Objective: To determine whether robot-assisted training is cost-effective compared with an enhanced upper limb therapy (EULT) programme or usual care. Design: Economic evaluation within a randomised controlled trial.Setting: Four National Health Service (NHS) centres in the UK: Queen’s Hospital, Barking, Havering and Redbridge University Hospitals NHS Trust; Northwick Park Hospital, London Northwest Healthcare NHS Trust; Queen Elizabeth University Hospital, NHS Greater Glasgow and Clyde; and North Tyneside General Hospital, Northumbria Healthcare NHS Foundation Trust. Participants: 770 participants aged 18 years or older with moderate or severe upper limb functional limitation from first-ever stroke.Interventions: Participants randomised to one of three programmes provided over a 12-week period: robot-assisted training plus usual care; the EULT programme plus usual care or usual care. Main economic outcome measures: Mean healthcare resource use; costs to the NHS and personal social services in 2018 pounds; utility scores based on EQ-5D-5L responses and quality-adjusted life years (QALYs). Cost-effectiveness reported as incremental cost per QALY and cost-effectiveness acceptability curves. Results: At 6 months, on average usual care was the least costly option (£3785) followed by EULT (£4451) with robot-assisted training being the most costly (£5387). The mean difference in total costs between the usual care and robot-assisted training groups (£1601) was statistically significant (p<0.001). Mean QALYs were highest for the EULT group (0.23) but no evidence of a difference (p=0.995) was observed between the robot-assisted training (0.21) and usual care groups (0.21). The incremental cost per QALY at 6 months for participants randomised to EULT compared with usual care was £74 100. Cost-effectiveness acceptability curves showed that robot-assisted training was unlikely to be cost-effective and that EULT had a 19% chance of being cost-effective at the £20 000 willingness to pay (WTP) threshold. Usual care was most likely to be cost-effective at all the WTP values considered in the analysis. Conclusions: The cost-effectiveness analysis suggested that neither robot-assisted training nor EULT, as delivered in this trial, were likely to be cost-effective at any of the cost per QALY thresholds considered.