Ultramarine 1 Flash vs DeepSeek V4 Flash: A Cheap Token Rate Is Not a Cheap Task

Ultramarine 1 Flash costs $0.15 per million input tokens and $0.50 output. D​eepSeek V4 Flash is listed at $0.44 and $1.32 on the Artificial Analysis record, or $0.22 and $0.66 at the routed rate our catalogue serves it at. Set the two rate cards side by side and the Paper Instruments model looks like the obvious bargain — three times cheaper going in and two and a half times cheaper coming out, at the list prices. The problem with that reading is that a per-million-token rate is only half of a bill, and the half it leaves out is the one where these two models differ most. D​eepSeek V4 Flash generates 240 million output tokens to complete the Intelligence Index. The median model in its comparison class generates 140 million. It is, in Artificial Analysis' phrasing, very verbose — ranked 36th of 40 in its price class on that axis.

Ultramarine 1 Flash makes the opposite claim about itself. Across all six benchmarks Paper Instruments evaluated, it generated 6.5% fewer tokens than the GLM-5.3 Flash model it was post-trained from, while scoring higher on every one of them. One model is cheap per token and expensive per task. The other went out of its way to become less expensive per task without changing its price. The comparison worth making is between those two approaches, not between two input rates.

Two models, two different routes into the same price tier

They arrived at the budget tier from opposite directions, and the difference explains most of what follows.

D​eepSeek V4 Flash is a from-scratch mixture-of-experts model: 284 billion total parameters, 13 billion active, released on 24 April 2026 under an MIT licence with weights published on Hugging Face. Its 0731 revision, re-post-trained from scratch on the same architecture, is what the dated identifier now serves. It holds a 1,048,576-token context window with up to 384,000 output tokens, takes text in and returns text out, and reports strong agent benchmarks on D​eepSeek's own published numbers — 82.7 on Terminal-Bench 2.1, 76.7 on Cybergym, 70.3 on Toolathlon Verified.

Ultramarine 1 Flash is a derivative. Paper Instruments took GLM-5.3 Flash — Z.ai's 320B-total, 18B-active multimodal model — and ran large-scale reinforcement learning on it over a synthetic curriculum of finance, legal, consulting and operations work. The API went live on 14 October 2026 at $0.15 / $0.03 cached / $0.50 per million tokens, the same rates GLM-5.3 Flash charges. The weights did not go live; the vendor's own page reads "Model weights coming soon."

So the honest framing of this pairing is a downloadable from-scratch MoE against a closed post-training of somebody else's open base, one of them six months into production and the other one day old.

The verbosity number neither side's marketing leads with

Artificial Analysis measures how many output tokens a model spends to finish the Intelligence Index, and reports it as a verbosity ranking. D​eepSeek V4 Flash used 240 million output tokens against a class median of 140 million — 71% above typical, 36th of 40 in its class. That figure is the reason to be careful with the $0.22 per task it posts. It is also, notably, a better cost-per-task number than most of its competitors achieve, which tells you how much the model's raw speed compensates: at 173.7 output tokens per second it is the eighth fastest of 74 models in its class, so the token volume arrives quickly even if there is more of it.

Ultramarine 1 Flash's verbosity is not published by Artificial Analysis, because Artificial Analysis has not evaluated the model. What Paper Instruments publishes instead is the comparison against its own base: 30.5 million tokens against GLM-5.3 Flash's 32.7 million across the six benchmark suites, reasoning included. That is a measurement both sides of the comparison share — same suite, same harness, same conditions — which makes it the most trustworthy figure in the release and the only one where a token-efficiency claim is verified against something other than itself.

Put those together and the arithmetic changes shape. D​eepSeek V4 Flash's advantage is a fast, capable, well-measured model you can run almost anywhere. Its disadvantage is that it writes a lot, and on a fixed output rate that shows up in the invoice. Ultramarine 1 Flash's advantage is that it writes less than the model it descends from — but only against that model, and only on suites its own maker wrote.

What the six benchmarks actually say

Paper Instruments ran six suites against its own model and several competitors. D​eepSeek V4.1 Flash appears in them; the original D​eepSeek V4 Flash, the model in this headline, does not.

• JobBench — Ultramarine 1 Flash 62.9, D​eepSeek V4.1 Flash 60.4. The Paper Instruments model leads by 2.5 points.

