Ultramarine 1 Flash Explained

Ultramarine 1 Flash is an AI model for professional knowledge work from Paper Instruments, built by post-training an open foundation model with reinforcement learning (RL) instead of training from scratch. It is unrelated to the Warhammer 40,000 tabletop gaming franchise that shares its name. The release targets tasks in finance, law, consulting, and operations.

Paper Instruments built the system by applying reinforcement learning on top of GLM-5.3 Flash, an open base model developed by Z.ai. CEO Daanish Khazi announced the release on X, describing it as a low-cost alternative to frontier systems such as Anthropic's Claude Fable 5. While Paper Instruments claims parity with proprietary frontier models on business tasks, the system trails on software development and relies on vendor-run evaluations that lack independent verification.

What Ultramarine 1 Flash Is (and What It Isn't)

Ultramarine 1 Flash is a specialized artificial intelligence (AI) model post-trained by Paper Instruments on top of an open foundation model to automate corporate knowledge work. Paper Instruments reports that it beats Claude Fable 5 on JobBench (62.9 versus 57.4) and FrontierFinance (53.3 versus 49.2), but scores below its own base model on both coding benchmarks it published. None of these vendor-run scores has been independently reproduced as of October 16, 2026.

The base model is GLM-5.3 Flash, an open model from Z.ai. Paper Instruments positions the release as an agentic system tailored for finance, law, consulting, and operations. Published pricing on the official pricing endpoint is $0.15 per million uncached input tokens, $0.03 per million cached input tokens, and $0.50 per million output tokens.

On social media, CEO Daanish Khazi claimed that Ultramarine 1 Flash delivers performance similar to Claude Fable 5 at one-sixtieth of the cost. That cost comparison is based on a proprietary internal benchmark suite rather than on general token pricing. As of October 16, 2026, no independent benchmark laboratory has replicated those numbers.

Weights for Ultramarine 1 Flash have not been released as of October 16, 2026. Paper Instruments stated that weights would follow in late October 2026, but the company has not published its planned software license terms.

• Developer: Paper Instruments, founded in 2025 and backed by Y Combinator
• Base model: GLM-5.3 Flash post-trained with reinforcement learning
• Official pricing: $0.15 per million uncached inputs, $0.03 cached inputs, and $0.50 outputs
• Primary focus: Finance, law, consulting, and operations
• Current access: Available through the Paper Instruments cloud API, with model weights pending release

Training Method and Non-Verifiable Reinforcement Learning

Paper Instruments trained Ultramarine 1 Flash using large-scale reinforcement learning (RL) guided by rubric rewards rather than automated unit tests or singular ground-truth answers. Standard reinforcement learning works well when code compiles or math problems have a single verified answer. In professional knowledge work, quality cannot be verified by an automated unit test.

According to the technical report, every training task is synthetic and comes with an instruction, a sandbox of files and MCP services, and its own rubric. Each output was graded against weighted pass or fail rubric criteria written for that task.

The training infrastructure used a modified CISPO objective in an asynchronous execution loop. Training parameters included Low-Rank Adaptation (LoRA) at rank 64, an AdamW optimizer, a learning rate of 5 × 10⁻⁵, a batch size of 64 × 8, and context lengths up to 131,072 tokens. Agent tasks allowed up to 250 interactive turns per synthetic scenario.

The curriculum spanned 4 core domains, 19 practice areas, and 88 synthetic environment types equipped with Model Context Protocol (MCP) integrations. Khazi reported that a team of 5 completed the post-training run with an all-in budget of approximately $300,000.

• Training paradigm: Reinforcement learning driven by multi-criteria qualitative rubrics rather than checkable code tests
• Hyperparameters: LoRA rank 64, AdamW, learning rate 5 × 10⁻⁵, and batch configuration 64 × 8
• Curriculum scope: 4 operational domains, 19 practice specialties, and 88 synthetic environment setups
• Execution window: Scenarios ran for up to 250 agent turns inside a 131,072-token training context
• Stated cost: Completed by a team of 5 with roughly $300,000 all-in, according to Paper Instruments

Knowledge-Work Strengths and Coding Weaknesses

Benchmark results reported by Paper Instruments show Ultramarine 1 Flash outperforming its base model on knowledge-work evaluations while declining on software engineering tasks. All reported scores reflect runs using maximum reasoning effort. As of October 16, 2026, third-party benchmark platform BenchLM lists the model as tracked but unranked, and no score has been independently verified.

Across six knowledge-work evaluations, Paper Instruments reports an average improvement of 4.1 points over the base GLM-5.3 Flash model. The model also generated 6.5% fewer output tokens while achieving those scores. On the internal PI Knowledge Work benchmark, Paper Instruments reports a score of 59.3, which the company states correlates with the Artificial Analysis Intelligence Index at r = 0.81 across 24 test runs.

