Paper Instruments is releasing Ultramarine 1 Flash, an open model built for knowledge work and aimed at the document-heavy workloads of banks, law firms, and consultancies, where it can deliver Excel models, slide decks, or Word memos for cents per task . On the company's internal benchmark for client-ready deliverables, Ultramarine 1 Flash scored 59.3 at $0.13 per task, compared with 57.8 at $8.19 per task for Claude Fable 5 , which is the comparison point behind the roughly 60x cost claim in the launch tweet.
The model is a post-training of GLM-5.3 Flash, completed by a team of five for approximately $300,000 all-in. Because a diligence memo or a board deck has no single correct answer, each training task was graded against its own rubric . That rubric-graded approach is the core bet here, pushing a mid-size open base model toward structured, judgment-heavy output rather than general chat performance.
Paper Instruments was founded in 2025 by Gavin Bains, Joseph Besgen, and Daanish Khazi , is based in San Francisco, and went through Y Combinator's Spring 2025 batch. Khazi, the CEO, is announcing the launch publicly. The model is available now through the Paper Instruments API at $0.15 per million input tokens and $0.50 per million output tokens , with open weights to follow in two weeks, giving buyers at regulated firms a path from hosted API trials to self-hosted deployments.
Comments (14)
Sign in to join the discussion.
N
Nadia V.4h ago
A team of 5 for 300k is either a flex or a recruiting ad. Probably both.
T
Tanya Belov3h ago
ok wait, 5 people and 300k and you got Fable 5 parity? Someone at the big labs is going to have a very uncomfortable Monday standup.
K
Kenji Mura4h ago
The tweet buried the lede. 'Weights in 2 weeks' should have been the hook, not the footnote.
P
Pablo Arenas3h ago
Also the thumbnail on that video is suspiciously good for a 5-person team. Who did the motion design, asking for my own launch next month.
P
Pablo Arenas4h ago
Rate limits on the API? Streaming support? Tool calling spec compatible with Fable's schema or do I have to rewrite my agent harness again?
O
oli4h ago
Every month someone claims frontier-class at 1/60th the cost. Then you plug it in and it forgets what a CSV is halfway through.
P
Priyanka R.4h ago
Been watching Bert ship for a while. The restraint in that launch copy is doing a lot of heavy lifting, no emojis, no benchmarks screenshot, just numbers.
M
mehmet the merciless4h ago
Ultramarine 1 Flash sounds like a Sherwin Williams paint chip, which, given the company name, is maybe the point.
Y
Yama O.4h ago
The knowledge work model market is cooked. In 18 months nobody will pay for inference, this will all be ambient.
O
oli3h ago
And 'similar performance to Fable 5' is doing Olympic-level gymnastics. Similar on what, MMLU from 2023?
I
Ingrid Nilsen4h ago
Building in the eval space, happy to run Ultramarine through our harness this week. Curious how it holds up on long-context retrieval, that's where most Flash tier models faceplant.
D
Doug P.3h ago
Is this one of those LLM things? I'm still evaluating whether chatbots have product-market fit.
M
Mikhail Orlov3h ago
Reminds me of the early days with one of my portfolio co's, same lean post-training setup. The 5-person team detail is what actually moves me, not the price.
A
amara zen3h ago
Open weights in 2 weeks is the real launch. Everything before that is a press release.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Paper Instruments
Ultramarine 1 Flash: frontier-class model for knowledge work
Ultramarine 1 Flash: frontier-class model for knowledge work
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Paper Instruments is releasing Ultramarine 1 Flash, an open model built for knowledge work and aimed at the document-heavy workloads of banks, law firms, and consultancies, where it can deliver Excel models, slide decks, or Word memos for cents per task . On the company's internal benchmark for client-ready deliverables, Ultramarine 1 Flash scored 59.3 at $0.13 per task, compared with 57.8 at $8.19 per task for Claude Fable 5 , which is the comparison point behind the roughly 60x cost claim in the launch tweet. The model is a post-training of GLM-5.3 Flash, completed by a team of five for approximately $300,000 all-in. Because a diligence memo or a board deck has no single correct answer, each training task was graded against its own rubric . That rubric-graded approach is the core bet here, pushing a mid-size open base model toward structured, judgment-heavy output rather than general chat performance. Paper Instruments was founded in 2025 by Gavin Bains, Joseph Besgen, and Daanish Khazi , is based in San Francisco, and went through Y Combinator's Spring 2025 batch. Khazi, the CEO, is announcing the launch publicly. The model is available now through the Paper Instruments API at $0.15 per million input tokens and $0.50 per million output tokens , with open weights to follow in two weeks, giving buyers at regulated firms a path from hosted API trials to self-hosted deployments.
Comments (14)
A team of 5 for 300k is either a flex or a recruiting ad. Probably both.
ok wait, 5 people and 300k and you got Fable 5 parity? Someone at the big labs is going to have a very uncomfortable Monday standup.
The tweet buried the lede. 'Weights in 2 weeks' should have been the hook, not the footnote.
Also the thumbnail on that video is suspiciously good for a 5-person team. Who did the motion design, asking for my own launch next month.
Rate limits on the API? Streaming support? Tool calling spec compatible with Fable's schema or do I have to rewrite my agent harness again?
Every month someone claims frontier-class at 1/60th the cost. Then you plug it in and it forgets what a CSV is halfway through.
Been watching Bert ship for a while. The restraint in that launch copy is doing a lot of heavy lifting, no emojis, no benchmarks screenshot, just numbers.
Ultramarine 1 Flash sounds like a Sherwin Williams paint chip, which, given the company name, is maybe the point.
The knowledge work model market is cooked. In 18 months nobody will pay for inference, this will all be ambient.
And 'similar performance to Fable 5' is doing Olympic-level gymnastics. Similar on what, MMLU from 2023?
Building in the eval space, happy to run Ultramarine through our harness this week. Curious how it holds up on long-context retrieval, that's where most Flash tier models faceplant.
Is this one of those LLM things? I'm still evaluating whether chatbots have product-market fit.
Reminds me of the early days with one of my portfolio co's, same lean post-training setup. The 5-person team detail is what actually moves me, not the price.
Open weights in 2 weeks is the real launch. Everything before that is a press release.