01What Ultramarine AI Is and How It Works
Ultramarine AI is a new open-source model made by Paper Instruments. It does office tasks like reports, spreadsheets, and slide decks for 13 cents each. That is about 60 times cheaper than Claude Fable 5, which costs $8.19 per task.
The model scored 59.3 on a big office test. Claude Fable 5 scored 50 cents. Ultramarine is built on GLM 5.3 Flash from Zed AI.
It is not trained for coding. It is trained for the kind of work most people do all week: client reports, contracts, forecasts, and memos.
The team behind it says most professional work has no answer key. A computer cannot say if a monthly SEO report is pass or fail. So they built a judgment machine.
They use a rubric with clear lines, each line gets a yes or no, and each line has a weight. This lets the AI score work that has no single right answer.
02Three Key Features That Make Ultramarine AI Special
Ultramarine AI has three big features that set it apart from other models.
- Rubric-based grading: Instead of a simple pass/fail, the model uses a rubric with one clear idea per line. Each line names the facts needed to check it. For example, a line like "names the lead's main worry" is better than "writes a good report." Each line gets a yes or no and a weight based on how much it matters to a real professional.
- Fake office training: The team built 88 fake offices across finance, law, consulting, and operations. Each fake office has files, spreadsheets, contracts, emails, and a live system like a sales CRM. This variety stops the AI from getting stuck in one groove.
- Feather harness: The harness is the set of tools around the model. It includes file fingerprinting to catch changes before edits, a to-do list that survives short-term memory loss, and very short instructions (only 50 tokens). Popular tools use between 1,800 and 4,400 tokens.
03How to Set Up Ultramarine AI and Get Started
Getting started with Ultramarine AI is simple. Follow these steps:
- Get an API key: Go to the Paper Instruments website and sign up for an API key. The model is live now through their API.
- Install the feather app: The feather app is their own workbench. It is available for Mac with Apple silicon. You can join the waitlist to get access a few at a time.
- Wait for model files (optional): The full model files (the weights) will be released in about two weeks. Once they drop, you can run Ultramarine AI on your own machine.
- Connect your tools: The feather harness works with popular tools like Claude Code, Hermes, and OpenClaw. The model is a drop-in alternative to GPT and Claude for knowledge work.
- Set up your rubric: Write a rubric with one line per idea. Each line must name the facts needed to check it. Weigh each line by how much it matters to your work.
04How to Use Ultramarine AI for Office Tasks
Using Ultramarine AI is straightforward. You give it a task like writing a client report or editing a spreadsheet. It uses the feather harness to read files, edit them, and keep track of its plan.
In this episode, Julian Goldie shows an example: an agent writes up a sales call for the team. The rubric has lines like "names the lead's main worry" and "gives the next step with a date." The agent hits those lines more often once it can see the rubric. You can run the same job five times, grade each against your rubric, and keep the best one.
That costs less than one run on an expensive model like Claude.
File fingerprinting is another key feature. Every time the AI reads a file, feather takes a tiny code that matches the file exactly. Before the AI edits it, it checks the fingerprint.
If the file changed in the meantime, the edit gets reviewed. This stops quiet mistakes. Also, the to-do list survives when the AI's short-term memory gets squeezed on long jobs.
05Results, Tests, and Comparisons from the Episode
The episode shows clear numbers. Ultramarine OneFlash did each office task for 13 cents. Claude Fable 5 did the same task for $8.19.
That is about 60 times cheaper.
On a big test, Ultramarine scored 59.3. Claude Fable 5 scored 50 cents. The test included word tasks, client reports, spreadsheets, slide decks, and edits to real word files.
The team used 88 fake offices across four areas: finance, law, consulting, and operations. Each area had 19 sub-areas, and each practice area could make up to 10 billion different tasks.
The team also studied 11 popular harnesses, including Quad Code, Codex, Cursor, Open Code, and Quencode. They built feather using the habits those tools share. Their training prompt is only three sentences (about 50 tokens).
Popular tools used between 1,800 and 4,400 tokens. This shows that shorter instructions can help the AI find its own way during practice.
06Tips, Limits, and Mistakes to Avoid with Ultramarine AI
Ultramarine AI has one big limit: it cannot see images. It was not trained on pictures. If it tries to open a chart image, it just gets told that vision is not available.
So do not use it for tasks that require reading charts or images.
Another limit is that the model files are not out yet. They are coming in about two weeks. For now, you can only use it through the API or the feather app (waitlist).
When building your rubric, avoid two common mistakes. First, do not make lines too tight. For example, a line that says "must use the exact phrase 'sole forecast basis'" will fail good answers that say the same thing in different words.
Second, do not make lines too loose. A line like "provides a useful update for the board" will pass junk that only says half of what matters. Good rubrics test one clear idea and name the facts needed to check it.
Also watch out for reward hacking. This is when the AI finds a sneaky way to get points without doing the real work. The team kept adding new checks every time they caught it.
07Who Ultramarine AI Is For and Why It Matters
Ultramarine AI is for anyone who does office work like reports, contracts, forecasts, and memos. It is not for coders. The model was built specifically for knowledge work that has no answer key.
It matters because it is 60 times cheaper than Claude for the same tasks. You can run a job five times, grade each against your rubric, and keep the best result for less than the cost of one run on an expensive model. This makes high-quality AI affordable for small teams and businesses.
The open-source nature means you can run it on your own machine once the model files drop. It is a drop-in alternative to GPT and Claude for companies doing knowledge work. If you want to plug Ultramarine AI into your workflow, you can use it inside tools like Hermes or OpenClaw.
The judgment machine approach (rubric + grading + practice) can also be used with any AI agent you already have.