On specialized domain benchmarks, Ultramarine 1 Flash scored 62.9 on JobBench and 53.3 on FrontierFinance, outperforming Claude Fable 5 on both evaluations. On DiligenceBench, a benchmark created by Paper Instruments, Ultramarine 1 Flash reached 50.6, trailing Claude Fable 5 at 52.9 and Gemini 3.8 Flash at 53.1. On APEX-Agents 1.1, the model scored 57.0, behind Claude Fable 5 at 63.6.

The trade-off of this specialized post-training is reduced coding capability. On Terminal-Bench 2.1, Ultramarine 1 Flash scored 83.2, dropping below the 84.3 scored by base GLM-5.3 Flash. On DeepSWE 1.1, the model scored 61.5, dropping nearly two points below the base model's 63.4 and falling well behind Claude Fable 5 at 69.9.

• JobBench: 62.9 for Ultramarine 1 Flash, leading Claude Fable 5 at 57.4 and base GLM-5.3 Flash at 56.3
• PI Knowledge Work: 59.3 at $0.13 per task, compared to 57.8 at $8.19 per task for Claude Fable 5
• FrontierFinance: 53.3 for Ultramarine 1 Flash, leading Claude Fable 5 at 49.2
• Terminal-Bench 2.1 (coding): 83.2 for Ultramarine 1 Flash, trailing base GLM-5.3 Flash at 84.3
• DeepSWE 1.1 (coding): 61.5 for Ultramarine 1 Flash, trailing base GLM-5.3 Flash at 63.4 and Claude Fable 5 at 69.9

Analysis of the Ultramarine 1 Flash Cost Claims

The claim that Ultramarine 1 Flash delivers comparable performance to Claude Fable 5 at sixty times lower cost applies specifically to Paper Instruments' internal task benchmark rather than direct token rates. Understanding that claim requires separating per-task agent run expenses from published token price sheets.

On the internal PI Knowledge Work evaluation, Paper Instruments measured total API costs for each model to complete assigned tasks using official October 2026 price lists. The company reported a per-task cost of $0.13 for Ultramarine 1 Flash against $8.19 for Claude Fable 5. Dividing $8.19 by $0.13 produces a multiple of 63, which forms the basis for the advertised sixty-fold claim.

Direct token pricing shows a different comparison ratio. On its official pricing page, Anthropic lists Claude Fable 5 at $10.00 per million input tokens and $50.00 per million output tokens, with cached input reads at $1.00. The newer Claude Fable 5.1 release maintains the same $10.00 and $50.00 rates while reducing cached input reads to $0.25 per million tokens.

By comparison, Paper Instruments sells Ultramarine 1 Flash at $0.15 per million input tokens, $0.03 for cached inputs, and $0.50 for output tokens. Comparing raw input token rates shows Ultramarine 1 Flash is 66.7 times cheaper ($10.00 divided by $0.15). Comparing output token rates shows it is 100 times cheaper ($50.00 divided by $0.50).

Because the PI Knowledge Work test suite remains an internal, proprietary benchmark, external developers cannot verify the task structures or token consumption numbers. A full assessment of production savings requires testing on public datasets, as covered in the Ultramarine 1 Flash vs Claude Fable 5 comparison.

• Origin of the sixty-fold claim: $0.13 per task on Ultramarine 1 Flash versus $8.19 on Claude Fable 5 inside Paper Instruments' private test suite
• Input pricing difference: $0.15 per million tokens for Ultramarine 1 Flash versus $10.00 for Claude Fable 5
• Output pricing difference: $0.50 per million tokens for Ultramarine 1 Flash versus $50.00 for Claude Fable 5
• Independent verification status: Unconfirmed by external parties as of October 16, 2026

API Access and the Feather Harness

Developers can access Ultramarine 1 Flash through an OpenAI-compatible API endpoint or through Paper Instruments' native Feather agent harness. Cloud API calls use the base URL api.paperinstruments.com/v1 (HTTPS) and the model identifier ultramarine-1-flash.

The endpoint supports standard Chat Completions at POST /v1/chat/completions, as well as a Responses API and an Anthropic-style Messages format described in the Paper Instruments documentation. The reasoning_effort parameter defaults to max, and reasoning cannot be turned off. Reasoning tokens stream within delta.reasoning_content, and developers must set max_completion_tokens high enough to contain both internal reasoning and the final response.

