Ultramarine 1 Flash: Paper Instruments Ships a Deliverable Model at $0.13 a Task

Paper Instruments published Ultramarine 1 Flash on 14 October 2026, and the number the company leads with is not a benchmark score. It is a price: $0.13 to complete one task on PI Knowledge Work, the firm's internal suite of finance, legal, consulting and operations work. The model that costs $8.19 to complete the same task — Claude Fable 5 at maximum reasoning — scores 57.8 there. Ultramarine 1 Flash scores 59.3. Read that pair the way the vendor frames it and you have a model that beats a frontier system at roughly one sixty-third of the cost. Read it more carefully and you have something narrower and more interesting: a model built for a specific kind of output, priced at the rate its own base model charges, with the benchmark that says so written by the company selling it.

The base model is worth stating up front, because it explains most of the price. Ultramarine 1 Flash is a post-training of GLM-5.3 Flash, the 320B-total / 18B-active native multimodal model from Z.ai, improved through large-scale reinforcement learning rather than built from scratch. Its API rates are $0.15 per million input tokens, $0.03 cached input and $0.50 output — identical to GLM-5.3 Flash's, which is what "the same as GLM-5.3 Flash" on the vendor's own pricing block means.

What shipped, and what is still a promise

Two things are true on the day of release, and the gap between them is the story a lot of the coverage will flatten.

The Ultramarine API is live. You can send it work today and pay the rates above. The landing page links to it under a plain "Ultramarine API →" line.

The weights are not. The same page says "Model weights coming soon," and the technical report describes GLM-5.3 Flash as "a strong open base model" — which is a statement about the base, not about the derivative. Ultramarine 1 Flash is, at this moment, a closed hosted endpoint with an open-weights intention. Whether it ends up genuinely open is the single largest open question about it, and it is the one a lot of secondhand summaries have already answered by assumption rather than by checking.

What did go open on release day is the scaffolding. The Feather agent harness, which is what the model was trained inside and what it is evaluated in, is published as a repository. So are paper-docx, paper-pptx and paper-xlsx, Paper Instruments' forks of the standard Python Office libraries, which the model uses to author Excel workbooks, PowerPoint decks and Word documents. Those forks exist because, in the company's own account, agents driving python-docx and friends "hit API gaps, drop into raw OOXML, and create silent corruptions and regressions."

The six benchmarks, including the two it loses

Paper Instruments evaluated on six suites. Ultramarine 1 Flash leads four of them and does not lead two, which is a more useful summary than the average.

• JobBench — 62.9 for Ultramarine 1 Flash, ahead of Gemini 3.8 Flash at 60.6 and DeepSeek V4.1 Flash at 60.4, with Claude Fable 5 at 57.4 and GPT-5.6 Sol at 45.4.

• PI Knowledge Work, the internal suite — 59.3, ahead of Claude Fable 5 at 57.8 and GPT-5.6 Sol at 55.1.

• FrontierFinance — 53.3, ahead of GLM-5.3 Flash at 50.7 and Gemini 3.8 Flash at 50.0.

• AutomationBench — second. DeepSeek V4.1 Flash takes it at 51.3 against Ultramarine's 50.2.

• DiligenceBench — third. Gemini 3.8 Flash leads at 53.1, Claude Fable 5 is second at 52.9, Ultramarine 1 Flash is at 50.6.

• APEX-Agents 1.1 — third. Gemini 3.8 Flash at 64.3 and Claude Fable 5 at 63.6 both clear it; Ultramarine 1 Flash sits at 57.0, still well ahead of GPT-5.6 Sol at 51.5 and DeepSeek V4.1 Flash at 39.5.

The pattern in the losses is consistent and worth naming: Ultramarine 1 Flash is strongest where the deliverable is a document and weakest where the task is a live agent loop over systems. DiligenceBench is an equity-research agent evaluation; APEX-Agents is an agentic suite. AutomationBench is automation. The two benchmarks its own publisher built around office deliverables are the two it wins by the widest margin.

Why the scores look like that

Almost all professional work is not checkable by a parser. A diligence memo does not compile and a board deck has no unit tests, so there is no regex that says whether one is right. That is the problem Paper Instruments set out to attack, and its answer is a verbatim-quotable one: "we scale non-verifiable task synthesis across agentic knowledge work."

