Ultramarine 1 Flash: A Small Team Claims Fable-Class Work at 60x Lower Cost

A small San Francisco lab says it has built an open model that does professional knowledge work as well as Anthropic's Claude Fable 5 for a sixtieth of the price. Ultramarine 1 Flash, from San Francisco lab Paper Instruments, launched on October 14, 2026 with an API, a technical report and a promise of weights. This is a claimed-versus-verified post: the lab's own Ultramarine page is detailed, but almost all numbers are the lab's own.

The release lands in a week when cheap open models keep pressing on frontier prices; see our coverage of Jefferies on cheap Chinese open models and the AI infrastructure reset and Coinbase's fine-tuned Qwen3.5 9B beating larger models on fraud.

Quick reference

What Paper Instruments says

The product page states that Ultramarine 1 Flash "makes the decks, models, and memos that people in finance, law, consulting, and operations put in front of clients, at a fraction of the cost of frontier models". It is trained for what the company calls non-verifiable reinforcement learning: tasks where there is no unit test, only a rubric. The research post is titled "Ultramarine 1 Flash: Scaling Non-Verifiable RL across Finance, Law, Consulting, and Operations". Paper Instruments' earlier work includes Kos-1 medical models, DiligenceBench for equity research, DRACO deep-research evaluations with Perplexity, and Paper Office, Python packages for editing Word, PowerPoint and Excel files.

The benchmark table, as published

Scores below are copied from the company's page. We did not run them.

Read it carefully. Ultramarine 1 Flash tops three of six (JobBench, PI Knowledge Work, FrontierFinance). It is second on AutomationBench behind DeepSeek V4.1 Flash, third on DiligenceBench behind Gemini 3.8 Flash and Fable 5, and third on APEX-Agents 1.1, well behind Gemini 3.8 Flash and Fable 5 (57.0 against 64.3 and 63.6). "Fable-class" therefore holds on some benchmarks and not others, and Fable 5 here appears to be run in the lab's own harness.

The Paper Instruments page shows one of the sample outputs the company shows: five periods of Nucor segment sales and earnings with subtotals that reconcile, and the headline that mill earnings swung from $7.2bn in 2022 to $2.4bn in 2025 while mills' share of sales rose to 62 percent. The sources line cites Nucor filings. We did not check the slide's figures against the filings, so it is an illustration of format, not proof of accuracy.

The cost claim

The page says that on PI Knowledge Work, at $0.13 per task, Ultramarine 1 Flash outscores Claude Fable 5 (max) and GPT-5.6 Sol (max), which cost $8.19 and $2.09 per task, and comes within three points of Claude Opus 5.5 (max) and GPT-6 Sol (max) at $6.96 and $1.26 per task. Costs are at each developer's official API pricing, with every model in the same harness. Simple division gives the headline: $8.19 divided by $0.13 is about 63, consistent with "60x cheaper".

Three caveats. First, PI Knowledge Work is the lab's internal benchmark; it says scores correlate with the Artificial Analysis Intelligence Index (r = 0.81), which is a correlation among models, not an audit of the benchmark. Second, "max" reasoning settings for rivals are expensive by design, and a lower setting would narrow the gap. Third, cost per task depends on tokens used. The page says Ultramarine uses 6.5 percent fewer output tokens than its base across the six benchmarks (30.5 million against 32.7 million), which suggests the efficiency comes from cheap per-token pricing of a small model, not from magic.

For comparison with our other cost-per-task coverage, see FrontierCode 1.1 and the Artificial Analysis maintainer-merge benchmark for Opus 5.5 and the five-model comparison of Muse Wonder, GPT-6.1 Sol, Opus 5.5, GPT-6 Astra and Grok 4.7.

Claimed versus verified

We could not open the full technical report page (the research URL returned an error), so any detail beyond the product page, the benchmark list and the research index remains unread.

Which model for which job, on the published numbers

This is our reading of the company's table, not an independent ranking.

What people are asking

Can a small lab really match a frontier lab? The company's claim is narrower: on knowledge-work benchmarks it designed or selected, a post-trained small model matches or beats a much larger one. That is plausible, because narrow post-training on well-built rubrics can lift a specific skill, and Coinbase's fraud result shows the same pattern. It is not a claim about general intelligence.

Why compare against Fable at "max" settings? Because that is how the price gap gets so large. Max reasoning settings spend many more tokens. A fair customer comparison would test each model at the setting you would actually deploy.

What does "non-verifiable RL" mean? Reinforcement learning normally needs a reward you can compute, such as a passing test. Professional work has no such check, so labs use rubric graders, often another model. Scaling that is the hard part, since a model can learn to please the grader rather than the client.

Will the weights really be open? The page says "coming soon" and the product is described as open. Until the weights and license are public, treat "open" as an intention. We will update this post when they appear.

Does this change anything for me today? If you already pay for a frontier model to draft finance or legal documents, it is worth a blind test of the API against your own samples. If you do not, there is nothing to act on yet.

What it means for builders

• Watch the weights. An open model post-trained for office-style deliverables (Excel, PowerPoint, Word) would be useful for teams that cannot send client data to a frontier API. Until weights ship, the API is the only way to test it.
• Run your own tasks. The category it targets, finished deliverables judged by rubric, is exactly where benchmark gaps hide. Pick 20 real tasks, score with your own rubric and compare cost per accepted deliverable, not per token.
• Base-model dependence. If it is a post-trained GLM-5.3 Flash, license terms and behaviour inherit from that base; check the model card when weights land. Our note on GLM-5.2's MIT release covers that family's licensing history.
• Non-verifiable RL is the interesting part. Training on rubric-graded work, not unit tests, is how labs are extending RL beyond code and math. The risk is a rubric judge that can be gamed; the lab's own post on choosing a rubric judge suggests it knows this.

How to check it yourself

1. Request API access from the Ultramarine page and run a fixed set of your own tasks against it and your current model.
2. Score blind, with the same rubric, and record tokens and dollars.
3. When weights are published, check the license and the model card for training data and base model.
4. Wait for independent results, such as Artificial Analysis, before treating the cost frontier as settled.

Related reading

• Jefferies on cheap Chinese open models: claimed vs verified
• Coinbase's fine-tuned Qwen3.5 9B fraud agent
• FrontierCode 1.1 and Opus 5.5
• Muse Wonder vs GPT-6.1 Sol vs Opus 5.5 vs Astra vs Grok 4.7
• GLM-5.2 MIT open-source release
• Microsoft Token Wars: cut cost per answer
• Official: Ultramarine 1 Flash page and Paper Instruments research

Details reflect Paper Instruments' published pages as of October 14, 2026 and may change.