Ultramarine 1 Flash: a $0.13-a-task model for decks, models and memos, with the caveats

Paper Instruments says Ultramarine 1 Flash, a post-trained version of Z.ai's GLM-5.3 Flash, scores 59.3 on its internal knowledge-work suite at $0.13 a task, against 57.8 at $8.19 for Claude Fable 5 at max reasoning. The weights are not out yet, and the main benchmark is the company's own.

Ultramarine 1 Flash (Paper Instruments)The model page. Benchmarks, pricing and links to the technical report and API.paperinstruments.com →

What it is

Ultramarine 1 Flash is Paper Instruments' model for professional knowledge work: the decks, financial models and memos that people in finance, law, consulting and operations hand to clients. Per its technical report, the model applies large-scale reinforcement learning, with LoRA adapters of rank 64, to Z.ai's GLM-5.3 Flash. Weights are listed as coming soon. The model is served on the Ultramarine API at $0.15 per million input tokens, $0.03 cached and $0.50 output, the same rates as GLM-5.3 Flash.

The training data is the unusual part. Most reinforcement learning on agents uses tasks a program can check, such as code. Paper Instruments built tasks whose result is a spreadsheet, deck or memo, scored against a rubric written for each task. The curriculum spans four domains (finance, law, consulting and operations), 19 practice areas and 88 environment types, from IPO and buyout rooms to SOX controls, plant shifts and warehouse operations. Each task runs in a sandbox with realistic documents and live systems of record, such as CRM, ERP, warehouse and manufacturing systems, reached over MCP. The model trained in Paper Instruments' own open-source agent loop, Feather, and in open-source forks of the Python libraries for Word, PowerPoint and Excel files.

What the numbers show

Ultramarine beats its base on all six knowledge-work benchmarks, by 4.1 points on average, while generating 6.5% fewer output tokens. It leads on JobBench and FrontierFinance. It trails Gemini 3.8 Flash and Claude Fable 5 on DiligenceBench and APEX-Agents, and DeepSeek V4.1 Flash on AutomationBench. On coding it scores a point or two below its base on both Terminal-Bench 2.1 and DeepSWE 1.1, so the specialization cost a little coding strength.

The cost claim, read closely

PI Knowledge Work is Paper Instruments' own suite, graded by GPT-5.6 Terra against a rubric for each task. At $0.13 a task Ultramarine scores 59.3. Claude Fable 5 at max reasoning scores 57.8 at $8.19, and GPT-5.6 Sol at max scores 55.1 at $2.09, so the 60 times cheaper claim compares against the dearest settings.

The company's own chart shows closer rivals. GPT-6 Luna at max reasoning scores 58.3 at $0.11 a task, GPT-6 Sol at max scores 61.4 at $1.26, Claude Opus 5.5 at max scores 61.9 at $6.96, and Claude Fable 5.1 at max scores 65.5 at $8.64. The report says the suite correlates with the Artificial Analysis Intelligence Index at r = 0.81 across the 24 runs both measure.

Where it could be used

These are inferences from the training domains and the report's worked example.

• First drafts of finance deliverables. The report's worked task has the agent update a board pre-read in place after a forecast changes, carry the revision through valuation multiples, and separate what the board can approve now from what still needs counsel clearance. Update-the-deck and update-the-memo jobs in M&A, capital markets, private equity and restructuring fit that shape.
• Diligence and consulting work in bulk. The curriculum covers diligence memos, workpapers, cost-saving plans and integration planning. At $0.13 a task, 1,000 deliverables would cost about $130 in model fees, my arithmetic from the report's figure for its own suite.
• Legal workbench tasks. The domains include commercial contract matters, discovery review, litigation motions, patent claim charts and tax. The base model is the same one behind GLM-Legal, and no legal-specific benchmark such as Harvey's is reported, so test it on your own documents first.
• Back-office and operations work. FP&A close and forecast, procurement, SOX controls, privacy compliance, plant shifts and warehouse operations run against live ERP, manufacturing and warehouse systems over MCP. AutomationBench at 50.2 is about a point behind the top score of 51.3.
• Office file automation. The open-source paper-docx, paper-pptx and paper-xlsx libraries expose file structure as data, validate supported changes and refuse unsafe ones, which helps any agent that edits Office files.
• A drafting tier in a mixed setup. Cheap drafts from Ultramarine with review by a stronger model or a person is a natural split at these prices.

Fine print

• The weights are not out. The model page lists Hugging Face weights as coming soon, and CEO Daanish Khazi says they follow in about two weeks, per AICoder.
• One route in so far. The model is served only on Paper Instruments' own API, which is OpenAI-compatible at https://api.paperinstruments.com/v1 with the model name ultramarine-1-flash. A SourceForge comparison says it also accepts Responses-format and Anthropic Messages-style requests, so Claude Code and the Claude Agent SDK can point at it, and bills tokens against a prepaid balance. The same listing shows OpenRouter as not supported.
• Reasoning cannot be switched off. The default effort is max, per AICoder's reading of the docs, which is the setting behind the $0.13 figure, so there is no cheaper fast mode to fall back on.
• No vision training. The base model takes images and the API reportedly accepts them, but Paper Instruments did not train on images, and its file tool returns a text note that vision is not available. Treat image quality as untested.
• Context. Training used 131,072 tokens, and the PI Knowledge Work runs used 262,144.
• Cost comparisons use each developer's official API prices as of October 2026, and the frontier models run at their highest reasoning setting.
• Graders are language models. GPT-5.6 Terra grades PI Knowledge Work and GPT-5.5 grades DiligenceBench, which is also Paper Instruments' own benchmark.
• Some comparison scores come from public leaderboards or developers' own reports, and Paper Instruments ran others itself.
• The training bill is the CEO's figure and has no audit behind it. He says a team of five post-trained the model for about $300,000 all in.

What to watch

• The weights release, which would put Ultramarine beside open models such as Reflection's Beam, and who hosts the model beyond Paper Instruments.
• Independent evaluations, since the headline suite is internal.
• Legal-specific results against benchmarks such as Harvey's, given the shared base with GLM-Legal.
• Whether the 4.1-point average gain over the base holds on other people's documents.

Sources

1. Ultramarine 1 Flash (Paper Instruments) and technical report, 14 Oct: benchmarks, prices, training method, curriculum and the worked task.
2. Ultramarine API, Paper Office libraries and Feather training code: the API, the file libraries and the agent loop.
3. AICoder's launch summary and SourceForge's comparison: API details, default reasoning effort, supported request formats and the weights timeline.