Ultramarine 1 Flash vs Gemini 3.5 Flash-Lite: Delivering an Artifact and Answering a Question Are Two Different Products

Start with the line that ends the argument rather than the one that starts it. Gemini 3.5 Flash-Lite accepts text, images, speech and video and returns text. Ultramarine 1 Flash takes text and returns an Excel workbook, a PowerPoint deck or a Word document, edited in place, having read and written to the business systems the work lives in. Both are cheap. G​oogle's model scores 22 on the Artificial Intelligence Index at $0.19 per completed task; Paper Instruments reports $0.13 per task on its own internal suite. The numbers are close enough that the cost story is a tie, and what is left is the question of whether you are buying an answer or buying a deliverable — because on that question these two models are not in the same market at all.

They were also not built at the same time or for the same reason. Gemini 3.5 Flash-Lite reached general availability in July 2026 as G​oogle's most cost-efficient tier, sitting at roughly a fifth of 3.6 Flash's price. Ultramarine 1 Flash was published on 14 October 2026 as a reinforcement-learning post-training of GLM-5.3 Flash, aimed at a client-ready deliverable rather than a reply.

The spec lines, both sides

Ultramarine figures are Paper Instruments' published API rates and its own published evaluations; its context figure is the training configuration stated in the technical report. G​emini figures are G​oogle's rate card plus Artificial Analysis measurements, labelled where they are measurements.

• Input price — Ultramarine 1 Flash $0.15 per million tokens vs Gemini 3.5 Flash-Lite $0.30. G​oogle is 2x more expensive going in.

• Output price — $0.50 vs $2.50 per million. 5x, and this is the line where a document job's cost actually accumulates.

• Cached input — $0.03 per million (80% off) vs $0.03 (90% off). The same absolute number, reached by different discounts.

• Context window — 131,072 tokens in the training configuration vs 1,048,576. G​oogle holds 8x more.

• Maximum output — not published vs 65,536 tokens.

• Inputs — text and document work vs text, image, speech and video.

• Cost per completed task — $0.13 on PI Knowledge Work vs $0.19 on the Intelligence Index. The second of the two numbers in the opening, and the one to be most careful with.

• Output speed — not published for Ultramarine 1 Flash vs 425.0 output tokens per second for Gemini 3.5 Flash-Lite, fourth fastest of 148 models in its class.

The latency number inside the speed number

Gemini 3.5 Flash-Lite is one of the fastest models anybody has measured. It is also slow to begin. Artificial Analysis records a time to first token of 8.13 seconds on G​oogle's own API, against a median of 1.85 seconds for models in a comparable price tier. The model streams at 425 tokens per second once it starts and takes eight seconds to start.

That is not a defect; it is a property of the serving path, and it tells you which workloads the tier was designed for. A batch classification job, an extraction pipeline, a background agent — none of them notice an eight-second start, and all of them benefit enormously from finishing a 500-token response in about a second. A human waiting on a chat reply notices nothing else. The word "Flash" is describing throughput, not responsiveness, and the two are not the same measurement.

Ultramarine 1 Flash has no published speed or latency figure at all. Not from Paper Instruments, not from any third party. There is no basis to claim it is faster or slower than G​oogle's model, and anyone who tells you otherwise is inferring from architecture rather than reading a number. What the technical report does disclose is the operating envelope it was trained inside: a 131,072-token context, a 250-turn maximum on the agent loop, and context compaction triggered at 85% of the window or when 30,000 tokens remain, whichever comes first. A harness that compacts its context by design is a harness built for long work that would otherwise overflow — which is a hint about the shape of its tasks and not a measurement of how quickly it produces them.

What a "task" is worth depends on who wrote the benchmark

The $0.13 and the $0.19 look comparable and measure different things, and this is the part of the comparison a side-by-side table will hide from you.

The Intelligence Index is a composite of ten evaluations — mathematics, coding, reasoning, agentic work. Its tasks have answers a grader can check. Gemini 3.5 Flash-Lite's 22 is a real score from an evaluator with no stake in the model, and its $0.19 is a measured cost per completed task across that suite. G​oogle also has the more flattering of the two verbosity records: 59 million output tokens to complete the Index, ranked 23rd of 60 and described as fairly concise against a median of 95 million. That is a model engineered not to ramble.

PI Knowledge Work is not that. It is a suite of client-ready deliverables graded against a rubric authored for each task, because a diligence memo has no correct answer to check against. Ultramarine 1 Flash scores 59.3 there at $0.13 per task, against Claude Fable 5 at 57.8 for $8.19 and GPT-5.6 Sol at 55.1 for $2.09. The vendor reports that its suite correlates with the Artificial Analysis Intelligence Index at r = 0.81 — a serious number, and a secondhand one. A task on one suite is not a task on the other: a reasoning question can be a few hundred tokens, and a board deck is a long agentic run producing a file. The two per-task figures are not units of the same thing, and subtracting them is arithmetic on a category error.

What survives that criticism is narrower but genuinely useful. Paper Instruments measured token efficiency against something other than itself: across all six of its benchmarks, Ultramarine 1 Flash emitted 6.5% fewer output tokens than the GLM-5.3 Flash model it descends from — 30.5 million against 32.7 million, reasoning included — while beating it on every one of the six. Same suite, same harness, both sides measured. On a per-token rate card, a model that writes less for a better result is a discount applied twice.

Two working models on one key

Gemini 3.5 Flash-Lite is served on OrcaRouter at $0.30 input and $2.50 output per million tokens, the provider's list price passed straight through with nothing added per token — so a G​oogle rate change is live here the same day, without a deploy on our side. Alongside it sit the models this comparison keeps referring to: GPT-5.6 Sol, Claude Fable 5 and GLM-5.3 Flash, the base Ultramarine 1 Flash was trained from, are all reachable from the same endpoint and the same key.

That matters more for this pairing than for a typical head-to-head, because the workload is mixed. A document pipeline built around a million-token window and four input modalities is a different pipeline from one built around a deliverable-producing agent, and the two usually coexist in the same team. A routing rule can hold scanned or audiovisual input on G​oogle's model and send the text-to-deliverable path elsewhere, with a declared fallback if the first choice fails — one contract instead of two, and no code change when you change your mind about which model gets the long jobs.

Ultramarine 1 Flash itself is not on our catalogue; the access route today is the vendor's own API and the third-party platforms that resell it.

How I would actually decide

Two situations, and they resolve cleanly.

If the input contains anything that is not text — a photographed contract, a call recording, a screen capture, a video walkthrough — G​oogle's model is the only one of the two that can accept the request. That is not a preference, it is a capability boundary, and it ends the comparison for that workload.

If the output is the thing being paid for — a model, a deck, a memo, a document that a person will open and edit — then the question becomes whether a post-trained specialist beats an efficient generalist, and Ultramarine 1 Flash is the specialist. The evidence for that is the vendor's, on the vendor's benchmarks, one day after release, and it should be read as a hypothesis rather than a finding. But it is a specific and testable hypothesis: on the four of six suites Paper Instruments leads, the largest margins are on JobBench and FrontierFinance, the two closest to real deliverable work; on APEX-Agents 1.1 and DiligenceBench, which are agent-loop and research tasks, Gemini 3.8 Flash and Claude Fable 5 lead and Ultramarine 1 Flash places third.

The unresolved question is not which model is better. It is whether the deliverable-producing behaviour survives contact with documents nobody at Paper Instruments wrote. The company has published its method in enough detail to invite exactly that test, and has not published the weights that would let someone run it without going through its API. Both of those facts are load-bearing, and only one of them is likely to change soon.

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.