoCraft, VectorCraft and EffectCraft had cut releases as recently as October 6.

What the repositories actually say about Claude Opus 5.5

The model attribution is not a tweet. It is written into the project's own planning documents. PhotoCraft's remaining work is scheduled in what its parity estimate calls "agent-hours" — total model work — with the file explicitly titled around "how much Opus 5.5 work remains." One line reads: "this session landed ~70 menu items + major subsystems in a handful of wall-clock hours using 8 parallel agents." The same document gives the parallelism it observed as roughly 6× to 8×, and quotes the project's own estimates in Opus 5.5 hours throughout.

So the honest description of the workflow is narrower than "an AI wrote Photoshop." A fleet of parallel agents was pointed at independent subsystems — the PSD parser, the GPU compositor, the text engine, the filter set — each with its own target directory, and a human ran the integration, the tests and the release. The project is also candid that the model is a paid subscription from a different vendor, so none of the pace it bought is free or universally available.

Claude Opus 5.5 is not a new model — Anthropic shipped it on September 22, 2026, priced at $4 and $20 per million input and output tokens, with a 1M-token context window and up to 128K output tokens. On the configuration Artificial Analysis tests as Max with default fallback, it scores 58 on the AA Intelligence Index. None of the seven apps is possible without a model that can hold a large multi-crate Rust codebase in context across a long agentic session, which is exactly what that 1M-token window and the low cache-read price are for: Anthropic lists cache reads at $0.20 per million tokens, and cache reads dominate agentic coding cost.

One number says 100%, another says 68%

The two figures that matter for a reader deciding whether to download anything are both produced by the project's own tooling, and they disagree in a way that is instructive.

• Menu and command breadth — 625 of 625 Photoshop menu items dispatch a live command. 100%, measured by the project's own cargo xtask parity on 2026-10-03.

• PSD rendering fidelity — 115 of 170 corpus files render pixel-identical to Photoshop's own output. About 68%, measured by the project's composite oracle.

• PSD round-trip — 169 of 169 files in its mixed corpus re-import identically, and 307 of 309 in the psd-tools test set.

• Overall functional parity on mainstream workflows — the project's own weighted estimate is roughly 86%; including Adobe's AI and Firefly features, roughly 75%.

• "A professional could switch for daily work" — the project's own assessment as of October 5 is roughly 25–35%.

That last line is the one to carry into a purchase decision. A menu item being wired means the command exists and runs; it does not mean the output matches. The roadmap names the difference directly: Camera Raw ships without masking or calibration, Vanishing Point without its dialog, and Select Subject and Select Sky are heuristic rather than trained models. PhotoCraft's own tooling had 214 keyboard-shortcut failures at one point, and its 0.2.x cycle fixed basics like text-selection offset and an immovable crop frame that its test suite had passed. Its current status banner says it is "not yet a Photoshop replacement for daily professional work."

The subscription maths, done properly

The pitch is about money, so the money is worth running. Adobe's US list prices, read from its own plans page on October 7, 2026:

• Creative Cloud Pro (20+ apps) — $69.99 per month, annual plan billed monthly.

• Photoshop, Illustrator, Premiere Pro, After Effects or InDesign, single app — $22.99 per month each.

• Acrobat Pro — $19.99 per month. Lightroom — $11.99 per month.

Buying the seven equivalents à la carte comes to $146.93 a month. The ArtCraft Crafting Apps cost nothing: they are downloads, not subscriptions, and the source is public.

The part the comparison usually hides is the build cost, and here the project publishes it. Reaching what it calls a 95% functional state is estimated at 140 to 220 Opus 5.5 agent-hours, or 25 to 40 wall-clock hours of parallel agents at the 6–8× figure the project observed. At Anthropic's listed rates that is a real but modest bill for the model time — and it lands on you, in tokens and electricity, rather than on Adobe's balance sheet. It is also not the only thing you would pay for. Adobe's moats are cloud sync, asset libraries, the plugin ecosystem, collaboration, and now its generative models; the seven projects touch none of that, and PrintCraft's own roadmap concedes that Adobe's cloud and AI assistant have "no clean-room equivalent."

Adobe announced Anil Chakravarthy as its next president and CEO on September 3, 2026, effective December 1 — he inherits a $23.77 billion fiscal-2025 revenue base. Nothing here dents that this quarter. What has changed is the floor price of a competent creative application, which is now zero for anyone willing to run an early alpha and file bug reports.

What this says about Opus 5.5 as a tool

Strip the Adobe framing out and the more portable finding is about the shape of the work, not the quality of any single file. Eight agents, independent subsystems, a clean-room constraint that forbids copying code, shaders, ICC profiles or assets, and a human-owned test oracle that decides whether a commit is real. The oracle is the load-bearing part: the project's 68% PSD pixel-match is measured against Photoshop's own merged image, which is why it reads so much lower than the 100% menu number. A fleet of agents can generate breadth fast — 398 menu items went from unwired to wired in a single day — and breadth is exactly what a shallow metric rewards.

If you want to run the experiment yourself, the model behind it is on the same key as everything else you route. Claude Opus 5.5 is served at Anthropic's own rate — $4 and $20 per million tokens, passed through with 0% markup — so a long agentic session costs what the vendor says it costs, cache reads included, and you can put automatic failover behind it rather than betting a production path on one provider on one afternoon. That matters more for this workload than for a chat call: a multi-hour parallel build is precisely the job you do not want to lose to a single rate-limited endpoint.

Who should download this today, and who should wait

If you are a developer or a technically confident hobbyist, the repositories are worth an hour. They build with cargo, they run headless through a CLI, they expose an MCP server, and they will open your real PSD files — which makes them useful as a second opinion on a file even when they are not your daily editor. Reading the parity documents is genuinely informative about what desktop creative software costs to build.

If you are running a studio, the answer is not yet. The missing 20-odd tools, the absent AI tier, the plugin gap and the fact that most of the fixes since 0.2.0 have not been validated by real users are all disqualifying for billable work. The project says so itself, at length, in public — which is the single most credible thing about it.

Three things would move this from a promising week to a thing to plan around. Workflow acceptance tests: 30 to 50 scripted real tasks — retouch a portrait, build a social post with type and effects, prep a CMYK file for print — checked against Photoshop's output on every build. That pass rate, not menu coverage, becomes the honest headline. Second, Photoshop 2026's own PSD files for the cases the corpus currently has no reference for. Third, a released 0.2.1 so that the fixes from this week's burst are confirmed by people who are not the author. Until then, the accurate one-line summary is the one the repository already prints under its own logo: early alpha, and it says so.