Paper Instruments AI Explained

Paper Instruments is a San Francisco AI company that post-trains open language models for professional knowledge work. The AI company Paper Instruments has no connection to the paper musical instruments people make as craft projects. On October 14, 2026, Paper Instruments announced Ultramarine 1 Flash, a post-trained reasoning model built on an open base model from Z.ai.

Paper Instruments went through the Y Combinator (YC) Spring 2025 batch under its original name, The LLM Data Company. Today, Paper Instruments sells per-token API access to Ultramarine 1 Flash, publishes the open-source Paper Office packages, and offers an agent harness called Feather as a downloadable app.

Paper Instruments in One Paragraph

Paper Instruments, Inc. is a three-founder startup founded in 2025 that describes itself as building "open-source frontier models for knowledge work" and says it "post-trains LLMs to accelerate knowledge work."

Paper Instruments went through the Y Combinator Spring 2025 batch as The LLM Data Company. The LLM Data Company sold post-training data, reinforcement learning (RL) environments, and evaluation tooling to AI labs and applied-AI companies. Paper Instruments subsequently transitioned from data sales to model development, releasing Ultramarine 1 Flash as its first frontier-class post-trained model on October 14, 2026.

YC lists a team of three. Y Combinator led its pre-seed round.

• Founded in San Francisco in 2025 by Daanish Khazi, Gavin Bains, and Joseph Besgen
• Participated in the Y Combinator Spring 2025 cohort under the name The LLM Data Company
• Transitioned from selling post-training data and evals to post-training open models
• Released Ultramarine 1 Flash on October 14, 2026 as its first frontier-class model

From The LLM Data Company to Paper Instruments

Paper Instruments began operations as an infrastructure supplier that sold datasets, training environments, and benchmark suites to other machine learning laboratories.

That background directly informs the technical approach Paper Instruments applies to model development today. Rather than pre-training a base model from scratch, Paper Instruments specializes in post-training open foundation models through automated reinforcement learning.

For Ultramarine 1 Flash, Paper Instruments selected GLM-5.3 Flash, an open base model created by Z.ai. According to the Paper Instruments technical report, Paper Instruments trained the model inside 88 synthetic environment types spanning 19 practice areas.

Each synthetic task in that curriculum pairs an instruction and a sandbox of files and MCP services with weighted pass/fail rubric criteria that score the model's work. This specialization in evaluation rubrics and simulated environments reflects the core tooling Paper Instruments developed during its time as The LLM Data Company.

• Original corporate identity: The LLM Data Company in the YC Spring 2025 batch
• Former product line: post-training data, RL environments, and evaluation suites for external AI labs
• Core technical strategy: post-training open base weights using automated reinforcement learning
• Base model selection: GLM-5.3 Flash from Z.ai
• Training environment design: 88 synthetic sandbox types evaluated with rubric-based rewards

Founders and Executive Leadership

Three co-founders lead Paper Instruments: Daanish Khazi, Gavin Bains, and Joseph Besgen.

Daanish Khazi serves as chief executive officer (CEO) and acts as the primary public technical voice for Paper Instruments. Khazi announced Ultramarine 1 Flash on his personal X thread, stating that a post-training team of five trained the model for approximately $300,000 all-in. His public statements and listed roles are collected in Who Is Daanish Khazi?.

Gavin Bains is a co-founder recorded on the Crunchbase profile for The LLM Data Company and the official Y Combinator directory. Paper Instruments has not published Bains's functional executive title or prior operational background.

Joseph Besgen is the third co-founder listed in official accelerator records. According to his public profile, Besgen holds a degree in electrical engineering and computer sciences with a minor in mechanical engineering from the University of California, Berkeley. The same profile says Besgen previously led backend engineering at Traba, earlier built platform infrastructure for Instagram's business messaging, and contributed to the open-source Magma project.

Other founders and researchers behind new AI labs are listed in AI People to Watch.

• Daanish Khazi: Co-founder and CEO, primary public spokesperson on X (@bertgodel)
• Gavin Bains: Co-founder, verified via Y Combinator and Crunchbase listings
• Joseph Besgen: Co-founder; according to his public profile, a UC Berkeley graduate who previously led backend engineering at Traba
• Published team size: 3 people per Y Combinator, with a post-training team of 5 cited by the CEO

Product Portfolio Across Models and Tooling

Paper Instruments makes one agentic model, an agent harness called Feather, three open-source packages for Office files, and a benchmark called DiligenceBench.

The primary release is Ultramarine 1 Flash, a post-trained checkpoint of GLM-5.3 Flash announced on October 14, 2026. Paper Instruments calls Ultramarine 1 Flash "an agentic model for professional knowledge work," and launch coverage names finance, law, consulting, and operations as the target work. Daanish Khazi claims "similar performance to Fable 5, while being 60x cheaper." That figure comes from one internal Paper Instruments suite, and the Ultramarine 1 Flash vs Claude Fable 5 comparison checks it benchmark by benchmark.

Paper Instruments also makes Feather, an agent harness for knowledge work written in Python and offered as a downloadable app. Feather gives a model a roughly 50-token system prompt and tools to read, edit, write, grep, glob, and run bash. Optional web search and Model Context Protocol (MCP) servers can be added per task, so the model can work through files and commands step by step.

