The Open-Prem Inflection Point V4: Companies Building Their Own Software
In early October, developer Brandon Thomas posted to Hacker News that "software is over." He had used Claude Opus 5.5 to write Rust reimplementations of seven Adobe applications, which he describes as clean-room, and said a cancellation fee from Adobe was what pushed him to start. Within about a week he added equivalents of Word, Excel, and PowerPoint. The suite is released as ArtCraft under MIT and Apache 2.0 licenses and is free to download at getartcraft.com/apps. By mid-October its Photoshop equivalent, PhotoCraft, had 64,700 GitHub stars, 11,300 forks, 2,429 commits, and more than 1,700 tests, and it could re-save 307 of 309 test PSD files with a matching render.
ArtCraft anchors a new section of The Open-Prem Inflection Point V4, which I publish on October 26. The earlier editions argued that companies could own their AI models, their agents, and the hardware those run on. V4 applies the same reasoning to application software. As coding models improve, I expect enterprises to build more of their own tools, starting from permissively licensed open-source code like ArtCraft's, adapting it to their own workflows, and keeping the result inside the company.
How ArtCraft was built to be operated
Two details of ArtCraft's design matter more for enterprises than its parity figures. Every PhotoCraft action is one of more than 500 commands that can be called from the interface, a command line, a JSON control channel, or an MCP server, the protocol AI agents use to call tools. An agent can operate the software as easily as a person can. The apps also run natively on macOS, Windows, Linux, FreeBSD, and in a browser, with no account and no cloud dependency. A company could run them on its own machines, with an agent on local hardware driving them against files that stay in the building.
Companies were already declining purchases
Survey data shows the shift began before ArtCraft. McKinsey's State of AI 2026 survey of 1,719 respondents, published August 25, found that 32% of organizations had decided against buying at least one software product or feature because they could build it in-house with agentic coding tools. Among technology companies the figure was 41%.
Retool's 2026 Build vs. Buy report surveyed 817 builders, mostly Retool customers. Thirty-five percent had replaced at least one SaaS tool with something they built, and 78% expected to build more of their own tools in 2026. The sample consists of people who already build software, so it likely overstates the rate for companies in general.
Individual companies have described the same thing. GoDaddy told investors it is testing internally built replacements for smaller third-party SaaS tools. West Monroe's chief people officer told Business Insider that buying equivalents of the payroll-checking and workforce-insight tools its employees built with ChatGPT and Codex would have cost more than $300,000 a year.
Open-weight models can do the coding
The capability needed for this work is now available in models a company can host itself. On the DeepSWE v1.1 coding benchmark, tables published by Xiaomi show DeepSeek V4.1-Flash at 74.2 and Xiaomi's MiMo-V2.6-Pro at 71.9, against 74.0 for both Claude Opus 5 and GPT-6 Astra. These are vendor-reported figures, and Xiaomi compared against Opus 5 rather than the current Opus 5.5. Both open models are released under MIT.
On the Artificial Analysis Intelligence Index, MiMo-V2.6-Pro, a 1.02-trillion-parameter model released September 22, leads open-weight models with a score of 46. Claude Opus 5.5 leads overall at 58. Developers already trust open models with production traffic. On OpenRouter, the largest public model-routing marketplace, open-weight models passed closed models in token volume on a seven-day average on June 4 and held about 60% of tokens by mid-June. That traffic runs on hosted APIs, and closed labs still collect most of the money: on Vercel's AI Gateway, open-weight models carried 29% of June tokens on under 4% of spend.
For a company building its own software, this means the coding agents can run on self-hosted weights, and the source code, specifications, and internal data involved in the build stay inside its network.
The license of the starting code decides what a company can keep
A company that starts from open-source code and modifies it can keep its version private only if the original license allows that. MIT and Apache 2.0, which ArtCraft uses, permit closed derivatives. The AGPL, used by the OnlyOffice suite, requires releasing source code for modified versions offered to users over a network. A team starting from AGPL code should plan to publish its changes or choose a different base.
Copyright gives a second reason to keep AI-written code unpublished. Plagiarism Today argues that code generated by AI without a human author is likely not protected by copyright at all. If that view holds, a company cannot rely on copyright to protect internal software its agents wrote, and its practical protection is to keep the code private and treat it as a trade secret.
