A 4B model edged the GPT-5 family at bargaining. The interesting number is the spread

Rank the four scores and the figure worth keeping is the spread, not the winner. Top to bottom is 0.014 utility points, about 2.2% of the leading score ^[2]. Inside that band, GPT-5.1 and GPT-5.2 both sit below GPT-4.1 ^[3], which is the part of the table nobody put in a headline: on these bargaining tasks, two later frontier releases scored worse than an earlier one. That is what makes the small-model result readable. Whatever this benchmark measures, it is not parameter count.

It does appear to measure the training objective. SocialRL rewards strategic behaviour rather than pure helpfulness, and the trained policy holds information back and anchors, conceding only where concession serves the user's overall utility ^[10]. The behaviour it corrects is not a knowledge gap but a disposition: unprompted disclosure and premature concession, which Microsoft casts as the obstacle to an agent that actually represents its principal ^[11]. A model that folds under pushback is doing what it was rewarded for.

The per-task figures say the advantage is uneven. Closure of the baseline-to-frontier gap runs from 73% to 122% depending on the task ^[8], and anything above 100% is an overshoot, so some of the six domains (Deal-or-No-Deal, CaSiNo, Craigslist, Job Interview, Calendar and Marketplace) ^[4] are carrying the average while others contribute nothing to it.

The cost argument everyone will draw from this is not actually in the work. As reported, the material contains utility scores, anchoring rates, gap-closure percentages and the parameter count of the model Microsoft trained, and no inference prices and no parameter counts for the GPT-5 family it was compared against ^[5]. So the claim on offer is a capability claim with the economics left to the reader to supply. The economics an operator would need are the ones nobody priced here: gathering domain transcripts and re-running the cascade when a counterparty's tactics change.

Two structural caveats before anyone reallocates an inference budget. Consolidating specialists into one policy worked within a family of similar scenarios and did nothing for domains with unique dynamics ^[9], so the single-model story holds inside a family and breaks between families. And this is Microsoft Research grading its own method, in a paper relayed through cryptobriefing.com's account of it ^[1][12]. The average lead will not survive a different scenario mix. The change in disposition probably will.