• PI Knowledge Work — Ultramarine 1 Flash 59.3, D​eepSeek V4.1 Flash 47.9. An 11.4-point gap, the widest in the set.

• FrontierFinance — 53.3 against 46.0.

• AutomationBench — the one suite where D​eepSeek wins: V4.1 Flash 51.3 against Ultramarine's 50.2.

• DiligenceBench — 50.6 against 50.3. Effectively level, with both trailing Gemini 3.8 Flash and Claude Fable 5.

• APEX-Agents 1.1 — 57.0 against 39.5, the largest margin anywhere in the table.

The caveats stack here and it is worth being explicit about all of them. These are the vendor's benchmarks. The D​eepSeek figures are for a different, newer model in the same family, which Paper Instruments identifies by name. And none of it has been checked by anyone outside the company. What the table does establish is narrow but real: on document-deliverable work, the post-trained specialist holds a large lead over an efficient generalist MoE, and on automated agent tasks the gap nearly disappears. If your work is a memo or a model, that split points one way. If it is a multi-step agent loop, it does not.

You can download one of these

This is the difference no benchmark captures, and it is the one most likely to decide a procurement conversation.

D​eepSeek V4 Flash is MIT-licensed with weights on Hugging Face. You can download it, run it on your own hardware, fine-tune it, ship it inside a product, and never send a token to a third party. For a law firm or a bank with client-confidentiality obligations, that is not a nice-to-have; it is frequently the whole requirement, and it takes D​eepSeek's model off the table from the hosted-API half of this comparison entirely and puts it on your own machines instead.

Ultramarine 1 Flash cannot do that today. "Model weights coming soon" is a statement of intent, not a licence. What Paper Instruments did release openly is the scaffolding around the model: the Feather agent harness it was trained and evaluated inside, and the paper-docx, paper-pptx and paper-xlsx forks that give an agent a safe interface to Office file formats. Those are genuinely useful — and they are also an advertisement for the model, since they exist to make its outputs less corrupt than what a generic agent produces with the standard libraries.

So the licence line reorganises the comparison. If you must run the weights yourself, there is no comparison: D​eepSeek V4 Flash is the answer and Ultramarine 1 Flash is a promise. If you can call an API, the two are genuinely competing, and the deciding evidence is the per-task cost on your own documents.

Where the routed price lands

On OrcaRouter, D​eepSeek V4 Flash is served at $0.22 input and $0.66 output per million tokens — the provider's list price passed through with nothing added per token, which is why it sits below the $0.44/$1.32 pair on the Artificial Analysis record: that pair tracks a different provider's card. The pass-through arrangement has a second consequence worth knowing if you are budgeting: when a provider moves a rate, the served rate moves with it the same day, without a config change on your side.

Against that routed rate, Ultramarine 1 Flash's $0.15/$0.50 is about 1.5x cheaper on input and 1.3x on output — much narrower than the list-price comparison suggests, and narrower still once D​eepSeek's 97% cache discount is applied to a document workload with a large shared prefix. Ultramarine 1 Flash's cached rate is $0.03 per million, also a strong discount, but the two are close enough at this point that neither wins on arithmetic alone.

Which leaves the decision where it belongs: on your own work. Send a batch of real deliverables to both, compare the invoices, and read the outputs. Ultramarine 1 Flash is not on our catalogue, so that test runs against the vendor's own endpoint on one side. The model it is post-trained from, GLM-5.3 Flash, is served here at $0.15/$0.50 — the identical rate card — which gives you a reasonable proxy for what the base produces before you add the post-training on top.

The position, without a winner

These two are not really substitutes, and forcing them into a winner obscures the useful finding. D​eepSeek V4 Flash is a downloadable, independently measured, unusually fast MoE that writes more than it needs to. Ultramarine 1 Flash is a closed, unmeasured post-training of an open base that was engineered to write less, sold on a benchmark suite nobody else has run.

The dismissible version of this article would say the cheaper token rate wins. The defensible version says the license decides procurement, the verbosity figures decide the bill, and neither model has been tested on your documents. Two of those three you can resolve this week. The third cannot be resolved until Paper Instruments either publishes independent numbers or ships the weights it says are coming.

Comparisons like this one are also where the failure modes show up: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.