Who Is Daanish Khazi, the Founder of Paper Instruments?

Daanish Khazi is the co-founder and chief executive officer (CEO) of Paper Instruments, an artificial intelligence (AI) research startup based in San Francisco, California.

At Layer3Labs, we follow AI labs like Paper Instruments because the models they release end up inside the business workflows we build for clients.

Khazi announced Ultramarine 1 Flash in October 2026, presenting the release as a cost-effective alternative to closed frontier models for legal, financial, and operational tasks.

Daanish Khazi at a Glance

Khazi is listed as Founder and CEO on Paper Instruments' Y Combinator profile, which places the company in the Spring 2025 batch. He runs it with co-founders Gavin Bains and Joseph Besgen, and he posts on X as @bertgodel, a verified account.

He announced Ultramarine 1 Flash on October 14, 2026, a model Paper Instruments describes as "an agentic model for professional knowledge work". His launch thread rested on two claims: a "team of 5" post-trained it for "~$300k all-in", and it offers "similar performance to Fable 5, while being 60x cheaper". Public records on Khazi's background are limited. Paper Instruments and commercial databases do not publish his age, birth date, educational degrees, prior employers, or hometown as of October 16, 2026.

Khazi's thread also told readers how to get access: create an application programming interface (API) key on the Paper Instruments platform, or "download the Feather app to try it in our native knowledge work harness". New users come off the waitlist on a rolling basis, so access is not instant as of October 16, 2026. Khazi also appears in the AI people to watch list.

Paper Instruments Founding and Corporate History

Paper Instruments was founded in 2025 in San Francisco and initially operated under the name The LLM Data Company during the Y Combinator Spring 2025 batch, designated X25. During that period, the team built post-training datasets, reinforcement learning environments, and evaluation tools for external artificial intelligence laboratories.

Paper Instruments now describes itself as "Open-source frontier models for knowledge work". According to Crunchbase, Paper Instruments raised one pre-seed round led by Y Combinator, with NextGen Venture Partners also listed as an investor. Neither Paper Instruments nor its investors publish the dollar amount raised in that pre-seed round. Y Combinator lists a team size of 3, while Khazi's launch thread credits a team of 5 with the Ultramarine 1 Flash post-training run.

Under Khazi's leadership, the company rebranded to Paper Instruments. The team moved from selling post-training data, reinforcement learning environments, and evaluation tools to other AI labs to post-training open models itself. The Paper Instruments explainer covers the company's other releases in more detail.

What Daanish Khazi Claimed About Ultramarine 1 Flash

Each of Khazi's launch claims rests on a different kind of evidence, and as of October 16, 2026 outsiders can check only the pricing. The "~$300k all-in" figure in his X thread covers post-training only: Ultramarine 1 Flash is GLM-5.3 Flash post-trained, so the cost of building that base model sits outside it. A second post called it "a drop-in alternative to GPT & Claude for AI companies serving knowledge workers".

The performance half of Khazi's claim rests on Paper Instruments' own benchmarks: the company reports 59.3 for Ultramarine 1 Flash against 57.8 for Claude Fable 5 on its internal PI Knowledge Work suite, which is not public. On APEX-Agents 1.1 in the same report, Claude Fable 5 leads, 63.6 to 57.0. The 60x figure comes from that internal suite, which recorded $0.13 per task for Ultramarine 1 Flash compared to $8.19 for Claude Fable 5. It is not a 60x difference in raw per-token pricing.

Official Paper Instruments pricing documentation lists rates of $0.15 per million uncached input tokens, $0.03 per million cached input tokens, and $0.50 per million output tokens. Output pricing includes reasoning tokens generated by the model, as outlined in the Paper Instruments billing documentation. Published list prices from Anthropic for Claude Fable 5 are $10.00 per million input tokens, $1.00 per million cached input tokens, and $50.00 per million output tokens. The Ultramarine 1 Flash vs Claude Fable 5 comparison works through the per-token and per-task math side by side.

Post-Training Methodology and Knowledge Work Tooling

The technical strategy behind Paper Instruments relies on post-training open-weight base models using non-verifiable reinforcement learning rather than training a model from scratch. The Paper Instruments technical report confirms that Ultramarine 1 Flash builds on top of the open GLM-5.3 Flash base model. Paper Instruments applied large-scale reinforcement learning across tasks that lack a single checkable answer.

The training curriculum spanned 4 domains, 19 practice areas, and 88 synthetic environment types. Each task evaluated agent actions using weighted pass-and-fail rubric criteria. Training employed an asynchronous loop with a modified CISPO objective, Low-Rank Adaptation (LoRA) rank 64, AdamW optimization, and an active context of 131,072 tokens. The technical report recorded an average gain of 4.1 points across six knowledge work benchmarks over base GLM-5.3 Flash, with 6.5 percent fewer output tokens.

Paper Instruments also reports that coding scores dipped slightly below the base model: 83.2 against 84.3 on Terminal-Bench 2.1, and 61.5 against 63.4 on DeepSWE 1.1. The Ultramarine 1 Flash guide covers the full benchmark table and API details. Paper Instruments also built Feather, a Python agent harness that ships as a downloadable app, so people can try Ultramarine 1 Flash on files and tools without writing code against the API. Paper Instruments released Paper Office on September 23, 2026. Paper Office provides three open-source Python libraries, paper-docx, paper-pptx, and paper-xlsx, designed to let agents edit documents, slide decks, and spreadsheets without corrupting their structure.

Paper Instruments Co-Founders and Leadership Team

Paper Instruments was founded by Daanish Khazi alongside co-founders Gavin Bains and Joseph Besgen. All three founders are listed on the official Paper Instruments Y Combinator profile.

Gavin Bains is listed only as Founder; his role title and background are not published on Y Combinator or Crunchbase as of October 16, 2026. More details are available for Joseph Besgen.

According to his public profile, Besgen holds a degree in electrical engineering and computer sciences with a mechanical engineering minor from the University of California, Berkeley. That profile notes that Besgen previously led backend engineering at Traba, engineered platform infrastructure for Instagram's business messaging systems, and contributed to the open-source Magma project.

What to Watch Next from Daanish Khazi

As of October 16, 2026, nobody outside Paper Instruments has reproduced its Ultramarine 1 Flash benchmark scores, and the model weights are not yet public. Khazi stated on X that weights would release approximately two weeks after the October 14 launch. Paper Instruments has not published the software license terms that will govern those weights.

As of October 16, 2026, the benchmark tracker BenchLM lists Ultramarine 1 Flash as tracked but unranked. Before sending queries, check organization limits in the Paper Instruments rate limits documentation.

Before acting on Khazi's 60x figure, run a sample of your own knowledge-work tasks through the Ultramarine 1 Flash API and compare cost per task with the model you are weighing it against.