By A five-person lab trained a model for $300,000 that outscores Claude Fable 5 on professional knowledge work — and charges $0.13 per task where Fable 5 charges $8.19. Paper Instruments released Ultramarine 1 Flash on October 14, and while the community is right to question the benchmark framing, the price gap is harder to argue with.
Built on a Strong Open Base
Ultramarine 1 Flash is a post-training of GLM-5.3 Flash, Z.ai’s 320B-parameter mixture-of-experts model released in August 2026. GLM-5.3 Flash has 18B active parameters per token, a 1.3M-token context window, and an MIT license — so Paper Instruments started from a genuinely capable and freely available foundation. They did not invent a new architecture. They trained the right model for the right domain.
The training budget was roughly $300,000 all-in, completed by a team of five. Paper Instruments is Y Combinator-backed and was founded in 2025. This is not a skunkworks experiment — it is a deliberate bet that professional knowledge work is narrow enough to be won with domain-specific post-training instead of scale.
The Technical Unlock: Non-Verifiable RL
Standard reinforcement learning for language models works well when tasks have verifiable answers — a coding benchmark passes or fails, a math problem is right or wrong. Professional knowledge work breaks this assumption. A board presentation or a diligence memo has no single correct answer. You cannot write a unit test for “did this slide deck make the point clearly?”
Paper Instruments solved this with per-task rubric rewards: each training task was graded against its own rubric rather than a single ground truth. The underlying RL objective is CISPO (Clipped Importance Sampling Policy Optimization), a GRPO-family method documented in NVIDIA’s NeMo RL library. The training ran on a 131,072-token context window with a maximum of 250 agent execution steps per task, and included explicit penalties for overly long outputs, excess agent steps, and incorrect tool calls.
The curriculum covered 4 domains, 19 professional practice areas, and 88 distinct environment types. The result: Ultramarine improved 4.1 points on average over the GLM-5.3 Flash base across six knowledge-work benchmarks while generating 6.5% fewer tokens — which matters when token costs determine whether a workflow is economically viable.
What It Does (and Doesnt Do)
Ultramarine 1 Flash generates finished Excel workbooks, PowerPoint decks, and Word documents. It works inside live business systems — ERP and MES environments are explicitly in scope. The four training domains are finance, law, consulting, and operations: DCF models, legal diligence memos, strategy decks, process documentation.
It is not a general-purpose model. It does not replace Claude or GPT-6 for software development, creative writing, or open-domain research. Narrowness is not a bug — it is what made training at $300,000 possible. If you are building a workflow that generates 200 financial models per day, Ultramarine’s constraints will not bother you.
One practical warning from early testers: the model is token-heavy. It submits large volumes of input tokens against the endpoint, which means aggressive caching is not optional — it is required to keep costs at the advertised $0.13 per task. The cached input rate is $0.03 per million tokens, so architect your caching before running at scale.
The Benchmark: Honest About What It Shows
| Model | PI Score | Cost per Task |
|---|---|---|
| Ultramarine 1 Flash | 59.3 | $0.13 |
| Claude Fable 5 (max) | 57.8 | $8.19 |
On Paper Instruments’ PI Knowledge Work benchmark, Ultramarine 1 Flash scores 59.3 versus Claude Fable 5 at 57.8 — a narrow lead on a benchmark that Paper Instruments built and administers themselves. The Hacker News thread was appropriately skeptical: independent verification has not been completed, and the cost-comparison charts in the launch materials were criticized for cherry-picking.
That is a fair reading. Ultramarine is at the front of a cluster of models on its own benchmark, not alone on a frontier. The benchmark score should be treated as a directional signal, not a hard proof of superiority.
What is harder to dismiss is the unit economics. At $0.15 per million input tokens versus Claude Fable 5’s effective $8.19 per task, the math changes for any team running professional work at volume. Even if Ultramarine is 5% worse on quality-sensitive edge cases, the cost structure may still justify the switch for high-throughput workflows.
Open Weights Coming in Late October
The API is live now at paperinstruments.com. Model weights are scheduled for late October 2026. When they drop, the MIT-licensed GLM-5.3 Flash base makes self-hosted deployment straightforward. For finance and legal teams with data residency requirements, on-premise deployment is the only viable path regardless of cost — the weights release makes it accessible.
What This Signals
Ultramarine 1 Flash is one data point in a pattern: specialized post-training is compressing the cost of professional AI by 50-100x on narrow domains. C3 AI’s Narwhal model does it for industrial systems. LinkedIn’s EON-8B does it for enterprise content. What Paper Instruments built with $300,000 and five people was not a fluke — it is what happens when strong open-weight base models meet domain-specific RL on the right tasks.
Frontier models are not going away. But the assumption that you need a $10-per-task frontier call for every professional workflow is no longer safe. Ultramarine is an argument for auditing your AI cost structure now, before late October, when self-hosting becomes an option too.
