Glonce

Ultramarine 1 Flash claims to beat pricier models on its own benchmark

modelsThe Glonce editorsHacker News🔥59Glonce rating59Impact40Novelty50Relevance100Surprise55Players20Credibility60
Ultramarine 1 Flash claims to beat pricier models on its own benchmark

A vendor page presents Ultramarine 1 Flash, a model built on the base model GLM-5.3 Flash, under the line "Fable-class model for cents per task." The page says the model sits on the cost-quality frontier at the most attractive price point.

The central claim: at $0.13 per task, Ultramarine 1 Flash outscores Claude Fable 5 (max) and GPT-5.6 Sol (max), which cost $8.19 and $2.09 per task respectively. It also comes within three points of Claude Opus 5.5 (max) and GPT-6 Sol (max), which cost $6.96 and $1.26 per task. The page gives no actual scores, only these costs and the gap.

The measurement is PI Knowledge Work, which the page calls a carefully built internal benchmark. It says the scores correlate strongly with the Artificial Analysis Intelligence Index (r = 0.81). The benchmark covers professional work in finance, law, consulting and operations. Its tasks ask the model to deliver finished Excel workbooks, PowerPoint decks and Word documents, and to work inside live business systems such as ERP and MES. A rubric written for each task scores the result on substance, accuracy and presentation. According to the page, every model works in the same harness, and costs are computed at each developer's official API pricing.

On efficiency, the page says that across six benchmarks Ultramarine 1 Flash generates 6.5% fewer tokens than its base, GLM-5.3 Flash, "taking a shorter path to better work." The six benchmarks are not named.

The Ultramarine API prices are, per million tokens: $0.15 for input, $0.03 for cached input and $0.50 for output. The page says these are the same as GLM-5.3 Flash. It also points readers to a technical report on the kind of professional data and evaluations that trained the model.

The page describes the model's style as "crafted with a sense of client readiness": it writes clean, simple, straightforward prose, has a strong sense of visual and information hierarchy, and is built to follow the same rules required of analysts in professional services. As a showcase it offers a Nucor example: five periods of segment sales and earnings in one table, with every subtotal and total reconciling. In that output, steel mills earnings fell from $7.2bn in 2022 to $2.4bn in 2025, while their share of sales rose from 58% to 62%.

Key facts

  • Ultramarine 1 Flash, built on GLM-5.3 Flash, is claimed to outscore Claude Fable 5 (max) and GPT-5.6 Sol (max) at $0.13 per task, against $8.19 and $2.09 for those two.
  • It is said to come within three points of Claude Opus 5.5 (max) and GPT-6 Sol (max), which cost $6.96 and $1.26 per task.
  • The yardstick is PI Knowledge Work, the vendor's internal benchmark of finance, law, consulting and operations work, reported to correlate with the Artificial Analysis Intelligence Index at r = 0.81.
  • API pricing per million tokens: $0.15 input, $0.03 cached input, $0.50 output, the same as GLM-5.3 Flash.
  • Across six benchmarks the model generates 6.5% fewer tokens than GLM-5.3 Flash.

Why it matters

The pitch is near-frontier quality on professional knowledge work at a tiny fraction of the cost. On the page's own figures, $0.13 per task compares with $8.19 and $2.09 for the two models it says it outscores, and $6.96 and $1.26 for the two it trails by up to three points. If the claim held up, it would move the cost-quality trade-off for office-style tasks. At this stage it is a vendor's assertion about a vendor's benchmark.

Who it affects

Teams that would run model-written spreadsheets, slide decks and documents at volume are the obvious audience, as is anyone working in finance, law, consulting or operations workflows, which are the areas PI Knowledge Work covers. Developers already using GLM-5.3 Flash may also take note, since the page says Ultramarine's API pricing matches it.

How to use it

The model is offered through the Ultramarine API at, per million tokens, $0.15 for input, $0.03 for cached input and $0.50 for output. The page says this is the same as GLM-5.3 Flash. It points to a technical report for more detail on the professional data and evaluations behind the model. No release date, availability or licence (open or closed weights) is given in the visible text.

How solid is it

Thin so far. The comparison rests on PI Knowledge Work, an internal benchmark of the vendor, and the benchmark is described only by the vendor. The claimed r = 0.81 correlation with the Artificial Analysis Intelligence Index is offered as support, but no independent evaluation is mentioned. The actual benchmark scores for any model are not given, only costs per task and a gap of "within three points." The page says every model works in the same harness and that costs use each developer's official API pricing. The six benchmarks behind the 6.5% token figure are not named, and it is not stated whether the figure is a mean or median.

Risks and caveats

A self-run benchmark can favour the model it was built around, and the page does not show scores that would let a reader judge the margins. The Nucor example shows a reconciled table, but it is not stated whether it was verified independently. No parameter count, context window or training details are stated on the page; the technical report is only referred to.

“Ultramarine 1 Flash sits on the cost-quality frontier at the most attractive price point.”

— Ultramarine product page