AI21 says post-training Nemotron 3.5 Super VL cut costs across three enterprise tasks

AI21 says post-training Nemotron 3.5 Super VL cut costs across three enterprise tasks

AI21’s post dated October 15, 2026, says its team worked with Gong, Deel and Alice to post-train NVIDIA Nemotron 3.5 Super VL for three enterprise tasks. AI21 argues that organizations can tailor an open model to their own data and requirements, reducing the cost of relying on general models while gaining control over model behavior. For Gong, AI21 says post-training matched the quality of frontier models on extracting information from customer interactions, at a small fraction of the cost; Gong’s own evaluations supplied the comparison. For Deel’s payroll questions, AI21 reports 72% accuracy on Emet’s eight-country evaluation, eight points above Claude Opus 5.5, at 33 times lower cost per query. For Alice’s security task, AI21 and Alice report a 41-point improvement in attack resilience without harming overall task performance. These are company-reported results on specific evaluations and tasks, not evidence here of performance across other uses. The post says the work took three weeks overall; it gives two weeks for Gong’s work. It does not provide the evaluation details needed to assess how the comparisons were measured or how the models perform in deployment.

Publication: 1e099861-12e1-66eb-f6fd-e773d8d4e266

attribution — pass: This is a single AI21 blog post on its own site. The reading consistently attributes the claims to AI21, or to AI21 and its partners, and notes that Gong's own evaluations supplied the Gong comparison. Saying AI21 and Alice 'report' the Alice result is a reasonable reading of the post's 'Together, AI21 and Alice' wording.

Together, AI21 and Alice improved Nemotron 3.5 Super’s attack resilience by 41 points

summary — pass: The reading leads with the actor (AI21), the partners and the model. It describes post-training in plain terms as tailoring an open model to an organization's own data and requirements, and frontier model is defined on the page. It summarizes the three use cases and their results accurately and keeps the 'own rather than rent' argument. The title's claim of cost cuts across all three tasks matches the post's closing statement. The question follows from the evidence gap.

took all three tasks from the baseline evaluated to specialized intelligence in three weeks, reducing cost

date — pass: originalPublished is 2026-10-15T11:50:52Z, and the reading states October 15, 2026, which matches the UTC day. The source is a publisher page, not an archive copy, so no archive disclosure is needed. There is no hindsight framing, and the three-week and two-week durations are given at the source's precision. dateModified (October 21) is later, but no edit flag is set, and the claims used appear to be part of the core post.

Over the past three weeks, our team of Forward-Deployed Researchers (FDRs) post-trained

support — pass: Every result claim matches a source excerpt: Gong's speed and accuracy parity at a fraction of the cost, Deel's 72% accuracy (8 points above Claude Opus 5.5) at 33x lower cost, and Alice's 41-point resilience gain without harming task performance. The reading correctly frames these as company-reported results on specific evaluations. It does not turn them into deployment reliability, and it notes that evaluation details are missing.

answer accuracy rose to 72% on Emet’s 8-country payroll benchmark evaluation, sitting 8 points above Claude Opus 5.5, and costing 33x less per query.

media — pass: No media was selected, and no verified media candidates were supplied. Selecting none is appropriate.