Ultramarine 1 Flash vs Claude Fable 5
A benchmark and pricing comparison across document analysis, agentic workflows, coding tasks, and operating costs.
On benchmark figures reported by Paper Instruments, Ultramarine 1 Flash edges Claude Fable 5 on specific knowledge-work evaluations at an execution cost of $0.13 compared to $8.19 per task. Claude Fable 5 leads on APEX-Agents 1.1, DiligenceBench and both coding benchmarks, though its Terminal-Bench 2.1 margin is 1.1 points. Neither model wins across every workload, and independent researchers have not yet reproduced the vendor-reported numbers as of October 16, 2026.
Paper Instruments announced Ultramarine 1 Flash on October 14, 2026, positioning the system as an agentic model for professional knowledge work in finance, law, consulting, and operations. Claude Fable 5 is Anthropic's most capable public Claude model with closed weights. Deciding between them requires separating token list prices from per-task execution costs and evaluating how specialized reinforcement learning compares against general frontier model capabilities.
Ultramarine 1 Flash vs. Claude Fable 5: Side-by-Side
| Dimension | Ultramarine 1 Flash | Claude Fable 5 |
|---|---|---|
| Developer organization | Developed by Paper Instruments, Inc., founded in 2025. | Developed by Anthropic. |
| Base model origin | Post-trained from GLM-5.3 Flash, an open weights foundation model developed by Z.ai. | Anthropic's own model. |
| Model weights availability | Weights are not released as of October 16, 2026, with public availability promised in late October 2026. | Weights remain closed. |
| Uncached input price per 1M tokens | Billed at $0.15 per 1 million input tokens through the official application programming interface. | Billed at $10.00 per 1 million input tokens on the Anthropic standard rate schedule. |
| Output price per 1M tokens | Billed at $0.50 per 1 million output tokens, which includes internal reasoning tokens. | Billed at $50.00 per 1 million generated output tokens. |
| Cached input price per 1M tokens | Billed at $0.03 per 1 million cached input tokens for prompt context reuse. | Billed at $1.00 per 1 million cached input tokens on Fable 5, reduced to $0.25 on Fable 5.1. |
| Per-task cost on PI Knowledge Work | Averages $0.13 per task across Paper Instruments' internal knowledge work evaluation suite. | Paper Instruments reports $8.19 per task on the same internal suite. |
| JobBench score | Scores 62.9 points under maximum reasoning effort in Paper Instruments evaluation runs. | Scores 57.4 points under comparative evaluation runs reported by Paper Instruments. |
| FrontierFinance score | Paper Instruments reports 53.3. | Paper Instruments reports 49.2. |
| AutomationBench score | Paper Instruments reports 50.2. | Paper Instruments reports 46.2. |
| DiligenceBench score | Paper Instruments reports 50.6 on DiligenceBench, a benchmark it built itself. | Paper Instruments reports 52.9, a 2.3-point lead. |
| APEX-Agents 1.1 score | Paper Instruments reports 57.0. | Paper Instruments reports 63.6, a 6.6-point lead. |
| Terminal-Bench 2.1 score | Paper Instruments reports 83.2, below the 84.3 of its GLM-5.3 Flash base. | Paper Instruments reports 84.3, a 1.1-point lead. |
| DeepSWE 1.1 score | Paper Instruments reports 61.5, below the 63.4 of its GLM-5.3 Flash base. | Paper Instruments reports 69.9, an 8.4-point lead. |
| Context window limit | Not published in the official documentation, though training used 131,072 tokens. | Not published here. |
| Reasoning control behavior | Reasoning effort defaults to maximum and cannot be disabled in application programming interface calls. | Not published here. |
| Platform access methods | Available through the Paper Instruments API and the downloadable Feather app, which admits users from a rolling waitlist. | Available through the Anthropic API. |
Are you one of these vendors? Update your listing
Assessment of the Claude Fable 5 Performance Claim
Paper Instruments reports that Ultramarine 1 Flash scores higher than Claude Fable 5 on four of the six knowledge-work benchmarks it compares (JobBench, PI Knowledge Work, FrontierFinance and AutomationBench) while reducing per-task execution costs from $8.19 to $0.13 on its internal evaluation suite. On those numbers the claim holds for four benchmarks, but Claude Fable 5 still leads on APEX-Agents 1.1, DiligenceBench and both coding benchmarks, and no outside group has reproduced any of the scores as of October 16, 2026.
In the Paper Instruments technical report, Ultramarine 1 Flash achieved 62.9 points on JobBench compared to 57.4 points for Claude Fable 5. On FrontierFinance, Ultramarine 1 Flash reached 53.3 points against 49.2 points for Claude Fable 5. On AutomationBench, Ultramarine 1 Flash registered 50.2 points versus 46.2 points for Claude Fable 5. On the private PI Knowledge Work benchmark, Ultramarine 1 Flash scored 59.3 points at $0.13 per task, while Claude Fable 5 scored 57.8 points at $8.19 per task.
