Ultramarine 1 Flash vs Claude Haiku 4.5: Two Cost-Per-Task Numbers That Do Not Measure the Same Thing

The two models in this comparison both publish a cost per task, and the temptation is to put the numbers next to each other and subtract. Paper Instruments says Ultramarine 1 Flash finishes a job for $0.13. Artificial Analysis measures Claude Haiku 4.5, in its reasoning configuration, at $0.28 per completed Intelligence Index task. One is less than half the other, and if that were the whole picture the decision would take ten seconds. It is not the whole picture, because those two figures come from different suites that ask the model to do different things, and understanding which suite resembles your work is the entire comparison. The rates themselves are further apart still: $0.15 per million input and $0.50 output for the Paper Instruments model against $1.00 and $5.00 for A​nthropic's — a 6.7x gap on input and 10x on output.

There is a second asymmetry. Claude Haiku 4.5 has a published score from a lab that is not its maker. Ultramarine 1 Flash has a published score from the company that trained it, on a suite that company built. Both facts matter more here than any single benchmark number, and neither shows up in a side-by-side.

The two rate cards, and the line that is not the same on both

Ultramarine figures are Paper Instruments' published API pricing and its own published evaluations. Claude Haiku 4.5 pricing is A​nthropic's rate card; its evaluation figures are Artificial Analysis measurements, labelled as such.

• Input — Ultramarine 1 Flash $0.15 per million tokens vs Claude Haiku 4.5 $1.00. A​nthropic's model costs 6.7x more going in.

• Cached input — $0.03 vs $0.10 per million, a 90% discount on the A​nthropic side against an 80% discount on the Paper Instruments side. The absolute gap stays wide, the proportional one narrows.

• Output — $0.50 vs $5.00 per million. 10x, and this is the line where the difference actually lands, because document work generates a great deal of output.

• Context — 131,072 tokens in the training configuration the technical report specifies vs 200,000 on the Haiku rate card. A​nthropic holds roughly 1.5x more.

• Maximum output — not published by Paper Instruments vs 64,000 tokens for Claude Haiku 4.5. One blank cell in the list.

• Inputs — Claude Haiku 4.5 accepts text, image and file uploads. Ultramarine 1 Flash is built around text and document work; its base, GLM-5.3 Flash, is natively multimodal, but Paper Instruments does not publish a modality list for the derivative.

• Measured cost per completed task — $0.13 on PI Knowledge Work for Ultramarine 1 Flash vs $0.28 on the Intelligence Index for Claude Haiku 4.5 (reasoning). The numbers people will quote at each other. The next section is why they are not the same unit.

What each score actually covers

Claude Haiku 4.5's published record comes in two configurations and they are worth separating, because a search result will usually show you the wrong one. Artificial Analysis lists the non-reasoning configuration at an Intelligence Index of 15 with 87.3 output tokens per second. The reasoning configuration — the one with a cost-per-task figure attached, and the one this piece uses throughout — measures at an Intelligence Index of 17, 90.0 output tokens per second, $1.00 input and $5.00 output per million, a 90% cache discount, $0.28 per completed Index task and 78 million output tokens of verbosity. Its context window is 200,000 tokens with a knowledge cutoff of July 2025, and it accepts text and image.

The Intelligence Index is assembled from mathematics, coding, reasoning and agentic evaluations. None of its component tests asks a model to produce a board deck, a diligence memo or an Excel model with a live forecast inside it, because those artifacts have no single correct answer and a benchmark that scored them would need a rubric per task.

Which is exactly what Paper Instruments built, and exactly why the comparison cannot be a subtraction. PI Knowledge Work is a suite of client-ready deliverables in finance, law, consulting and operations, graded task-by-task against authored rubrics by a rubric-grading model. Ultramarine 1 Flash scores 59.3 there at $0.13 per task. A score of 59.3 on a document-deliverable rubric and a score of 17 on a mathematics-and-coding index are not comparable quantities. Neither is the cost figure, because a task on one suite can be an order of magnitude longer than a task on the other.