The mechanism is a curriculum. Four domains — finance, law, consulting, operations — expand into 19 practice areas and 88 distinct environment types, each of which samples a taxonomy of axes: deliverable format, source conflicts, stakeholders, transaction stage. The company puts the nominal configuration space at somewhere between 10⁷ and 10¹⁰ combinations per practice area, which is a mode-collapse defence as much as a scale claim. Each environment is a real workspace: a filesystem built from SEC EDGAR filings, analyst reports and synthesized contracts, plus live systems of record — ERP, CRM, WMS, MES, GRC — reached through MCP, HTTP and CLI.

Grading is where this gets unusual. Each task ships with a rubric, and each criterion is written to draw a boundary over acceptable answers rather than to demand one. The report's own worked example contrasts an overfit criterion that rejects two of three correct answers against a vague one that accepts three wrong ones, then shows the rewritten version matching the boundary. Deliverables are parsed and handed to a rubric grader — the company's open-source package, with GPT-5.6 Terra as the judge model. Training ran at a 131,072-token context with a 250-turn agent ceiling and compaction at 85% of the window.

That design is the source of both the result and the caveat. It is the reason a model this cheap can score near a frontier system on document work: the task distribution was manufactured to resemble document work. It is also the reason the numbers need a label attached. These are the vendor's own benchmarks, graded by the vendor's own judge, on tasks the vendor wrote. Paper Instruments reports that PI Knowledge Work correlates with the Artificial Analysis Intelligence Index at r = 0.81, which is a serious number and a partial answer — but a correlation on the vendor's side of the fence is still on the vendor's side of the fence. No third party has published an independent evaluation of Ultramarine 1 Flash yet, and the model is one day old.

The claim that does not hold on the chart

The most defensible thing about the release is also the least quoted: the model generates 6.5% fewer output tokens than its base across all six benchmarks. 30.5 million against 32.7 million, reasoning included. On a per-task bill, that is the difference between a cheaper model and a cheaper model that is also cheaper to run, and it is the sort of engineering that shows up in a bill rather than a leaderboard.

The claim that does not quite hold is the cost-quality frontier. On PI Knowledge Work, GPT-6 Luna at maximum reasoning scores 58.3 at $0.11 per task. Ultramarine 1 Flash's best mark is 59.3 at $0.13. One point and two cents is a narrow enough gap that "sits on the cost-quality frontier at the most attractive price point" is a positioning statement rather than a fact — which the vendor's own chart, to its credit, prints for anyone who reads past the headline entry. A model one point behind at two cents less is a genuine alternative for volume work, and the honest version of this release is that Ultramarine 1 Flash is at the front of a small cluster rather than alone on a frontier.

Where it fits if you want to call it

Ultramarine 1 Flash is not a general-purpose model and does not pretend to be. It is text in, and Excel workbook, PowerPoint deck or Word document out, with the ability to act inside a business system while it does. If your workload is a document, this is now one of the cheapest credible ways to produce it.

Ultramarine 1 Flash is not on our catalogue. The model it is post-trained from is: GLM-5.3 Flash is served on OrcaRouter at $0.15 input and $0.50 output per million tokens with nothing added per token, and the rate is the provider's list price passed straight through, so if Z.ai moves the number the served number moves with it the same day. That matters more than usual for a release like this one, because the entire argument for Ultramarine 1 Flash is arithmetic done against a base model's rate card. Separately, four of the models it is measured against — GPT-5.6 Sol, DeepSeek V4.1 Flash, Gemini 3.8 Flash and Kimi K3 — are reachable from the same endpoint, which means you can put the vendor's comparison to the test on your own documents rather than on theirs.

What would change the assessment

Three things, in descending order of how much they would matter.

One independent evaluation on a suite nobody at Paper Instruments wrote. The vendor has invited this by publishing the correlation with a third-party index; the next step is someone else publishing the score.

The weights. "Coming soon" is a placeholder, and until it resolves, "open model" is a description of intent that the release does not substantiate. If they land, a 18B-active model tuned for document work becomes something you can run yourself, and the $0.13 figure stops being the only way to buy this capability.

A benchmark it does not win by construction. The four suites it leads tell you it is good at the thing it was trained on. The two it trails tell you where a general agent still beats a specialist. Anyone planning a legal or financial agent pipeline should read those two rows and the $0.13 row together, and decide which one describes their workload.

Comparisons like this one are also where the failure modes show up: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.