Paper Instruments releases Ultramarine 1 Flash, a GLM-5.3 Flash model trained with rubric-graded reinforcement learning for office work

• Model: Ultramarine 1 Flash, a LoRA (rank 64) adapter on GLM-5.3 Flash with a 131K-token context window (Paper Instruments). AICoder reports it can run up to 250 agent turns (AICoder).
• Access: an OpenAI-compatible API, live now. Weights are promised in about two weeks (AICoder).
• Price: $0.15 per million input tokens and $0.50 per million output tokens (daily.dev).
• Text only: the model has no vision, by design (Paper Instruments).
• Apps: the video says a Feather app for Apple Silicon Macs is opening from a waitlist (video).

The following details come from the company's write-up (Paper Instruments):

• Rubrics: each grading criterion should test one condition and name the facts, thresholds and context needed to check it. Criteria are graded yes or no and weighted by importance. The goal is to avoid rejecting valid alternative answers while still catching plausible errors.
• Training environments: 88 simulated office environments across 19 practice areas: finance (34), legal (15), consulting (28) and operations (11). Each practice area allows roughly 10⁷–10¹⁰ task configurations. Tasks are generated synthetically and filtered for grounding, realism and reward hacking.
• RL objective: a modified CISPO objective with importance-sampling ratios clipped at 5, mean-centered advantages, asynchronous updates and a length penalty.
• Feather harness: file reads take a content fingerprint so stale edits are rejected, and a to-do list survives context compaction. The training system prompt is about 50 tokens, against 1,800–4,400 in the production harnesses studied.
• Paper Office: forks of the docx, pptx and xlsx Python libraries that validate edits so Office files don't break silently.

• Headline result: 59.3 at $0.13 per task versus Claude Fable 5 at 57.8/$8.19 and GPT-5.6 Sol at 55.1/$2.09 (Paper Instruments). This is vendor-reported and measured on a private suite (AICoder). The video's "60 times cheaper" figure comes from these per-task costs (video).
• Other benchmarks: an average gain of 4.1 points over the base model across six benchmarks, with 6.5% fewer output tokens. Scores include JobBench 62.9 and FrontierFinance 53.3 (Paper Instruments).
• Cost and team: the $300,000 cost and five-person team are the CEO's own figures (AICoder).
• Independent corroboration: none was found in the reviewed sources.

• Coding regressed slightly: Terminal-Bench 2.1 fell from 84.3 to 83.2, and DeepSWE 1.1 from 63.4 to 61.5 (AICoder).
• Chart presentation is contested: commenters called the charts cherry-picked, saying stronger competitors appear only in cost plots (daily.dev).
• Openness is unclear: sources disagree on what is released. One coverage summary says training code will be published (AICoder). A read of the company write-up found no release commitment in it.
• Reward hacking: the company says it kept adding new checks each time it caught the model gaming the score (video).
• Name errors in commentary: the CEO's name is garbled in the Reddit text and the auto-captions.

• Rubrics work with any agent: writing one criterion per idea, each tied to named facts, can be adopted without changing models. The video gives sales-call examples such as "names the lead's main worry" (video). This advice comes from commentary and has not been tested.
• Cheap models allow best-of-N: at about $0.13 per task, running a job five times and keeping the best rubric-graded output costs less than one run on a premium model (video).
• Test on your own work first: before switching, compare it on in-house documents, and leave out work that depends on charts or images.