What the pairing does support is narrower and still useful. Claude Haiku 4.5 has been measured by someone with no stake in it, on tests nobody at A​nthropic designed. Ultramarine 1 Flash has not. That is not a criticism of the Paper Instruments numbers — the company discloses its method in unusual detail — but it is a difference in the kind of evidence on offer, and a buyer putting the choice in front of a risk committee will hit it immediately.

Where Claude Haiku 4.5 wins without argument

Three places, and none of them is about being smarter.

Multimodal input is the clearest. A scanned contract, a screenshot of a broken spreadsheet, a photograph of a whiteboard — all of that goes into Claude Haiku 4.5 through an interface that has supported it, unchanged, for a year. Ultramarine 1 Flash rests on a base model that reads images and video, but the vendor publishes no modality list for its derivative, and no example in the technical report turns on an image the model has to look at. If your documents arrive scanned, this decides the question on its own.

Operational maturity is the second. A model with a year of production traffic behind it has had its failure modes found and documented by people who were not being paid to find them. A model that was published on 14 October 2026 has not, and the first week of any new endpoint is when you learn what it does under load.

Independent verification is the third, and it is the one that compounds. Every additional month of Haiku 4.5's life is another month in which somebody ran it on their own work and wrote down what happened.

Where the gap closes to almost nothing — and why the arithmetic is misleading

The headline 10x on output is real, and on a long document job it produces exactly the bill you would predict. What the headline hides is that the per-task economics of the two models can converge, because a task on PI Knowledge Work is not the same quantity of work as a task on the Intelligence Index — and because the A​nthropic model is a heavily cached one in practice.

There is a counter-intuitive number on the Paper Instruments side worth reading properly. Across its six benchmarks, Ultramarine 1 Flash generates 6.5% fewer output tokens than the GLM-5.3 Flash model it is post-trained from — 30.5 million against 32.7 million, reasoning included — while scoring better on all six. Token efficiency that improves is not the usual direction of travel for a post-training run, and on a per-million-token rate card it is a second discount stacked on top of the first. That is the one comparison on the Paper Instruments page where both sides are measured by the same method on the same suite, which makes it the most trustworthy number in the release.

Against that, Claude Haiku 4.5 is the cheaper model on one comparison nobody advertises: at $0.28 per completed Index task it is priced for a suite whose tasks are shorter and less agentic than a diligence memo. Put a real legal or financial deliverable in front of each and the $0.13 figure was never measuring the same unit of work as the $0.28 one.

Both models are on one key through OrcaRouter if you want to settle this empirically rather than by reading. Claude Haiku 4.5 is served at $1.00 and $5.00 per million tokens with nothing added per token — the provider's list price passed straight through, so a vendor rate change is live on our side the same day, which for a model whose price has moved before is not a trivial property. Running the same twenty documents through both and reading the invoices costs less than the meeting about which one to pick, and it produces a number instead of an argument.

The decision, stated as a fork

If the input is a document someone scanned or a screenshot, the answer is Claude Haiku 4.5 and there is nothing further to discuss.

If the input is text and the output is a finished deliverable — a model, a deck, a memo — and the volume is high enough that $0.15 versus $1.00 on input changes the shape of the project, Ultramarine 1 Flash is worth a pilot on your own work. It is a day old. The $0.13 is vendor-reported on a vendor-built suite, and no independent evaluation exists yet. But the rate card is real, the token-efficiency comparison is measured the only way it can be measured honestly, and the gains over its own base model on the four benchmarks it leads are consistent rather than cherry-picked.

If your workload includes both, the useful move is not to choose. A routing rule can hold the scanned-document path on the A​nthropic model and send the text-to-deliverable path to whichever of the two clears your quality bar on a sample — and if the newer one falls over, the request goes to the older one without a code change and without a second contract to sign.

What would resolve the open question is a single independent evaluation of Ultramarine 1 Flash on a suite its maker did not write. Until that exists, the honest position is that Claude Haiku 4.5 is the model whose behaviour you can predict and Ultramarine 1 Flash is the model whose price you would like to be able to use. Nothing published today says you cannot. Nothing published today, except the vendor's own page, says you can.

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.