Paper Instruments releases Ultramarine 1 Flash: RL on GLM-5.3 Flash for knowledge work, near Fable 5 on its own suite at ~$0.13 per task, API now and weights later

Why it matters

【Key points】

Paper Instruments, a small San Francisco team, released Ultramarine 1 Flash on October 14: an agentic model for professional knowledge work (finance, law, consulting, operations). Per its technical report, it applies large-scale non-verifiable-reward RL (per-task rubrics graded by a model) to Zhipu's open GLM-5.3 Flash base across 88 synthetic work-environment types, producing Excel models, decks and Word/PDF memos. The company says it beats the base by 4.1 points on average across six knowledge-work benchmarks while using 6.5% fewer tokens. On its internal PI Knowledge Work suite it scores 59.3 at about $0.13 per task, above Claude Fable 5 (57.8, $8.19) and GPT-5.6 Sol (55.1), and 2.6 points behind Claude Opus 5.5 and 2.0 behind GPT-6 Sol. The API is live at $0.15 input / $0.50 output per million tokens (same as GLM-5.3 Flash), with OpenAI- and Anthropic-compatible endpoints that plug into Claude Code. CEO Daanish Khazi says on X that a team of 5 did the post-training for about $300k all-in, and that weights will follow (the product page says "Model weights coming soon").

【Verifiable】

Technical report: https://www.paperinstruments.com/blog/ultramarine (datePublished 2026-10-14T19:00Z)

Product page (pricing, weights coming soon): https://www.paperinstruments.com/ultramarine

API docs: https://docs.paperinstruments.com/docs/ultramarine

CEO launch post: https://x.com/bertgodel/status/2110400987622265232 (Oct 14 16:02Z, ~1.08M views, 5.7k likes)

Table 8 shows coding below the base: Terminal-Bench 2.1 83.2 (GLM-5.3 Flash 84.3), DeepSWE 1.1 61.5 (base 63.4).

【Our read】

1. The story is a cheap open base plus targeted post-training: a tiny team running RL on non-verifiable professional tasks gets close to top closed models on knowledge work at a small fraction of the cost.

2. "Fable-class" rests mostly on the company's internal benchmark. On public APEX-Agents 1.1 and DiligenceBench it still trails Fable 5 and Gemini 3.8 Flash, and coding is slightly below the base. The CEO himself says not to use it for coding.

3. For now it is API-only. "Open-source" counts once the weights actually ship; check the license then.