The percentages are the product
Unisound leads this upgrade with usage, not a capability score. U2-Flash-1118, released on November 18, 2026, is pitched on using about 37.0% fewer tokens and making about 44.8% fewer model calls than the U2-Flash it upgrades, measured, the company says, on the same task set under the same evaluation conditions. For an agent model, that is a meaningful claim if it holds: tokens and calls are the two usage lines that feed an invoice, and a model that reaches the same result in fewer steps uses less of both.
The reports of the filing then stop one step short of making the claim checkable in money terms. They convert the percentages into no cost saving, state no price for the model, and give no timetable for when anyone outside the company can use it. Token efficiency is an input to serving cost, not a disclosed saving, and actual savings would depend on how tokens and calls are billed. A reader can conclude that U2-Flash-1118 uses fewer tokens and calls per task than its predecessor, on the company's own measurement. A reader cannot conclude what that is worth in currency, or whether any saving reaches a customer as a lower price or stays with Unisound as margin.
Rankings without scores
The capability side of the announcement has the opposite shape: plenty of placement, no numbers. Unisound reports top-two finishes in three visual evaluations, second place in DeepSWE v1.1, and first place in GPQA-Diamond, and it names the models it compared against, including DeepSeek-V4.1-Flash-1017, GLM-5.3-Flash, MiniMax-M3.1-Flash, Qwen3.8-Max-0902 and Kimi K3. Naming the field is better than an unnamed comparison. But a ranking without the scores behind it hides the margin. First place by a wide gap and first place within rounding look identical in this format, and the evaluations themselves were selected by the company, with no independent verification disclosed in the reports reviewed.
That matters more for this release than for a pure capability launch, because the efficiency claim depends on quality holding steady. Unisound says the token reduction comes while "maintaining the quality of task completion". The reported rankings establish placement in the cited comparisons. They do not establish that the upgrade maintained quality on the tasks used to measure the token and call reductions: relative placement on chosen tests is not matched-task quality on the efficiency task set.
What actually changed in the model
The concrete change is visible. The language base keeps its original parameter size, and a vision module of about 1 billion parameters is added, so the model can take images natively: interface screenshots, design drafts and document layouts, with image-text question answering and screenshot-assisted development among the stated uses. For a coding and agent model, that is a practical addition rather than a decorative one.
The efficiency is attributed by the company to a Recursive Self-Improvement mechanism, which converts execution feedback into targeted training tasks, and to trajectory optimization with the company's Agent Harness, reducing repeated exploration, redundant tool calls and context transfer. Those are plausible places for a 44.8% reduction in calls to come from. They are also described only in the company's own terms, in a voluntary filing the exchange does not stand behind.
What would settle it
Three disclosures would turn this announcement from a milestone claim into evidence: the benchmark scores behind the rankings, a price or a dated availability plan, and any sign of outside use, whether a named customer or third-party evaluation results. Until one of those appears, U2-Flash-1118 is best read as a statement of direction from a listed company to its investors: agent economics, not leaderboard position, is where it wants to be judged. On that stated test, it has published the usage percentages and left readers unable to assess the three things that would price them: benchmark margins, matched-task quality, and actual cost per task.
Sources
- [1] Reuters, "Unisound AI Technology Releases U2-Flash-1118, Upgraded Version Of U2-Flash", via TradingView, published November 18, 2026 (secondary report of the filing, read first-hand)Read source
- [2] Minichart, "Unisound Releases U2-Flash-1118, Claims 37% Token Cost Cut vs Original Model", November 18, 2026 (detailed summary of the announcement, read first-hand in full)Read source
- [3] TipRanks, "Unisound Releases Upgraded U2-Flash-1118 Large Model", published November 18, 2026 (corroborates the vision module, rankings and RSI description)Read source