Introduction: Observational studies report lower COVID-19 mortality among metformin users, but most used prevalent-user designs comparing long-term survivors with heterogeneous controls. Randomised trials suggest little or no mortality effect. No large observational study has used a new-user design to estimate the effect of starting versus not starting metformin on severe and Long COVID outcomes in adults with type 2 diabetes (T2D). We aimed to estimate this effect on COVID-19-related hospitalisation or death, COVID-19 diagnoses, and Long COVID among adults with T2D in England. Research design and methods: Using OpenSAFELY-TPP linked NHS electronic health records, we conducted a population-based cohort study with a landmark new-user design. Adults aged 18-85 years newly diagnosed with T2D (August 2018-August 2019) were classified as starting metformin monotherapy within 6 months or not starting any antidiabetic drug, based on prescription data. Follow-up began 6 months after diagnosis and before the UK pandemic (avoiding immortal time) for up to 2 years. We estimated intention-to-treat and per-protocol effects using weighted Cox and pooled logistic regression, with negative (bone fracture) and positive (diabetes-related death) control outcomes. Results: Of 52,341 eligible people, 19,599 (37.4%) started metformin. For COVID-19-related hospitalisation or death, the maximally adjusted hazard ratio (HR) of metformin versus no antidiabetic was 1.17 (95% CI 1.02-1.35; 2-year risk difference 0.36%, 95% CI 0.09-0.62; intention-to-treat). The per-protocol HR for COVID-19-related hospitalisation or death was 1.11 (95% CI 0.96-1.29). The Long COVID estimate was imprecise (HR 1.39, 95% CI 0.83-2.33; intention-to-treat). For bone fracture (negative control) the estimate was null (HR 1.00), but for diabetes-related deaths (positive control) it was elevated (HR 1.41, 95% CI 1.09-1.81), indicating unmeasured confounding. Conclusions: The risks of severe and long COVID were slightly higher among people who started metformin, compared with those who did not start. These results plausibly reflect unmeasured confounding. These findings align more closely with randomised evidence than earlier prevalent-user studies and argue against a clinically meaningful benefit of starting metformin for COVID-19 prevention in newly diagnosed T2D.