Paper Instruments releases Ultramarine 1 Flash, an open model claiming Fable-class work at 60x lower cost

What was released

Paper Instruments has launched Ultramarine 1 Flash, an open model tuned for professional knowledge work in finance, law, consulting and operations. The company claims it performs on par with Claude Fable 5 while costing roughly 60 times less per task. The announcement, which reached the front page of Hacker News, is accompanied by a technical report, a GitHub repository and an API.

The model is built on GLM-5.3 Flash as its base. According to Paper Instruments, the fine-tuned model generates 6.5% fewer tokens than that base across the six benchmarks it reports, meaning it reaches higher scores while producing less output rather than more.

The numbers behind the claim

The central claim is that at $0.13 per task, Ultramarine 1 Flash outscores Claude Fable 5 at maximum reasoning ($8.19 per task) and GPT-5.6 Sol at maximum reasoning ($2.09 per task), and lands within three points of Claude Opus 5.5 (max, $6.96) and GPT-6 Sol (max, $1.26). On the company's own chart the model scores 59.3 at max reasoning, against 57.8 for Fable 5 — which makes the 60x figure arithmetically reasonable at the task level, since $8.19 divided by $0.13 is about 63.

The chart also contains results that soften the headline. Claude Fable 5.1 at max reasoning scores 65.5 at $8.64 per task, the highest result shown, and GPT-6 Astra at max reasoning reaches 64.2 at $5.64. The Fable comparison is against Fable 5, not the newer 5.1. The closest competitor on price is GPT-6 Luna, which scores 58.3 at max reasoning for $0.11 per task — slightly cheaper and slightly lower.

Against its own base, the gains are clear: GLM-5.3 Flash scores 52.0 at max reasoning for $0.10 per task, so the fine-tune adds 7.3 points for a marginal increase in cost.

What the benchmark measures

The scores come from PI Knowledge Work, an internal benchmark maintained by Paper Instruments. Its tasks ask a model to deliver finished Excel workbooks, PowerPoint decks and Word documents, and to work inside live business systems such as ERP and MES. Each task is scored against a written rubric covering substance, accuracy and presentation. The company says every model runs in the same harness and that costs are calculated using each developer's official API pricing. It also reports that the benchmark's scores correlate with the Artificial Analysis Intelligence Index at r = 0.81.

There is fine print worth noting: where a competitor's score is public, the token counts used in the comparison come from Paper Instruments' own runs that came closest to that score.

Pricing and tuning

On the Ultramarine API, pricing is $0.15 per million input tokens, $0.03 for cached input and $0.50 for output — identical to GLM-5.3 Flash. The per-task savings therefore come from scoring near frontier-level models while paying flash-tier token prices, helped along by the reduced token generation.

Paper Instruments also emphasizes how the model writes. It is described as producing plain, well-structured prose with a strong grasp of visual and information hierarchy, reflecting the formatting and presentation standards expected of analysts doing client-facing work in professional services.

Why it matters

If the numbers hold up outside the vendor's own harness, the economics of high-volume professional document work change: scores close to maximum-reasoning frontier models at cents per task make workflows that were previously too expensive to automate suddenly viable. The evaluation design is also notable — scoring finished deliverables produced inside real business software is closer to how these models are actually deployed than chat-style question answering.

The standard caveats apply in full. This is a developer-run comparison on an internal benchmark, validated only by correlation with an external index rather than independent replication. The useful counterweight is that the model is open, so teams can download it and test it against their own workloads before trusting the chart.