Calls support an optional x-session-id header to maintain sticky routing across multi-turn agent interactions. Account billing operates on a prepaid balance funded through Stripe Checkout, as outlined in the billing guide. Generated output tokens include internal reasoning tokens at the standard $0.50 per million rate. Paper Instruments does not list a free credit tier.

Rate limits are assigned per organization and can be checked using GET /v1/usage, as noted in the rate limits documentation. Organizations requiring custom quotas must contact Paper Instruments directly. Router provider OrcaRouter reported that Paper Instruments' per-token rates match the published rates of base GLM-5.3 Flash.

For local testing, Paper Instruments offers Feather, an agent harness written in Python and also offered as a downloadable app. Feather operates with an approximately 50-token system prompt and includes tools for reading, writing, editing, file searching, terminal commands, and task tracking. It includes file integrity checks and supports MCP servers.

Paper Instruments also released Paper Office, a collection of open-source Python packages (paper-docx, paper-pptx, and paper-xlsx) published on September 23, 2026. These tools fork common document libraries so agents can edit corporate office files without corrupting document formatting. API keys are created on the platform portal, and Paper Instruments admits users off its waitlist on a rolling basis.

• Base URL: api.paperinstruments.com/v1 (HTTPS) with model ID ultramarine-1-flash
• API compatibility: OpenAI Chat Completions, Responses API, and Anthropic-style Messages API
• Reasoning behavior: Permanently enabled, defaults to max, and billed at normal output rates
• Payment model: Prepaid account credits via Stripe Checkout with optional automatic top-ups
• Local harness: Feather, a Python agent harness also offered as a downloadable app

Unpublished Ultramarine 1 Flash Specifications

Several technical and operational parameters for Ultramarine 1 Flash remain unpublished by Paper Instruments following the initial launch. Organizations scoping a production rollout must account for these documented gaps before committing to deployment.

Model weights have not been released as of October 16, 2026. Paper Instruments stated that weights would be published approximately two weeks after the October 14 announcement, but the company has not published its intended license.

The production context window for the API is not stated in official documentation. Paper Instruments used 131,072 tokens during training runs and conducted one internal evaluation at 262,144 tokens, but the production API ceiling is not listed. The maximum allowed output token limit is also omitted from documentation.

Baseline organization rate limits are not published. While documentation displays sample numbers, Paper Instruments clarifies that those numbers are non-binding examples. New accounts must inspect their assigned limits programmatically using GET /v1/usage.

Company funding details remain partially disclosed. Paper Instruments participated in Y Combinator's Spring 2025 batch under the operating name The LLM Data Company, as recorded on Crunchbase. While Crunchbase lists a pre-seed round led by Y Combinator with participation from NextGen Venture Partners, total funding is not published. Y Combinator lists 3 team members, whereas Khazi noted a post-training team of 5.

• Weights release schedule: Expected in late October 2026, but the download repository is not live
• API context limit: Not published in developer documentation, despite a 131,072-token training window
• Output token ceiling: Not published for single API calls
• Total capital raised: Pre-seed round closed, but specific financial amounts are not published

Evaluation Criteria for Early Testing

Engineering and operations teams should evaluate Ultramarine 1 Flash for multi-step office document tasks while avoiding it for software development pipelines. Paper Instruments describes it as an agentic model for professional knowledge work rather than coding.

Teams handling knowledge-work tasks in finance, law, consulting, or operations can benefit from the low token pricing. At $0.15 per million input tokens and $0.50 per million output tokens, running exploratory document pipelines costs less than deploying Claude Fable 5. The Ultramarine 1 Flash pricing guide breaks down per-token and per-task costs.

Conversely, teams should bypass Ultramarine 1 Flash for software development, code refactoring, or script generation. On Terminal-Bench 2.1 and DeepSWE 1.1, the model scored below base GLM-5.3 Flash and behind Gemini 3.8 Flash and DeepSeek V4.1 Flash. Organizations that require self-hosted weights must also wait until Paper Instruments issues its weights and legal licensing terms.

Two events would alter this assessment. If independent evaluations reproduce Paper Instruments' knowledge-work scores, Ultramarine 1 Flash becomes a credible low-cost default for office workloads. Similarly, releasing model weights under a permissive open-source license would make local deployment feasible for security-sensitive organizations.

Teams planning an evaluation should test Ultramarine 1 Flash on a small sample of historical documents to compare real completion quality directly against current frontier models.

• Recommended workloads: Tasks across finance, law, consulting, and operations
• Excluded workloads: Code generation, repository maintenance, bug fixing, and script development
• Self-hosting constraint: Weights remain unreleased as of October 16, 2026
• Immediate testing step: Fund an API account with minimal credits and benchmark actual office outputs against existing models