The comparison shifts on agentic and coding benchmarks. Claude Fable 5 scored 63.6 points on APEX-Agents 1.1 compared to 57.0 points for Ultramarine 1 Flash. On DiligenceBench, a benchmark created by Paper Instruments itself, Claude Fable 5 scored 52.9 points while Ultramarine 1 Flash scored 50.6 points. Claude Fable 5 also led on Terminal-Bench 2.1 with 84.3 points against 83.2 points, and led on DeepSWE 1.1 with 69.9 points against 61.5 points.
Knowledge Work Benchmark Advantages
Paper Instruments reports Ultramarine 1 Flash ahead of Claude Fable 5 by 5.5 points on JobBench, 4.1 on FrontierFinance, 4.0 on AutomationBench and 1.5 on the internal PI Knowledge Work suite. Paper Instruments post-trained the system from GLM-5.3 Flash, an open model from Z.ai, using large-scale reinforcement learning (RL) on non-verifiable tasks. Paper Instruments defines non-verifiable tasks as tasks with no single checkable answer, graded with per-task rubric rewards.
The training curriculum spanned 4 domains, 19 practice areas, and 88 synthetic environment types. Each synthetic task paired a prompt instruction with a sandboxed environment containing reference files, Model Context Protocol (MCP) integrations, and weighted pass or fail rubric criteria. Paper Instruments used an asynchronous training loop with a modified CISPO objective, a low-rank adaptation (LoRA) rank of 64, an AdamW optimizer, and agent interactions running up to 250 turns within a 131,072-token context window.
This specialization produced an average gain of 4.1 points over the GLM-5.3 Flash base model across six knowledge-work benchmarks, while reducing output token volume by 6.5 percent. To assist document workflows, Paper Instruments also created Paper Office, a collection of open-source Python packages including paper-docx, paper-pptx, and paper-xlsx released on September 23, 2026. These libraries allow automated systems to modify Word, PowerPoint, and Excel files without breaking them.
Agentic and Software Engineering Workloads
Paper Instruments' own report shows Claude Fable 5 ahead of Ultramarine 1 Flash on APEX-Agents 1.1, DiligenceBench, Terminal-Bench 2.1 and DeepSWE 1.1. On APEX-Agents 1.1, Claude Fable 5 leads by 6.6 points.
On DiligenceBench, Claude Fable 5 scored 52.9 points compared to 50.6 points for Ultramarine 1 Flash.
On coding, the size of the gap depends on the benchmark. On Terminal-Bench 2.1, Claude Fable 5 scored 84.3 points against 83.2 points for Ultramarine 1 Flash. On DeepSWE 1.1, Claude Fable 5 achieved 69.9 points while Ultramarine 1 Flash managed 61.5 points. The base GLM-5.3 Flash model scored 63.4 on DeepSWE 1.1 and 84.3 on Terminal-Bench 2.1, so Ultramarine 1 Flash scores slightly below its own base model on both coding benchmarks.
Token Pricing Versus Task Pricing Dynamics
The claim that Ultramarine 1 Flash is 60 times cheaper than Claude Fable 5 reflects full task execution costs on an internal evaluation suite rather than an identical per-token pricing multiple across every request. Pricing structures differ across input, cached input, and output token categories, producing variable cost ratios depending on workflow architecture.
On raw token pricing published on the Paper Instruments pricing endpoint and the Anthropic pricing page, Ultramarine 1 Flash charges $0.15 per 1 million uncached input tokens, $0.03 per 1 million cached input tokens, and $0.50 per 1 million output tokens. Claude Fable 5 charges $10.00 per 1 million uncached input tokens, $1.00 per 1 million cached input tokens ($0.25 on Claude Fable 5.1), and $50.00 per 1 million output tokens. Uncached input tokens are 66.7 times cheaper on Ultramarine 1 Flash, while output tokens are 100 times cheaper.
The per-task ratio of about 63 times differs from the per-token ratios because cached input is only about 33 times cheaper ($0.03 versus $1.00) and each model can use a different number of tokens on the same task. Across the internal PI Knowledge Work evaluation, Paper Instruments recorded an average execution cost of $0.13 per task for Ultramarine 1 Flash compared to $8.19 for Claude Fable 5. On that same benchmark, GPT-6 Sol averaged $1.26 per task, GPT-5.6 Sol averaged $2.09 per task, and Claude Opus 5.5 averaged $6.96 per task.
Verification Gaps and Unproven Dimensions
Every published benchmark comparing Ultramarine 1 Flash directly against Claude Fable 5 originates from Paper Instruments rather than independent academic or industry testing groups. CEO Daanish Khazi says on X that a team of five post-trained the model for about $300,000 all-in, and no outside group has reproduced the results as of October 16, 2026.
The PI Knowledge Work benchmark remains an internal evaluation harness that is not publicly downloadable. Paper Instruments reported that performance on the suite correlates with the Artificial Analysis Intelligence Index at r = 0.81 across 24 test runs, but third-party developers cannot inspect the underlying prompt set. Furthermore, independent tracking platform BenchLM lists Ultramarine 1 Flash as tracked but unranked as of October 16, 2026.