Paper Instruments also develops Paper Office, a collection of three open-source Python packages launched on September 23, 2026. The packages, named paper-docx, paper-pptx, and paper-xlsx, fork standard Python document libraries so automated agents can edit existing Word, PowerPoint, and Excel files without corrupting document formatting.

The technical report also names Kos-1 Lite, Kos-1 Experimental, and DiligenceBench.

• Ultramarine 1 Flash: Agentic model for finance, law, consulting, and operations
• Feather: Knowledge-work agent harness with a roughly 50-token system prompt, shell tools, and MCP support
• Paper Office: Open-source forks of python-docx, python-pptx, and openpyxl designed for agent edits
• Kos-1 Lite and Kos-1 Experimental: Other releases named in the technical report
• DiligenceBench: Benchmark created by Paper Instruments (details not published)

Capitalization and Organization Size

Paper Instruments operates as a privately held software business headquartered in San Francisco with pre-seed venture backing.

According to Crunchbase, Paper Instruments raised a pre-seed venture funding round led by Y Combinator, with NextGen Venture Partners participating as an institutional investor. Neither Paper Instruments nor its investors have published the total capital raised in that round.

Published directory listings show a compact organization. The Paper Instruments Y Combinator directory entry lists an overall company size of three people, matching the three co-founders.

However, in his public launch statement on X, CEO Daanish Khazi stated that a team of five people completed the post-training run for Ultramarine 1 Flash. Paper Instruments has not published whether the additional contributors are full-time employees, contractors, or advisory researchers.

• Lead investor: Y Combinator (Spring 2025 cohort)
• Participating investor: NextGen Venture Partners
• Financing stage: Pre-seed round (total dollar figure not published)
• Reported headcount: 3 people on Y Combinator; 5 people cited for the post-training project team

Pricing and Access for Paper Instruments Products

Paper Instruments offers its products three ways: per-token paid API access to Ultramarine 1 Flash, the downloadable Feather app, and the open-source Paper Office packages.

The hosted API for Ultramarine 1 Flash is a paid, commercial service accessible through the Paper Instruments platform. According to the Paper Instruments pricing card, standard input costs $0.15 per million tokens, cached input costs $0.03 per million tokens, and output costs $0.50 per million tokens. These rates are broken down in the Ultramarine 1 Flash pricing guide.

According to the Paper Instruments billing documentation, customer balances are prepaid through Stripe Checkout, with optional automatic card top-ups. Paper Instruments does not advertise a free trial tier or complimentary API credit balance for new registrations. Rate limits are assigned per organization, and developers can inspect their current quota by calling GET /v1/usage, as described in the Paper Instruments rate limits documentation.

Paper Office is three open-source Python packages with no usage fee. Feather is offered as a downloadable app. Daanish Khazi's launch thread points new users to the Feather app as a way to try Ultramarine 1 Flash, with people admitted from a waitlist on a rolling basis.

Daanish Khazi announced that model weights for Ultramarine 1 Flash will be released approximately two weeks after the October 14 announcement. As of October 16, 2026, Paper Instruments has not released the model weights and has not published the software license that will govern self-hosted deployments.

• API pricing: $0.15 per million input tokens, $0.03 cached input, $0.50 output (including reasoning tokens)
• Billing model: Prepaid balance funded via Stripe Checkout; no free usage tier published
• Feather access: Downloadable app; Paper Instruments has not published a price for it
• Model weights status: Unreleased as of October 16, 2026; release promised for late October 2026

Benchmark Verification and Technical Outlook

Paper Instruments reported every Ultramarine 1 Flash benchmark score itself. The 60x cost claim comes from its non-public PI Knowledge Work suite.

Paper Instruments' technical report shows an average gain of 4.1 points over GLM-5.3 Flash across six knowledge-work benchmarks. The report also counts 6.5 percent fewer output tokens. On its non-public PI Knowledge Work suite, Paper Instruments reports 59.3 at $0.13 per task. Paper Instruments reports Claude Fable 5 at 57.8 and $8.19 per task on the same suite. However, on software engineering benchmarks such as Terminal-Bench 2.1 and DeepSWE 1.1, Ultramarine 1 Flash scored slightly lower than its base model.

No independent evaluation firm or academic laboratory has reproduced these benchmark scores as of October 16, 2026. The independent evaluation directory BenchLM lists Ultramarine 1 Flash as tracked but unranked. Furthermore, the claimed sixty-fold cost reduction represents task-level arithmetic on the internal PI Knowledge Work suite rather than a uniform per-token price difference against competing frontier models.

Teams planning an implementation should monitor whether Paper Instruments releases model weights under an unrestricted open license and whether independent platforms verify its knowledge-work evaluation scores. Prospective enterprise users evaluating Paper Instruments can inspect the Paper Instruments documentation and execute pilot tasks against its paid API endpoints to measure output quality directly.

• Vendor-reported gains: +4.1 average score across six knowledge-work suites compared to GLM-5.3 Flash
• Coding tradeoff: Slightly lower performance on Terminal-Bench 2.1 and DeepSWE 1.1 than the base model
• Cost claims: 60x cheaper figure applies to task-level runs on the private PI Knowledge Work suite, not uniform token rates
• Independent verification status: Zero third-party benchmark replications published as of October 16, 2026
• Unranked status: Tracked without an official rank on BenchLM as of October 16, 2026