Ultramarine 1 Flash added to BenchGecko · knowledge-work model at $0.15/$0.50 per M
Paper Instruments launched Ultramarine 1 Flash on 2026-10-14, a GLM-5.3 Flash post-train for professional work. BenchGecko lists it at $0.15/$0.50 per M tokens.
What landed
Ultramarine 1 Flash is a model from Paper Instruments, launched on 2026-10-14. It is GLM-5.3 Flash post-trained with large-scale reinforcement learning on non-verifiable professional tasks in finance, law, consulting and operations. The API model id is ultramarine-1-flash on an OpenAI-compatible endpoint. Paper Instruments says open weights are coming soon on Hugging Face.
Pricing and limits
From Paper Instruments' live price card:
- Input · $0.15 per M tokens
- Cached input · $0.03 per M tokens
- Output, reasoning included · $0.50 per M tokens
Paper Instruments states these match GLM-5.3 Flash. Completion tokens are capped at 32,768 per request, reasoning included. Training used a 131,072 token context and the internal eval harness used 262,144, but the API docs do not state a public context limit, so BenchGecko leaves that field empty.
Training notes
The RL curriculum spans 4 domains, 19 practice areas and 88 environment types, run in Paper Instruments' own Feather agent harness. The model was not trained with vision.
Vendor-reported benchmarks
From the 2026-10-14 technical report, all models at max reasoning. Vendor-reported, not BenchGecko tests. PI Knowledge Work is an internal Paper Instruments suite.
- JobBench · 62.9
- PI Knowledge Work · 59.3
- FrontierFinance · 53.3
- AutomationBench · 50.2
- DiligenceBench · 50.6
- APEX-Agents 1.1 · 57.0
- Terminal-Bench 2.1 · 83.2 (coding reference)
- DeepSWE 1.1 · 61.5 (coding reference)
Paper Instruments reports an average gain of 4.1 points over GLM-5.3 Flash across the six knowledge-work benchmarks with 6.5% fewer output tokens.
Cost per task
On PI Knowledge Work the report puts Ultramarine 1 Flash at $0.13 per task for a 59.3 score. Claude Fable 5 scores 57.8 at $8.19 and GPT-5.6 Sol scores 55.1 at $2.09. Claude Opus 5.5 leads by 2.6 points at $6.96 and GPT-6 Sol leads by 2.0 at $1.26.
Sources
- Ultramarine 1 Flash technical report
- Ultramarine product page
- BenchGecko model page: /model/ultramarine-1-flash
Source: BenchGecko
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.