The paper includes a guide to which software to build, adopt, or keep buying. Internal admin tools, dashboards, and workflow apps are the simplest to build with coding agents, and the surveys show them moving first. Workflows that reflect how a particular company operates are candidates for building on a permissive open-source base and keeping closed. For commodity desktop software such as office suites, image editors, and PDF tools, LibreOffice and OnlyOffice are the stable choices today, and AI-built suites like ArtCraft are worth piloting. Systems of record, including ERP, payroll, core banking, and electronic health records, should stay with vendors, because their compliance, audit, and integration requirements outweigh the license savings.
The same approach applies to AI agents. OpenClaw's public skill registry, ClawHub, grew to about 52,700 skills by June with no mandatory security review, and a March audit by RankClaw found 1,103 malicious skills across 14,706. The paper recommends treating public skills as reference material and having coding agents rebuild the ones a company needs from reviewed source, then keeping them in a private registry.
Cost, hardware, and compliance
The rest of V4 updates the model, hardware, cost, and compliance picture since April. Several of the strongest open models now ship under custom licenses. Kimi K3, at 2.8 trillion parameters, requires branding for products above 100 million monthly users or $20 million in monthly revenue. Meta's Muse Spark Open, an 816-billion-parameter model released October 15, requires a separate agreement for companies with more than $20 million in annual revenue that offer it as a service. GLM-5.3, Qwen3.8-Max, Step 5, and MiniMax M3 attach conditions of their own. Legal review of each model license now belongs in any self-hosted deployment, as it has long belonged in the use of open-source code.
Hardware got more expensive. H100 rental was up about 20% for the year by May, and standard RTX 5090 cards sold for $4,930 to $5,090 in early September against a $1,999 list price. Apple's M5 Ultra Mac Studio, with up to 512GB of unified memory and a starting price of $5,499, is the main new option for small deployments, though independent benchmarks were limited when I finished the paper.
Hosted open-weight APIs are cheap enough to change the cost comparison. MiMo-V2.6-Pro costs $0.435 per million input tokens and $0.87 per million output tokens, against $10 and $50 for GPT-6 Astra. In the paper's worked example, a $40,000 hardware purchase handling 20 million tokens a day at a 3:1 input-to-output ratio pays back in about 3.3 months against GPT-6 Astra and in about 10 years against MiMo-V2.6-Pro through a hosted API, before counting power, staff time, and maintenance. Against open-weight APIs, the case for self-hosting rests on data residency, regulation, latency, fine-tuning, and protection of internal code and data.
On regulation, the EU's Digital Omnibus moved the AI Act's standalone high-risk obligations from August 2, 2026 to December 2, 2027, while the Article 50 transparency rules took effect on schedule in August. IBM's 2026 breach report put the global average cost of a breach at $4.99 million, up from $4.44 million the year before.
Where these projects slow down
Every internal tool needs an owner. One analysis of the Retool data estimates 12.5 to 25 hours of weekly maintenance for a team that replaces five SaaS tools. Governance is a larger concern. In Retool's survey, 60% of builders had built something outside IT oversight in the past year, and 49% had not yet shipped working software with AI. IBM found shadow AI involved in 43% of breaches, up from 20%, and software that employees generate with coding agents and run without review adds to the same exposure. The paper recommends giving employees approved coding tools that run on infrastructure the company controls, so the sanctioned route is the easier one to take.
Legal questions remain open. Plagiarism Today judges the ArtCraft apps to be on reasonable footing against copyright claims, since copyright protects code rather than functionality. It also notes that Adobe's and Microsoft's terms of service prohibit studying their software's inputs and outputs in order to recreate it, that Adobe holds thousands of patents, and that close interface copies could draw trademark claims. It does not expect Adobe to act quickly. Alternatives such as GIMP, Krita, and Affinity have been available for years without displacing Adobe.
The largest cost is reliability. Plagiarism Today cites the ninety-ninety rule, under which the last 10% of the code takes the other 90% of the development time. Thomas has said he expects other developers to follow his approach, and his own estimate for when ArtCraft reaches full parity has moved from within a month to within months.
