Paper Instruments' Ultramarine 1 Flash is on the API, weights still promised

On 14 October 2026, Paper Instruments published Ultramarine 1 Flash: Scaling Non-Verifiable RL. The page’s publication time is 19:00 UTC. The report presents Ultramarine 1 Flash as an agentic model for professional knowledge work: decks, financial models, memos, and similar deliverables. It says the model is a large-scale reinforcement-learning post-train of GLM-5.3 Flash, which the report calls a strong open base model. That sentence is about the base, not about Ultramarine’s own weights.

Daanish Khazi, whose X profile describes him as CEO of Paper Instruments, posted at 16:02 UTC the same day that a team of five post-trained the model for about $300,000 all-in, that the API was available, and that weights would release in two weeks. Those figures are in his post. They are not in the technical report.

The API

The quickstart tells callers to use the OpenAI SDK. The base URL is https://api.paperinstruments.com/v1. The model id is ultramarine-1-flash. The docs recommend reasoning_effort set to max, streaming on, and default sampling. The messages route can also be used with the Anthropic SDK. Any other model name returns an error, the docs say.

The billing page lists prices per million tokens: $0.15 for uncached input, $0.03 for cached input, and $0.50 for output, including reasoning. It gives a worked example: 20,000 uncached input tokens and 1,000 output tokens cost $0.0035. A prepaid balance is required. The unauthenticated endpoint GET /v1/pricing returned the same three rates on 15 October, for ultramarine-1-flash, with effective_at corresponding to 13 October 2026 at 22:00 UTC. Those are list rates, not a bill measured here.

What the company claims on its own suite

Every score below is Paper Instruments’ chart, not an independent rerun. The report says Ultramarine improves on GLM-5.3 Flash across all six knowledge-work benchmarks, by 4.1 points on average, while generating 6.5% fewer tokens. On the internal suite it labels PI Knowledge Work, it reports 59.3 at $0.13 per task. The same chart puts Claude Fable 5, at maximum reasoning, at 57.8 and $8.19 per task, and GPT-5.6 Sol, at maximum reasoning, at 55.1 and $2.09. The footnote says cost is per task at each developer’s official API pricing as of October 2026. Khazi’s post calls the Fable comparison similar performance at about 60 times lower cost. The report’s own dollars are $0.13 and $8.19.

The six chart headers, with Ultramarine’s score and the leader when it is not Ultramarine:

Suite	Ultramarine	Where it sits on the company’s chart
JobBench	62.9	First. Gemini 3.8 Flash 60.6, GLM-5.3 Flash 56.3
PI Knowledge Work (internal)	59.3	First. Claude Fable 5 57.8, GLM-5.3 Flash 52.0
FrontierFinance	53.3	First. GLM-5.3 Flash 50.7
AutomationBench	50.2	Second. DeepSeek V4.1 Flash 51.3
DiligenceBench	50.6	Third. Gemini 3.8 Flash 53.1, Claude Fable 5 52.9
APEX-Agents 1.1	57.0	Third. Gemini 3.8 Flash 64.3, Claude Fable 5 63.6

The report says comparison models run at the highest reasoning setting they support, and that some comparison scores come from public leaderboards rather than from Paper Instruments’ own reruns. The notes name which rows those are. PI Knowledge Work is the company’s internal suite of client-ready Excel, PowerPoint, and Word tasks plus live business systems, graded by GPT-5.6 Terra, in a harness it calls Feather, with a 262,144-token window.

Orca Router’s same-day note says the model leads four of the six suites and trails on two. The charts on the company page, read on 15 October, show three leads and three trails. Orca also describes an 18 billion active-parameter model. That size does not appear in the technical report or the API docs opened here.

Weights are promised, not shipped

Khazi’s post calls Ultramarine open-source and says weights release in two weeks. The homepage uses the phrase “Open-source frontier models for knowledge work.” The technical report and the API docs do not link a weights download. Open weights are promised, not shipped. This is not an open-source release.

The docs do point at a public Feather Harness tree for prompt examples, and at Paper’s document libraries for Word, PowerPoint, and Excel. A Feather app appears as a usage filter on the billing page. No waitlist for that app is stated on the pages opened here.

What to treat as settled

The model id, the OpenAI-compatible base URL, and the three token prices are written down and, for the prices, returned by the pricing endpoint. The 4.1 point gap, the 6.5% token cut, the $0.13 task cost, the $300,000 training bill, and the team of five are Paper Instruments’ own account. Until a weights repository is actually public, there is nothing to self-host.