Ultramarine 1 Flash vs Claude: Paper Instruments lists $0.15/$0.50 API vs Claude Fable 5 at $10/$50. Vendor JobBench claims labeled; who should try it.

Paper Instruments launched Ultramarine 1 Flash on October 14, 2026 as an open model for professional knowledge work — finance, law, consulting, operations — with an API today and weights promised soon. The marketing line is blunt: Fable-class quality at a fraction of Claude Fable 5’s cost. We checked the Ultramarine page and Anthropic pricing on October 16, 2026, and we label every bench as the vendor’s.

Ultramarine 1 Flash vs Claude at a glance

Ultramarine 1 FlashClaude Fable 5Claude Opus 5.5
API input / 1M$0.15$10$4
API output / 1M$0.50$50$20
Cache read / 1M$0.03(see Anthropic page)$0.20
WeightsComing soon (vendor)ClosedClosed
Best fit (our take)Cheap knowledge-work API trialsHighest Anthropic list tierDaily agentic coding on Anthropic
API list prices from Paper Instruments and anthropic.com/pricing, Oct 16, 2026.
Ultramarine 1 Flash vs Claude API list prices per 1M tokens
List rates from Paper Instruments and Anthropic pricing, Oct 16, 2026.

Verified API prices (Oct 16, 2026)

On the Ultramarine product page, Ultramarine 1 Flash lists $0.15 input, $0.03 cached input, and $0.50 output per million tokens on the Ultramarine API (same list rates the page shows for GLM-5.3 Flash).

On Anthropic pricing, Claude Fable 5 (legacy) lists $10 input and $50 output per MTok; Claude Fable 5.1 lists the same $10 / $50. Claude Opus 5.5 lists $4 / $20 with prompt-cache read $0.20 and write $5. Claude Sonnet 5.5 lists $2 / $10.

Paper Instruments also publishes a per-task cost on its internal PI Knowledge Work harness: about $0.13 per task for Ultramarine versus $8.19 for Claude Fable 5 (max) and $6.96 for Claude Opus 5.5 (max). Those task dollars are vendor methodology — useful as a directional claim, not as our invoice.

Vendor benchmarks (Paper Instruments’ numbers)

Selected rows from the Ultramarine page comparison tables:

Benchmark (vendor)Ultramarine 1 FlashClaude Fable 5Notes
JobBench62.957.4Ultra leads on PI table
PI Knowledge Work59.357.8Internal PI harness
FrontierFinance53.349.2Ultra leads
AutomationBench50.246.2DeepSeek V4.1 Flash leads at 51.3
DiligenceBench50.652.9Gemini 3.8 Flash leads at 53.1
APEX-Agents 1.157.063.6Gemini 3.8 Flash leads at 64.3
Source: paperinstruments.com/ultramarine, Oct 16, 2026. Not TechPilot.ai evaluations.
Paper Instruments own Ultramarine vs Claude Fable 5 benchmark numbers
Vendor-reported scores from paperinstruments.com — not TechPilot.ai tests.

Reading the sheet honestly: Ultramarine leads several knowledge-work rows Paper Instruments highlights, while Claude Fable 5 and Gemini Flash still lead agent-style rows such as APEX-Agents on the same page. Do not flatten that into “Ultramarine beats Claude everywhere.”

Weights, openness, and who should try it

Ultramarine is positioned as an open model, but weights are not downloadable yet — the page says “Model weights coming soon” with Hugging Face linked for when they land. Until then you are evaluating an API, not a local checkpoint.

  • Try Ultramarine if you want a low list-price API for document-heavy professional tasks and you are fine trusting vendor benches while weights are pending.
  • Stay on Claude if your agents, evals, and compliance already sit on Anthropic — especially Opus 5.5 for coding agents. See our Claude Sonnet 4.5 page for the broader Claude stack context.
  • Compare elsewhere with our best AI model guide and GPT-6 Astra vs Sol pricing context.
  • Watch the tool page Ultramarine 1 Flash once it is published and deployed.

Our take: on sticker price, Ultramarine undercuts Claude Fable 5 by a wide margin on list $/1M tokens. Quality is still a Paper Instruments claim until independent evals or open weights land. Use it as a cheap API experiment, not as a proven Fable replacement.

Frequently asked questions

Which Claude model does Ultramarine compare against?

Paper Instruments compares Ultramarine 1 Flash mainly to Claude Fable 5 on its product page (JobBench, PI Knowledge Work, and per-task cost charts). Anthropic also lists Fable 5.1 and Opus 5.5 on the public pricing page.

How much does Ultramarine 1 Flash cost?

Paper Instruments lists Ultramarine API at $0.15 input, $0.03 cached input, and $0.50 output per million tokens — same list as GLM-5.3 Flash on that page.

How much is Claude Fable 5 on the API?

On anthropic.com/pricing (checked Oct 16, 2026), Claude Fable 5 lists $10 input and $50 output per MTok. Fable 5.1 lists the same $10 / $50. Opus 5.5 lists $4 / $20.

Are Ultramarine’s benchmark wins proven?

No. JobBench, PI Knowledge Work, FrontierFinance and the $0.13 vs $8.19 per-task figures are Paper Instruments’ own numbers. Treat them as vendor claims.

Can I download Ultramarine weights today?

Not yet. The product page says API is available now and model weights are coming soon (about two weeks in launch coverage). Do not claim local weights until they appear on Hugging Face.