Paper Instruments Launches Ultramarine 1 Flash: 60x Cheaper Than Fable 5
Paper Instruments has released Ultramarine 1 Flash, an open knowledge work model post-trained by a 5-person team for ~$300k, matching Fable 5 at 60x lower cost.
On October 14, 2026, Daanish Khazi and the Paper Instruments research team officially announced Ultramarine 1 Flash, an open foundation model specifically engineered for enterprise knowledge work. Post-trained by a five-person research team for approximately $300,000 all-in, the model claims to match the knowledge-work performance of Fable 5 while operating at 60 times lower inference cost.
Image source: @bertgodel / X / paperinstruments.com
The release specifically targets enterprise AI companies building software for demanding professional sectors such as banking and legal services. By focusing strictly on complex document reasoning and workflow execution, the creators position Ultramarine 1 Flash on the Pareto frontier for knowledge work, demonstrating that targeted post-training can achieve frontier-class utility without the immense multi-million-dollar compute budgets of generalist labs.
Lean Team Post-Training and 60x Cost Reduction Compared to Fable 5
The primary achievement behind Ultramarine 1 Flash lies in its compute efficiency and cost-to-performance profile.
Rather than relying on massive pre-training infrastructure, the team performed intensive post-training tailored exclusively to professional office workflows. The entire project was developed by a team of five researchers with an all-in expenditure of roughly $300,000.
- Pareto Frontier Performance: According to the developers' internal benchmarks, the model matches the performance of Fable 5 on knowledge-work evaluations while cutting inference expenses by a factor of 60.
- Drop-in Commercial Alternative: Paper Instruments pitches the model as a direct, drop-in replacement for proprietary models like GPT and Claude for AI vendors serving knowledge workers.
Because the 60x cost savings and benchmark parity claims originate from the development team's internal evaluations, independent third-party evaluations will be essential to verify these metrics under varied production workloads.
Non-Verifiable Agentic RL and Professional Enterprise Formatting Standards
From a technical standpoint, Ultramarine 1 Flash represents one of the largest public experiments in non-verifiable agentic reinforcement learning (RL).
Unlike code synthesis or mathematical proofs where solutions can be validated by unit tests or compilers, knowledge work lacks automated, deterministic verifiers. Paper Instruments implemented prolonged reinforcement learning pipelines designed to guide the model toward institutional enterprise standards.
- Visual Hierarchy: The model autonomously structures long-form analyses, research memos, and legal briefs with clear hierarchical headers, consistent bulleting, and logical section breaks.
- Straightforward Prose: The training curbed conversational verbosity, instilling a direct, concise tone aligned with corporate communications in financial analysis and legal review.
This specialized post-training prepares the model for multi-step agentic workflows where output structure and precision take precedence over generic conversational style.
API Waitlist Rollout, Feather App Integration, and Weights Release Timeline
Paper Instruments has outlined a phased distribution plan combining managed cloud access with future open-source releases.
Access to the model API is currently available through the Paper Instruments platform, where API keys are distributed via a rolling waitlist as hosting capacity expands. Consequently, open public access remains gated during the initial rollout.
- Feather Desktop Harness: Users can test the model immediately in its native knowledge work environment via the downloadable Feather desktop application.
- Open-Source Release in Two Weeks: While model weights are not downloadable today, the team committed to releasing the full model weights, training codebase, and a detailed technical report within two weeks.
Once the weights and technical report are released, organizations in compliance-heavy industries such as banking and legal services will be able to host and evaluate the model within their own on-premise infrastructure.
Sources
- Daanish Khazi Official X (@bertgodel): Ultramarine 1 Flash Announcement Thread
- Paper Instruments Official Blog: Ultramarine 1 Flash Overview