BACKGROUND: The aim of the COSMIC-312 trial was to evaluate cabozantinib plus atezolizumab versus sorafenib in patients with previously untreated advanced hepatocellular carcinoma. In the initial analysis, cabozantinib plus atezolizumab significantly prolonged progression-free survival versus sorafenib. Here, we report the pre-planned final overall survival analysis and updated safety and efficacy results following longer follow-up. METHODS: COSMIC-312 was an open-label, randomised, phase 3 study done across 178 centres in 32 countries. Patients aged 18 years or older with previously untreated advanced hepatocellular carcinoma were eligible. Patients must have had measurable disease per Response Evaluation Criteria in Solid Tumours version 1.1 (RECIST 1.1), and adequate marrow and organ function, including Child-Pugh class A liver function; those with fibrolamellar carcinoma, sarcomatoid hepatocellular carcinoma, or combined hepatocellular cholangiocarcinoma were ineligible. Patients were randomly assigned (2:1:1) using a web-based interactive response system to a combination of oral cabozantinib 40 mg once daily plus intravenous atezolizumab 1200 mg every 3 weeks, oral sorafenib 400 mg twice daily, or oral single-agent cabozantinib 60 mg once daily. Randomisation was stratified by disease aetiology, geographical region, and presence of extrahepatic disease or macrovascular invasion. Dual primary endpoints were for cabozantinib plus atezolizumab versus sorafenib: progression-free survival per RECIST 1.1, as assessed by a blinded independent radiology committee, in the first 372 randomly assigned patients (previously reported) and overall survival in all patients randomly assigned to cabozantinib plus atezolizumab or sorafenib. The secondary endpoint was progression-free survival in all patients randomly assigned to cabozantinib versus sorafenib. Outcomes in all randomly assigned patients, including final overall survival, are presented. Safety was assessed in all randomly assigned patients who received at least one dose of study drug. This trial is registered with ClinicalTrials.gov, NCT03755791. FINDINGS: Between Dec 7, 2018, and Aug 27, 2020, 432 patients were randomly assigned to combination treatment, 217 to sorafenib, and 188 to single-agent cabozantinib, and included in all efficacy analyses. 704 (84%) patients were male and 133 (16%) were female. 824 of these patients received at least one dose of study treatment and were included in the safety population. Median follow-up was 22·1 months (IQR 19·3-24·8). Median overall survival was 16·5 months (96% CI 14·5-18·7) for the combination treatment group and 15·5 months (12·2-20·0) for the sorafenib group (hazard ratio [HR] 0·98 [0·78-1·24]; stratified log-rank p=0·87). Median progression-free survival was 6·9 months (99% CI 5·7-8·2) for the combination treatment group, 4·3 months (2·9-6·1) for the sorafenib group, and 5·8 months (99% CI 5·4-8·2) for the single-agent cabozantinib group (HR 0·74 [0·56-0·97] for combination treatment vs sorafenib; HR 0·78 [99% CI 0·56-1·09], p=0·05, for single-agent cabozantinib vs sorafenib). Grade 3 or 4 adverse events occurred in 281 (66%) of 429 patients in the combination treatment group, 100 (48%) of 207 patients in the sorafenib group, and 108 (57%) of 188 patients in the single-agent cabozantinib group; the most common were hypertension (37 [9%] vs 17 [8%] vs 23 [12%]), palmar-plantar erythrodysaesthesia (36 [8%] vs 18 [9%] vs 16 [9%]), aspartate aminotransferase increased (42 [10%] vs eight [4%] vs 17 [9%]), and alanine aminotransferase increased (40 [9%] vs six [3%] vs 13 [7%]). Serious adverse events occurred in 223 (52%) patients in the combination treatment group, 84 (41%) patients in the sorafenib group, and 87 (46%) patients in the single agent cabozantinib group. Treatment-related deaths occurred in six (1%) patients in the combination treatment group (encephalopathy, hepatic failure, drug-induced liver injury, oesophageal varices haemorrhage, multiple organ dysfunction syndrome, and tumour lysis syndrome), one (<1%) in the sorafenib group (general physical health deterioration), and four (2%) in the single-agent cabozantinib group (asthenia, gastrointestinal haemorrhage, sepsis, and gastric perforation). INTERPRETATION: First-line cabozantinib plus atezolizumab did not improve overall survival versus sorafenib in patients with advanced hepatocellular carcinoma. The progression-free survival benefit of the combination versus sorafenib was maintained, with no new safety signals. FUNDING: Exelixis and Ipsen.