Key operational specifications also remain unverified or unpublished in official resources. Public model weights have not been released as of October 16, 2026, though chief executive officer Daanish Khazi stated in an announcement thread on X that weights will release in approximately two weeks. The underlying model license remains unpublished. The Paper Instruments documentation does not publish an official application programming interface context window limit, nor does it list standard organization rate limits or maximum output token thresholds. The Paper Instruments rate-limits page says each organization can read its own limits with GET /v1/usage and must contact Paper Instruments for higher ones.
Implementation and Side-by-Side Testing Strategy
Engineering teams can test Ultramarine 1 Flash alongside Claude Fable 5 by sending the same sample of real tasks to both, using an API key from the Paper Instruments platform and the OpenAI-compatible endpoint at api.paperinstruments.com/v1. The application programming interface accepts standard Chat Completions requests at POST /v1/chat/completions, as well as a Responses interface and an Anthropic-style Messages endpoint using model identifier ultramarine-1-flash.
Developers integrating Ultramarine 1 Flash must account for model reasoning mechanics. The reasoning_effort parameter defaults to max and cannot be disabled in requests. Reasoning tokens stream through delta.reasoning_content and are billed as output tokens according to the Paper Instruments billing documentation. The max_completion_tokens parameter covers reasoning plus answer tokens. If requests terminate with a finish_reason of length, developers must increase this limit. Developers can also supply an optional x-session-id header to enforce sticky server routing across multi-turn dialogs.
For interactive evaluation, Paper Instruments provides Feather, a Python agent harness available as a downloadable app with admission granted from a rolling waitlist. Feather uses a compact system prompt of approximately 50 tokens and tools such as file reading, writing, regex grep, glob matching, shell execution, and task tracking, alongside Model Context Protocol (MCP) server support. Teams should run a representative sample of internal document extraction, quantitative analysis, and reporting tasks through both endpoints before committing to wide-scale workload migration.
The Verdict
Trial Ultramarine 1 Flash if your team runs high volumes of document and financial analysis work, where Claude Fable 5's $8.19 per-task cost adds up. At $0.13 per task, Ultramarine 1 Flash costs about 63 times less on Paper Instruments' internal suite, and Paper Instruments reports it ahead of Claude Fable 5 on JobBench and FrontierFinance.
Choose Claude Fable 5 if your work depends on software engineering or on agentic tasks of the kind APEX-Agents 1.1 measures, or if your team needs a proven vendor. Claude Fable 5 holds an 8.4-point lead on DeepSWE 1.1 and a 6.6-point lead on APEX-Agents 1.1. Engineering teams building autonomous developer agents or high-stakes autonomous workflows should not treat Ultramarine 1 Flash as a direct replacement.
The most practical production architecture is a model routing tier. Teams can direct cost-sensitive document and analysis workloads to Ultramarine 1 Flash to capture lower execution costs, while sending coding work and agentic tasks of the kind APEX-Agents 1.1 measures to Claude Fable 5. If independent evaluation suites fail to reproduce Paper Instruments' reported knowledge-work scores, teams should keep all critical tasks on Claude Fable 5.
Researched from primary vendor documentation and public regulator sources. Pricing and availability are accurate as of Oct 16, 2026 and can change — confirm current terms with each vendor before you buy.
Frequently Asked Questions
- Yes, but only on per-task execution costs within Paper Instruments' internal knowledge work benchmark suite. Ultramarine 1 Flash averaged $0.13 per task compared to $8.19 for Claude Fable 5. On basic per-token pricing, uncached input tokens are 66.7 times cheaper ($0.15 versus $10.00 per million) and output tokens are 100 times cheaper ($0.50 versus $50.00 per million).
- No. Paper Instruments' own report shows Claude Fable 5 ahead of Ultramarine 1 Flash on both coding benchmarks, scoring 84.3 versus 83.2 on Terminal-Bench 2.1 and 69.9 versus 61.5 on DeepSWE 1.1.
- Yes. Paper Instruments provides an Anthropic-style Messages endpoint alongside its OpenAI-compatible Chat Completions interface, both documented in the Ultramarine docs. Developers can point existing application code to the new endpoint by updating the base URL, providing a Paper Instruments application programming interface key, and using model identifier
ultramarine-1-flash. - No. As of October 16, 2026, all comparative benchmark figures originate directly from Paper Instruments' technical report. The PI Knowledge Work evaluation suite is internal and not publicly available, though Paper Instruments reports an r = 0.81 correlation with the Artificial Analysis Intelligence Index. Third-party benchmark platform BenchLM lists the model as tracked but unranked.
- Paper Instruments reports that Ultramarine 1 Flash scored 53.3 points on FrontierFinance compared to 49.2 points for Claude Fable 5, while operating at a fraction of the cost per task. However, Claude Fable 5 scored higher on DiligenceBench (52.9 versus 50.6).
Plan Your Model Routing Architecture
At Layer3Labs, we help teams benchmark emerging models against their production workloads, calculate cost thresholds, and build automated routing layers.
Book a Consultation